Files
paper-how-to-use-this-techn…/research/05-heatmap-forecasting.md
T

128 KiB
Raw Blame History

Sparse Urban Sensing to Heatmaps and Spatiotemporal Forecasts

State of the art, failure modes, and a reference architecture for the PulseCast paper

Status: research memo, not paper prose. Key quantitative claims are tagged [PROJECT] (from the project data contract or generated metrics), [ESTIMATE] (a design calculation shown in the text), or cited; simple arithmetic from those inputs is left untagged. Literature claims are stated at the strength the cited evidence supports; contested findings are marked as contested.

Data context. Eight fixed footfall sensors in an inner-city cluster of roughly 580 m × 280 m; about five months of 15-minute data (June–November 2026); weather, public holidays, and events are available. Values are estimated devices, not people; device penetration is unknown. [PROJECT]

One-sentence thesis. With eight sensors, the defensible product is a probabilistic, network-aware heatmap at roughly 100–200 m support, plus calibrated sensor-level forecasts and an explicit decision layer — not a city-scale deep model and not a 10 m raster.

Contents. 0. Executive summary · 1. Sparse-to-heatmap · 2. Forecasting models · 3. Covariates · 4. Transfer/foundation models · 5. Evaluation · 6. Recommended architecture · 7. Citations


0. Executive summary

  1. Separate four estimands. (i) the expected count at each sensor, (ii) the latent spatial intensity field / heatmap, (iii) the forecast of either, and (iv) the decision (staffing, routing, operations). Most overclaiming in urban sensing comes from letting one model silently answer all four. The heatmap is a state estimate; the staffing plan is a decision under uncertainty; they need different validation.

  2. Spatial resolution is data-limited, not display-limited. Eight sensors over a 580 m × 280 m bounding box give an equivalent uniform spacing of about 142 m [ESTIMATE]. If the sensors lie along one shopping street, the effective spacing is closer to 580/8 ≈ 72 m along the street and there is no cross-street information [ESTIMATE]. A 100–150 m cell is defensible; a 25–50 m raster is a visualization, not an inference. Report the distance to the nearest sensor and the prediction standard deviation next to every heatmap.

  3. Use a likelihood that matches the target, not Gaussian kriging alone. Footfall is non-negative, overdispersed, and often zero-inflated. If the target is an integer count, use a negative-binomial/Poisson likelihood; if it is a continuous density index (as in this project), use a Gamma or lognormal likelihood. Either way, a hierarchical model with a low-rank spatial field (Matérn GP / SPDE or a small set of spatial basis functions) is the statistically correct default. Ordinary kriging on raw counts is a baseline; IDW is a visualization baseline. With eight sensors, a free CAR/BYM2 field on a fine grid is not identifiable — the prior would be doing almost all the work.

  4. Put the heatmap on the walkable network. Footfall is generated by street geometry, entrances, and public space, not by an isotropic Gaussian over building footprints. Euclidean smoothing across buildings produces plausible-looking but wrong maps and can reveal activity in private space. Use network distance / graph smoothing, or at minimum mask non-walkable cells and report the support.

  5. The forecast core should be a global (panel) gradient-boosted model with quantile outputs and conformal calibration. Calendar, weather forecast, holidays, events, POI, and lag features; sensor identity as a pooled effect. This is the strongest small-data baseline and is likely to beat an LSTM/Transformer/GNN on 8 series × 5 months. Deep spatiotemporal models and time-series foundation models should be benchmarks and ablations, not the default.

  6. The largest covariate leakage risk is weather. Training on observed weather and serving with forecast weather is optimistic. Evaluate with archived forecast weather (or a calibrated forecast-error simulation), and report both the observed-weather oracle and the forecast-weather number. Holidays are safe if official; events are safe only if scheduled and announced before the forecast issue time; POI snapshots must be as-of.

  7. Foundation-model evidence is mixed. Chronos, TimesFM, and Moirai report zero-shot performance comparable to supervised models on broad benchmarks, but independent evaluations find them weak on high-frequency, lumpy, transport-like series, and tuned statistical/tree ensembles often beat them. Use two or three open models as zero-shot benchmarks and possible ensemble members; do not make them the core of a covariate-rich, 8-sensor, 5-month task.

  8. Recommended architecture: a modular hierarchy, not a monolithic shared encoder. Data/provenance → feature store → spatial state model (low-rank Bayesian/GP) → global quantile GBM forecast → reconciliation → decision layer → evaluation harness. A shared deep encoder with task heads is justified at roughly 50–100+ sensors, multiple cities, or multiple tasks with abundant unlabeled data. At eight sensors, the statistically principled version of a "shared encoder" is a spatial basis + temporal factor model, which is what this memo recommends.


1. Sparse-to-heatmap

1.1 What is being estimated

Let (y_{it}) be the observed count or device-density proxy at sensor (i \in {1,\dots,8}), location (s_i), time (t). Let (\lambda(s,t)) be the latent intensity (expected devices per unit area per unit time) over a domain (\mathcal{D}), and let (c) index a target cell or street segment. The heatmap target is one of:

[ \mu_{c,t} = \int_{c} \lambda(s,t),ds \quad\text{(expected devices per cell per 15 min)} ]

or the intensive density

[ \bar\lambda_{c,t} = \frac{1}{|c|}\int_c \lambda(s,t),ds \quad\text{(devices per hectare per 15 min)}. ]

A third, privacy-friendly target is a relative index (r_{c,t} = \mu_{c,t} / \mu_{c,t_0}) against a fixed baseline. A fourth target — people — is not identifiable from device counts without external penetration data. The paper should say "estimated devices" or "density index," never "people," unless a calibration study exists.

Two support mismatches must be handled explicitly:

  • Point-to-area (change of support). Sensors are points; cells are areas. Counts are extensive (they add over area); densities are intensive (they average). Do not interpolate counts as if they were a continuous Gaussian field, then sum them over cells without accounting for the support. The change-of-support problem is classical in geostatistics and areal interpolation (Goodchild & Lam 1980; Gotway & Young 2002).
  • Network vs plane. If the sensors are along a street, the domain is effectively one-dimensional along the pedestrian network. A 2D Euclidean variogram will overstate the information available across the street and will smooth over building blocks. The heatmap should be indexed by street segments or walkable cells, not by a building-footprint raster.

Recommendation. Define the target as expected devices per 15 min on a walkable 100–150 m grid or street-segment graph. Publish the support, the aggregation rule, and a reliability mask. Use a relative index for public display if the absolute device counts are sensitive.

1.2 Geometry first: is this a 1D or 2D problem?

Before choosing an interpolator, compute and report:

  • sensor coordinates, nearest-neighbor distances, and the convex hull;
  • the fraction of the hull area that is actually walkable public space;
  • the eigenvalues of the sensor coordinate matrix (a dominant first eigenvalue means the network is essentially 1D);
  • the street-network distance matrix vs the Euclidean distance matrix (if they differ materially, use network distance).

A simple density calculation gives the scale:

Geometry assumption Equivalent spacing Defensible support Caveat
2D areal, 580 m × 280 m = 162,400 m² (\sqrt{162{,}400/8} \approx 142) m 150–250 m cells Only if sensors fill the area
1D street, 580 m long (580/8 \approx 72) m 70–100 m along-street segments No cross-street inference
Sparse/irregular cluster compute nearest-neighbor distances nearest-neighbor scale Report hull coverage; do not extrapolate outside

[ESTIMATE] These are order-of-magnitude heuristics, not precision guarantees. The fitted variogram and leave-one-sensor-out (LOSO) error are the real arbiters.

1.3 Interpolation methods: what each one actually assumes

All spatial smoothing rests on Tobler's first law — near things are more related than distant things (Tobler 1970) — but the methods differ in what they assume about the mean, the covariance, and the likelihood. Classical geostatistics is treated in Cressie (1993) and Diggle & Ribeiro (2007); Bayesian hierarchical spatial models in Banerjee et al. (2014). Reviews of environmental interpolation find no universally best method: performance is data-specific, and combined methods such as regression kriging are often competitive (Li & Heap 2011, 2014).

Inverse distance weighting (IDW)

[ \hat y(s_0) = \frac{\sum_i w_i, y(s_i)}{\sum_i w_i}, \qquad w_i = d(s_0,s_i)^{-p}, \quad p \approx 1\text{–}3. ]

IDW is transparent, fast, and has no distributional assumptions. It also has no uncertainty estimate, produces "bull's-eye" artifacts around sensors, and cannot use covariates. It is a visualization baseline, not a scientific spatial model. If it is used, report (p), the search radius, and a LOSO error; do not present the map as a measured surface.

Splines and radial basis functions (RBF)

Thin-plate splines and penalized regression splines fit a smooth surface by minimizing data misfit plus a roughness penalty (Wahba 1990; Wood 2017). They are the deterministic cousin of a Gaussian process with a specific kernel. They handle non-uniform sensor density better than IDW and can include covariates in a generalized additive model (GAM). They do not naturally respect count likelihoods, and uncertainty requires bootstrapping or a Bayesian extension. Use them when the target is a smooth density index and the covariate model is additive; do not use them to claim fine-scale features.

Kriging / Gaussian processes

Ordinary kriging assumes a constant unknown mean and a stationary covariance. In matrix form,

[ \hat{\mathbf{y}}_0 = \mathbf{c}0^\top \mathbf{C}^{-1}\mathbf{y}, \qquad \sigma^2{OK}(s_0) = C(0) - \mathbf{c}_0^\top \mathbf{C}^{-1}\mathbf{c}_0 + \mu, ]

where (\mathbf{C}) is the sensor covariance matrix, (\mathbf{c}_0) the sensor-to-target covariance, and (\mu) the Lagrange multiplier for the unbiasedness constraint. Universal kriging replaces the constant mean with a linear trend in covariates; regression kriging fits a covariate model first and kriges the residuals.

A GP is the Bayesian version: (f \sim \mathcal{GP}(m(s), k_\theta(s,s'))), (y_i = f(s_i) + \varepsilon_i) (Rasmussen & Williams 2006). The posterior predictive variance is the natural uncertainty estimate; for sparse GMRF representations, see Rue & Held (2005). The Matérn covariance is the standard choice for spatial data because its smoothness parameter is interpretable:

[ k_\theta(d) = \sigma^2 \frac{2^{1-\nu}}{\Gamma(\nu)} \left(\frac{\sqrt{2\nu},d}{\ell}\right)^\nu K_\nu!\left(\frac{\sqrt{2\nu},d}{\ell}\right), ]

with range (\ell), marginal variance (\sigma^2), and smoothness (\nu). For areal data or a fine grid, the SPDE approach represents a Matérn field as a sparse Gaussian Markov random field, which makes Bayesian inference scalable (Lindgren et al. 2011; Rue et al. 2009).

The small-data problem. A variogram needs enough pairs per lag bin; conventional guidance is at least ~30 pairs per bin (Webster & Oliver 2007). Eight sensors give 28 pairs per time slice, so a single-slice empirical variogram is unstable. Pooling over time helps, but the field may be non-stationary (weekday vs weekend, day vs night). With 8 sensors, estimate a parametric Matérn covariance with a hierarchical prior, or fix the smoothness and estimate only range and nugget; do not fit a free-form variogram with many bins.

Counts. Gaussian kriging on raw counts violates the mean–variance relationship and can produce negative predictions. Options: (i) transform (\log(1+y)), krige, back-transform with a bias correction; (ii) Poisson/NB kriging (a generalized linear spatial model); (iii) regression kriging with a count GLM mean; (iv) a hierarchical count model (next subsection). Option (iv) is the most principled.

Bayesian hierarchical models

The canonical model for areal counts is

[ y_{it} \sim \text{NegBin}(\lambda_{it}, \phi), \qquad \log \lambda_{it} = \log E_{it} + \alpha + \mathbf{z}{it}^\top\boldsymbol{\beta} + u_i + v_i + \gamma_t + \delta_t + \epsilon{it}, ]

where (E_{it}) is an offset (catchment, exposure, or device-detection factor), (u_i) is a spatially structured effect, (v_i) is unstructured heterogeneity, (\gamma_t) is a temporal trend, (\delta_t) is seasonal/time-of-week structure, and (\epsilon_{it}) is a space–time interaction. If the project's density readings are continuous rather than integer counts, replace the negative binomial with a Gamma or lognormal likelihood; the linear predictor and spatial field are unchanged. If zeros are structural (store closed, sensor downtime), use a hurdle/two-part model rather than letting one likelihood conflate "no activity" with "no data." The Besag–York–Mollié (BYM) model and its scaled BYM2 parameterisation are the standard choices for (u_i+v_i) (Besag et al. 1991; Riebler et al. 2016). Penalised-complexity priors make the variance components interpretable and weakly informative (Simpson et al. 2017). Knorr-Held (2000) defines four types of space–time interaction, from independent to fully inseparable; with five months and eight sensors, only the simpler types are identifiable.

What this buys. A likelihood that respects counts, a principled way to borrow strength across sensors and time, posterior predictive intervals, and a natural handle for missing data and covariates. What it costs. Identifiability. With eight sensors, a free spatial random effect plus a sensor intercept plus a fine-grid CAR field is not identified. Use one of:

  • a low-rank spatial basis: (u(s,t) = \sum_{k=1}^{K}\phi_k(s)a_k(t)) with (K \ll 8) (e.g., 3–6 factors), where (\phi_k) are Matérn/RBF/network-diffusion basis functions and (a_k(t)) are temporal factors;
  • a Matérn GP with 3–4 hyperparameters and a sum-to-zero constraint;
  • regression kriging: a covariate mean model plus a low-rank residual GP.

The low-rank factor form is the most useful for this project: it is a statistical "shared encoder" and it gives coherent heatmaps and forecasts.

Spatial confounding. A spatially smooth random effect can absorb the effect of a spatially smooth covariate (e.g., POI density), biasing the covariate coefficient. This is the spatial analogue of omitted-variable bias and is well documented (Hodges & Reich 2010). With eight sensors, do not interpret a POI coefficient as causal while also including a flexible spatial field. Either fix the spatial field to a prespecified basis or report the sensitivity of the coefficient to the spatial model.

Graph-based smoothing

Represent sensors (and optionally grid cells) as nodes of a graph with edge weights (w_{ij}) from distance, network travel time, or co-visitation. Graph signal processing provides the vocabulary and the spectral tools for this representation (Shuman et al. 2013; Ortega et al. 2018). Graph Laplacian smoothing minimizes

[ \sum_i (y_i - f_i)^2 + \lambda \sum_{i,j} w_{ij}(f_i - f_j)^2 = |\mathbf{y}-\mathbf{f}|^2 + \lambda,\mathbf{f}^\top \mathbf{L}\mathbf{f}, ]

where (\mathbf{L}) is the graph Laplacian. This is equivalent to a Gaussian process with a graph-based kernel. It handles non-Euclidean geometry (a street network, a mall corridor) better than isotropic kriging, and it can encode barriers (e.g., a river, a building block) by setting (w_{ij}=0).

Graph neural network kriging (e.g., IGNNK) learns message passing from random masked subgraphs and can be inductive to new sensors (Wu et al. 2021). It is attractive when there are many sensors and a meaningful graph. With eight nodes, the graph is nearly complete, and a GNN reduces to a learned (8\times 8) mixing matrix with little data to train it. Use graph smoothing as a structured prior (Laplacian regularization) rather than a deep GNN at this scale.

Matrix/tensor completion

Low-rank plus temporal smoothing can recover missing sensor values, but it is transductive (a new sensor requires retraining), usually ignores covariates, and assumes a low-rank structure that may not hold across a heterogeneous cluster. It is a useful imputation baseline, not a heatmap method.

Comparison table

Method Output UQ Count support Covariates Data need Main failure mode Role here
IDW Smooth surface None No No Very low Bull's-eyes; no uncertainty Visualization baseline
RBF / thin-plate spline Smooth surface Bootstrap No Yes (GAM) Low Oversmooths; boundary artifacts Baseline / mean model
Ordinary kriging BLUP + variance Yes (Gaussian) No No Moderate Variogram unstable with 8 sensors Baseline
Universal/regression kriging Trend + residual field Yes With GLM mean Yes Moderate Mean model misspecification Strong practical baseline
GP / SPDE Posterior field Yes Via likelihood Yes Moderate Stationarity; smoothness mismatch Core spatial prior
Hierarchical NB + BYM2/GP Posterior counts Yes Yes Yes Moderate Identifiability with 8 sensors Core if carefully parameterised
Low-rank factor model Factors + map Yes Yes Yes Low–moderate Rank choice; interpretability Recommended core
Graph Laplacian smoothing Network field Via Bayesian form Yes Yes Low Graph construction; oversmoothing Network-aware prior
GNN kriging (IGNNK) Reconstructed field Usually no Yes Limited High Overfits 8 nodes; no calibrated UQ Ablation only
Matrix/tensor completion Imputed matrix Sometimes Approx. No Moderate Transductive; ignores covariates Imputation baseline

A nonparametric alternative worth testing is random forest (or gradient boosting) on spatial coordinates plus covariates, followed by kriging of residuals; it can capture nonlinear covariate effects but does not extrapolate and needs careful spatial cross-validation (Hengl et al. 2018). In practice, the ranking of methods is dataset-specific; the only defensible approach is to compare them on the same LOSO folds.

1.4 Uncertainty quantification

There are at least three uncertainties, and they should not be collapsed into one interval:

  1. Observation noise. Device counting, sensor detection radius, Wi-Fi/BLE probe variability, and aggregation. This is estimable from repeat measurements or assumed.
  2. Spatial interpolation uncertainty. The GP/kriging posterior variance, or the posterior predictive variance of the hierarchical model. This grows with distance to the nearest sensor and with the nugget-to-sill ratio.
  3. Representativeness / structural uncertainty. Device penetration, tourist vs resident mix, privacy noise, and the difference between "devices" and "people." This is not identified from the sensor data. It should be reported as a bias/limitation, not as a random interval.

For a heatmap, publish:

  • posterior mean or median;
  • 80% and 95% credible intervals per cell;
  • a reliability mask for cells where the interval width exceeds a stated tolerance (e.g., 50% of the prior standard deviation, or a fixed device-count threshold);
  • a distance-to-nearest-sensor layer;
  • calibration diagnostics: empirical coverage by distance band and by sensor.

A useful rule: if the 80% interval for a cell is wider than the cell's mean, the cell should be masked or aggregated upward, not colored.

Calibration under LOSO. Hold out one sensor, predict its series from the other seven, and compute interval coverage at the held-out location. If coverage is below nominal, widen the intervals (conformal or posterior predictive recalibration); if coverage is above nominal but intervals are very wide, the model is underconfident and the map is not sharp enough to be useful. Report both.

Conformal option. For the spatial model, split conformal on LOSO residuals can calibrate intervals without a full Bayesian model, but exchangeability across space is questionable. Use spatial-block conformal (hold out contiguous regions) or report LOSO coverage as an empirical check rather than a guarantee.

1.5 How many sensors are needed, and what resolution is achievable

This is the question reviewers will ask, and it has a defensible empirical answer.

Step 1: geometry and variogram. Compute the empirical variogram (pooled over time, stratified by day type), fit a Matérn or exponential model, and report range, nugget, and sill. The range (\ell) is the distance over which a sensor informs a cell. If (\ell) is 100 m, a 50 m grid is not independently informed between sensors. If (\ell) is 500 m, the whole cluster is one smooth field and fine cells are cosmetic.

Step 2: LOSO learning curve. Subsample the eight sensors: for each (k=2,\dots,7), train on (k) sensors and predict the held-out sensors (or use (k)-fold LOSO). Plot error and interval width against (k). This gives the marginal value of the 9th sensor and a saturation point. With 8 sensors there are only 255 non-empty subsets (247 of size 2–7), so the experiment is cheap.

Step 3: design calculation. In the independent-sampling limit, prediction variance scales roughly as (1/n), so halving the standard error needs about (4\times) the sensors. With a smooth spatial field, gains saturate once the spacing is below the correlation range; the constant depends on the variogram. Do not quote the (1/n) rule as a law; use it as a prior for the learning-curve fit.

Step 4: target-error inversion. Decide the tolerable error for the decision (e.g., staffing needs ±20% at the store, not ±5% in every cell). Invert the learning curve to find the sensor count that meets it at the target support. This is more honest than a universal "you need N sensors" answer.

**A practical answer for this project.**Eight sensors over a 580 m × 280 m cluster support a 100–200 m heatmap with useful uncertainty near sensors and wide intervals between them. A 50 m grid is a visualization; a 25 m grid is misleading. For a 50 m inferential grid, a density argument suggests roughly 60–100 sensors (one per 50 m × 50 m cell is 65 sensors; redundancy for a smooth field may reduce this to 30–60) [ESTIMATE]. State this as a design calculation, not a guarantee.

Placement matters more than count. If sensors are added, place them to maximize mutual information or minimize kriging variance, not uniformly (Krause et al. 2008). With a street network, the optimal placement is usually at junctions, store entrances, and the ends of the street, not at the centroid of a grid.

Resolution limits. Three constraints bind:

  • Sampling/Nyquist-like: a feature of scale (L) needs spacing (\lesssim L/2) to be resolved; with 142 m spacing, features smaller than ~280 m are not independently resolved.
  • Variogram range: beyond (\ell), the prediction reverts to the mean; the map is a prior, not data.
  • Noise: if the nugget is large relative to the sill, even dense sensors cannot resolve fine spatial variation.

MAUP and the ecological fallacy. Results depend on the cell boundaries and aggregation scale (Openshaw 1984). Report the heatmap at two or three supports (e.g., 50, 100, 200 m) and show that the qualitative conclusion is stable. Do not compare a 50 m map to a 200 m map without aggregation. Do not infer individual behavior from cell-level averages.

1.6 Fusion of heterogeneous sources

A "heatmap from heterogeneous sources" is a data-fusion problem, not a single interpolation. Data fusion has a long taxonomy of architectures and trade-offs (Castanedo 2013), and urban computing provides the application context (Zheng et al. 2014). Sources differ in support, latency, bias, and privacy constraints; mobile-phone datasets in particular have well-documented sampling and representativeness issues (Blondel et al. 2015).

Source What it measures Support Strengths Weaknesses / biases
Fixed footfall counters Devices/persons at a point Point, fixed Consistent, calibrated, high cadence Sparse; detection radius; device penetration
Cameras / computer vision Persons or vehicles in view Polygon/segment Rich attributes (direction, dwell) Privacy; occlusion; calibration drift; compute
GPS/probe (app, fleet) Trajectories from a sample Point/line High volume; movement Sample bias (app users, drivers); sparse in pedestrian areas
Telco / mobile network Devices associated with cells Coarse area Broad coverage Operator bias; coarse spatial; penetration; privacy
Wi-Fi/BLE Nearby devices Point, short range Indoor/street presence MAC randomization; range; dwell bias
Transit ridership Entries/exits Station Reliable, scheduled Only transit users; no path
POI / land use Attractors Polygon/point Explains spatial pattern Static; opening hours; future changes
Weather Exogenous driver Station/grid Available forecast Forecast error; microclimate mismatch

Fusion architectures. Four patterns, in increasing sophistication:

  1. Feature-level fusion. Concatenate source features into one model. Simple, but ignores source-specific bias and support.
  2. Late fusion / ensemble. Fit one model per source and combine predictions. Robust, but does not propagate shared latent structure.
  3. State-space / data assimilation. A latent field evolves in time; each source has an observation operator (H_k) and noise model: (\mathbf{y}_k = H_k \mathbf{x} + \boldsymbol{\epsilon}_k). Kalman/ensemble Kalman or particle filters update the field (Evensen 2003; Carrassi et al. 2018). This is the principled way to combine timely but noisy proxies with sparse accurate counters.
  4. Hierarchical observation model. Each source has its own likelihood and detection bias, all conditioned on a shared latent intensity. This is a generalization of (3) and is the best fit for a paper that wants a coherent heatmap with uncertainty.

Calibration. Do not treat a proxy as truth. If (y_{\text{proxy}}) measures the target with multiplicative bias and noise,

[ y_{\text{gold}} = a + b,y_{\text{proxy}} + \mathbf{c}^\top\mathbf{z} + e, ]

and regressing gold on proxy without accounting for measurement error in the proxy attenuates (b) toward zero. Use regression calibration, Deming/orthogonal regression, or a Bayesian errors-in-variables model (Carroll et al. 2006). Validate on held-out days and held-out locations, not on the same proxy. Report calibration curves by source and by context (weekday/weekend, rain/dry).

Normalization. Define a common currency before fusion: devices per 15 min per cell, or a relative index against a fixed baseline. Per-sensor z-scoring destroys the spatial pattern and makes the heatmap uninterpretable; if used for model training, invert it before mapping. For cross-city transfer, use a relative-to-baseline index (like the Google mobility reports) and document the baseline period.

Fusion evaluation. Leave-one-source-out: does adding source (k) improve LOSO error and calibration? If not, drop it. Report the fusion gain relative to the best single source. Do not evaluate a fused map only against one of its inputs.

1.7 Privacy-preserving aggregation

Footfall data are personal data at the device level, even when aggregated. The paper should describe a privacy budget and an aggregation rule.

  • k-anonymity / aggregation thresholds. Suppress cells with fewer than (k) devices (e.g., (k=10) or 20) and time bins with insufficient devices (Sweeney 2002). This prevents re-identification in small cells but does not protect against differencing attacks; (l)-diversity or DP is needed for stronger guarantees (Machanavajjhala et al. 2007).
  • Differential privacy (DP). Add calibrated noise to counts or to the sufficient statistics. DP composes across queries; repeated heatmap releases can exhaust the budget (Dwork 2006). Use a consistent noisy count and post-process (non-negativity, hierarchical reconciliation) rather than independent noise per cell.
  • Geo-indistinguishability. Perturb locations with planar Laplace noise before aggregation; this is DP for location with a distance-dependent guarantee (Andrés et al. 2013).
  • Suppression vs noise. Suppression is simple but biases the map (missing small cells); DP is principled but noisy for small counts. With 8 sensors and a small cluster, cell-level DP at fine resolution will destroy the signal. Aggregate first, then add noise at the published support.
  • Relative indices. Publishing percentage change from a baseline, rather than absolute counts, reduces disclosure risk and is the approach used by Google's Community Mobility Reports (Aktay et al. 2020).
  • Representativeness is separate. DP protects individuals; it does not fix device-penetration bias. A DP-noisy map of a biased sample is still a biased map of devices. Say both.

Recommended disclosure policy for this project. Publish sensor-level aggregates and a 150 m heatmap with cells below (k=10) devices suppressed or merged; publish relative indices for public display; keep raw device-level data out of the repository; document the privacy rule in the data card.

  1. Fix the estimand and support. Expected devices per 15 min on a walkable 100–150 m grid / street-segment graph. Publish the aggregation rule.
  2. QC and missingness. Flag sensor outages, gaps, and battery/network failures. Distinguish missing at random (network) from missing because the site was closed; a zero and a missing value are different. Do not silently interpolate; a missing sensor is informative. Report coverage per sensor.
  3. Geometry and variogram. Compute nearest-neighbor distances, hull coverage, network vs Euclidean distances; fit a Matérn variogram with pooled day-type strata.
  4. Mean model. Fit a pooled count GLM/GBM with calendar, weather, holiday, event, POI, and sensor effects. This captures the deterministic structure.
  5. Spatial residual model. Fit a low-rank Matérn/RBF/network-basis field to the residuals (or a hierarchical NB model with the same basis). Use 3–6 factors; do not fit a free 1000-cell CAR field.
  6. Posterior predictive map. Sample the field on the grid; aggregate to cells; compute mean and 80/95% intervals; mask unreliable cells.
  7. Validate LOSO. Hold out each sensor; predict its series; report RMSE/MAE/WAPE, CRPS, coverage, and interval width. Compare IDW, ordinary kriging, regression kriging, the low-rank model, and a GNN kriging ablation.
  8. Report resolution. Show LOSO error vs cell size and vs distance to nearest sensor. State the effective resolution and the sensor count needed for a finer one.

2. Forecasting models

2.1 Regimes and baselines

The project has four horizons with different physics:

Horizon Regime Dominant signal Model emphasis
15 min Nowcast / very short Current level, lagged counts, time of day Lags, state-space, persistence
1 h Short Intraday profile, weather now Lags + calendar + weather
48 h Medium Daily/weekly cycle, forecast weather, holiday Calendar + weather forecast + events
7 d Long Weekly cycle, season, holidays, events Calendar + climatology + scenario

A model that wins at 15 min may lose at 7 d. Report metrics per horizon, not only averaged. The very short horizons are a nowcasting problem — estimating the current state from timely but incomplete indicators — which has its own literature and evaluation logic (Giannone et al. 2008). If 15-minute values are sparse or noisy, model at hourly resolution and disaggregate with a learned profile; publish hourly heatmaps and treat 15-minute maps as animation, not as independent estimates.

Baselines that must be reported (tuned, not strawmen):

  • Seasonal naive: (\hat y_{t+h} = y_{t+h-s}), with (s=96) (same time yesterday) or (s=672) (same time last week). This is surprisingly strong for footfall.
  • Daily profile: average of the same hour-of-week over the training window.
  • ETS / SARIMA: per-sensor or global; linear, interpretable, handles autocorrelation.
  • Prophet / GAM: trend + weekly + holiday; robust to missing data.
  • Poisson/NB GLM with calendar and weather; a strong interpretable count baseline.
  • Global HistGradientBoosting: the main baseline for this project.

The baseline table is not optional. A paper that claims a deep model wins without a tuned seasonal naive and a GBM is not credible.

2.2 Classical statistical models

SARIMA/SARIMAX. SARIMA models autocorrelation and seasonality with AR/MA terms and differencing (Box et al. 2015); SARIMAX adds exogenous regressors. With 8 series and ~14,400 points each, per-series estimation is feasible but unstable; a global/panel model or a hierarchical shrinkage is better. SARIMA assumes linearity and Gaussian errors; prediction intervals are symmetric and can go negative. Use it as an interpretable baseline and for residual diagnostics.

Prophet. A generalized additive model with piecewise-linear trend, Fourier seasonality, and holiday dummies (Taylor & Letham 2018). It is robust to missing data and irregular spacing and is easy to explain. It is not designed for 15-minute count data with strong interactions, and its uncertainty intervals are not calibrated for overdispersed counts. The Forecasting: Principles and Practice text notes that Prophet rarely beats ETS/ARIMA on standard examples (Hyndman & Athanasopoulos 2021). Use it as a baseline, especially for daily/48h horizons, not as the core.

State-space / exponential smoothing. ETS and TBATS are strong automatic baselines. A structural time series with a local level, weekly season, and regression effects is a natural small-data model and gives interpretable uncertainty. For counts, a Poisson state-space model is preferable.

Poisson/NB GLM with splines. A generalized additive model with a log link, cyclic splines for hour-of-week, linear/spline terms for weather, and holiday/event dummies is a strong, interpretable count model. It can be fitted with mgcv/statsmodels/Stan. It does not capture nonlinear interactions automatically; add domain-informed interactions (rain × weekend, temperature × season) sparingly.

2.3 Gradient boosting

Gradient boosting fits an additive ensemble of shallow trees by stagewise optimization (Friedman 2001). Implementations: XGBoost (Chen & Guestrin 2016), LightGBM (Ke et al. 2017), scikit-learn HistGradientBoostingRegressor. For this project, a global model across the 8 sensors with sensor identity as a categorical feature is the right default: it pools the calendar/weather response while allowing sensor-level intercepts and interactions.

Why it is strong here.

  • Tabular features with nonlinearities and interactions (hour × weekend, rain × temperature) are its home turf.
  • It handles missing values, mixed types, and irrelevant features gracefully.
  • It trains in seconds, which matters for a hackathon and for reproducible rolling-origin evaluation.
  • Tree-based models remain state of the art on medium-sized tabular data; deep learning does not reliably beat them at ~10k–100k rows (Grinsztajn et al. 2022; Shwartz-Ziv & Armon 2022).
  • M5 results showed strong performance of LightGBM-based methods and combinations; the best ML method was about 22.4% more accurate than the best statistical benchmark (Makridakis et al. 2022).

Design details.

  • Features: hour-of-week (cyclic or categorical), day-of-year/season, holiday type and proximity, school term, weather (temp, precip, wind, humidity, cloud, lags), event features, POI densities, sensor ID, lagged counts and rolling means (respecting the horizon), and optionally neighboring-sensor lags.
  • Direct multi-horizon: train one model per horizon (or per horizon bucket) rather than recursive one-step forecasting. Recursive models accumulate error and are harder to calibrate at 7 d. Direct models use only information available at the issue time.
  • Quantile loss: fit separate models for (\tau \in {0.05, 0.1, 0.25, 0.5, 0.75, 0.9, 0.95}), or use a multi-output quantile GBM. Enforce monotonicity by sorting/rearranging quantiles.
  • Poisson/NB objective for counts if available; otherwise train on (\log(1+y)) with a Gaussian/quantile loss and invert. Report the transformation.
  • Early stopping on a time-based validation window, not a random split.
  • Sample weights: recency weighting can help under drift, but it reduces effective sample size; test it.
  • Categorical handling: native categorical support in LightGBM/HistGB; do not one-hot 168 hour-of-week categories with 5 months of data.

Weaknesses. Trees cannot extrapolate beyond the training range (e.g., a heatwave, a new event scale); they do not enforce spatial smoothness; they can overfit rare holidays/events; and they have no native coherent probabilistic output. Address these with quantile + conformal calibration, hierarchical priors for rare events, and spatial reconciliation.

NGBoost. Natural-gradient boosting predicts the parameters of a conditional distribution (e.g., Normal, LogNormal, NegativeBinomial) and is trained with a proper scoring rule (Duan et al. 2020). It is a principled way to get probabilistic GBM forecasts without fitting many quantile models. It can be slower and less mature than LightGBM; use it as an alternative to quantile GBM and compare CRPS.

2.4 Neural sequence models

LSTM/GRU. Recurrent networks can learn nonlinear temporal dependence and long-range interactions (Hochreiter & Schmidhuber 1997). A global LSTM across 8 series (with sensor embeddings) is a reasonable neural baseline. With ~14k time steps per series but strong autocorrelation, the effective sample is much smaller than 115k; LSTMs are prone to overfit and are hard to tune. Use them as a benchmark, not a default. DeepAR is the probabilistic version: an autoregressive RNN trained on many related series with a likelihood (NegativeBinomial/Poisson/Gaussian), which handles counts and covariates and produces Monte Carlo predictive samples (Salinas et al. 2020). With only 8 series, the "many related series" advantage is limited.

TCN. Temporal convolutional networks use dilated causal convolutions and residual blocks; they are parallel, stable, and have a long receptive field (Bai et al. 2018). A small TCN (2–4 layers, 16–64 channels) is a sensible neural encoder for a shared spatiotemporal model if you go that route. It is still data-hungry relative to a GBM.

Transformers. The vanilla Transformer (Vaswani et al. 2017), Informer (Zhou et al. 2021), Autoformer, FEDformer, and PatchTST (Nie et al. 2023) are strong on long-horizon benchmarks with large datasets. Two cautions: (i) self-attention is permutation-invariant, and positional encodings must carry the temporal order; (ii) simple linear models (DLinear) often match or beat Transformer forecasters on long-horizon benchmarks (Zeng et al. 2023). PatchTST's channel-independent patching is a strong design, but it still needs more data than this project has. Use a small PatchTST/Transformer as an ablation if compute allows; do not expect it to beat a tuned GBM on 8 series.

N-BEATS / N-HiTS. MLP-based residual stacks with basis expansion; strong on M4/M5, fast, and interpretable (Oreshkin et al. 2020; Challu et al. 2023). They are univariate by design, so exogenous covariates require an extension; they do not use cross-sensor information. Good for long-horizon univariate baselines, not for the core covariate-rich model.

TFT. The Temporal Fusion Transformer is the neural architecture whose input design best matches this problem: static covariates (sensor, POI), known future inputs (calendar, holiday, weather forecast), observed past inputs (lags, observed weather), and quantile outputs, with variable selection and interpretable attention (Lim et al. 2021). It is a plausible benchmark and a good reference for how to separate covariate types. It is parameter-heavy; with 8 series, it needs careful regularization and a global training setup. If it beats the GBM on rolling-origin and LOSO with multiple seeds, it is a legitimate paper contribution; if not, report that.

Diffusion models. TimeGrad and CSDI generate probabilistic forecasts by learning a denoising process (Rasul et al. 2021; Tashiro et al. 2021). They are flexible and can capture multimodal, heteroscedastic distributions, but they are computationally expensive, need many samples, are hard to calibrate, and are overkill for 8 series with 5 months. Use them only if scenario generation with rich joint distributions is a core requirement.

2.5 Spatiotemporal graph neural networks

DCRNN, STGCN, Graph WaveNet, ASTGCN, and GMAN are the canonical models (Li et al. 2018; Yu et al. 2018; Wu et al. 2019; Guo et al. 2019). They combine graph convolution (spatial message passing) with recurrent or convolutional temporal layers. They are designed for tens to thousands of nodes (traffic networks) and often assume a meaningful adjacency matrix.

With 8 nodes, three problems arise:

  1. The graph is nearly complete. A learned (8\times 8) adjacency has 64 parameters, but the effective spatial signal is weak; the model can memorize sensor pairs.
  2. Data hunger. STGNNs have many parameters and are trained on months of high-frequency data; 8 series × 5 months is small. The benchmark gains in the literature often come from datasets with hundreds of sensors and years of data.
  3. Evaluation pitfalls. GNN benchmarks are sensitive to random splits, graph construction, and baseline tuning; simple baselines are often competitive (Shchur et al. 2018; Corradini et al. 2026). Building the graph from target correlations over the full period leaks information; build it from external geometry (street network, travel time, POI) or from training-period data only.

Verdict. Use a GNN only as an ablation to test whether spatial message passing adds value over a sensor-embedding GBM. If it does not, that is a useful negative result. Do not make it the core architecture for eight sensors.

2.6 Model comparison table

Family Representative Data need Covariates Probabilistic output Strengths Weaknesses / verdict
Seasonal naive / ETS statsmodels ETS Very low No ETS yes Strong, transparent baseline No covariates; verdict: mandatory baseline
SARIMAX statsmodels Low–moderate Yes (linear) Gaussian intervals Interpretable; autocorrelation Linear; unstable per series; baseline
Prophet / GAM Prophet, mgcv Low Yes (additive) Approx. Holidays, missing data Weak on 15-min counts; baseline
Poisson/NB GLM statsmodels, Stan Low Yes Yes Correct count likelihood; interpretable Limited interactions; baseline/core mean model
Gradient boosting LightGBM, HistGB, XGBoost Moderate Yes Quantile/NGBoost Strong tabular accuracy; fast No extrapolation; no spatial coherence; core
NGBoost ngboost Moderate Yes Full distribution Proper scoring; flexible Slower; less mature; alternative core
Global LSTM/DeepAR PyTorch, GluonTS High Yes Yes (likelihood) Nonlinear; probabilistic Data-hungry; tuning; benchmark
TCN PyTorch High Yes Quantile Stable, parallel, long receptive field Data-hungry; benchmark/encoder
Transformer / PatchTST PyTorch Very high Limited Quantile Long horizon; scaling Overfits small data; ablation
N-BEATS / N-HiTS neuralforecast High Limited Quantile Strong univariate long-horizon No cross-sensor; ablation
TFT PyTorch, pytorch-forecasting High Yes (static/future/past) Quantile Best neural covariate design Parameter-heavy; benchmark
STGNN DCRNN/STGCN/Graph WaveNet/ASTGCN Very high Yes Usually point Spatial message passing 8 nodes; overfits; ablation only
Diffusion TimeGrad, CSDI Very high Yes Yes (samples) Flexible multimodal Expensive; overkill; not recommended
Foundation model Chronos, TimesFM, Moirai Zero/few-shot Usually no Yes No training; cold start No covariates; mixed evidence; benchmark

2.7 Small-data evidence and model selection

The empirical literature is consistent on a few points, although concerns about fair comparison and reproducibility of ML forecasting are long-standing (Makridakis et al. 2018):

  • M4/M5: top results are dominated by hybrids, combinations, and strong gradient-boosting methods, not monolithic deep nets; the best M5 ML method was ~22.4% better than the best statistical benchmark (Makridakis et al. 2020, 2022).
  • Tabular data: tree ensembles remain state of the art on medium-sized data; deep learning's advantage is not established (Grinsztajn et al. 2022; Shwartz-Ziv & Armon 2022).
  • Long-horizon Transformers: simple linear models are competitive; the Transformer advantage is dataset-dependent (Zeng et al. 2023).
  • Global vs local: a global model can be as expressive as local models and generalizes better when series are related (Montero-Manso & Hyndman 2021). With 8 related sensors, a global/panel model is sensible.
  • GNNs: benchmark gains are sensitive to evaluation protocol; simple baselines are often competitive (Shchur et al. 2018; Corradini et al. 2026).

Promotion rule. A complex model replaces the GBM only if it wins on (i) rolling-origin test folds, (ii) LOSO, (iii) at least 5 seeds for stochastic models, (iv) a paired significance test, and (v) the decision-level cost metric — not just MAE on one holdout. Otherwise, keep the simpler model and report the negative result.

2.8 Coherent spatial forecasts

Forecasting each sensor independently and then interpolating can produce a heatmap whose cell values do not sum to the forecast cluster total, and whose spatial pattern is noisy. Two remedies:

  1. Forecast the latent factors, then map. If the spatial field is (u(s,t)=\sum_k \phi_k(s)a_k(t)), forecast the factors (a_k(t)) jointly (VAR/state-space/GBM with covariates), then map to sensors and cells. This yields coherent sensor and cell forecasts by construction.
  2. Reconcile after forecasting. Treat the sensors and the cluster total as a hierarchy; apply MinT reconciliation to make forecasts coherent (Hyndman et al. 2011; Wickramasuriya et al. 2019). For probabilistic forecasts, reconcile samples (simulate, reconcile each sample, then summarize) or use a probabilistic reconciliation method. Enforce non-negativity.

Reconciliation is a small addition that materially improves the credibility of a heatmap. Report the coherence error before and after.


3. Covariates

3.1 Taxonomy: availability determines what can be used

The single most useful discipline for covariates is to label each one by when it is known. A feature that is observed at time (t) but not available at the forecast issue time (t_0) is leakage, even if it is "exogenous."

Covariate Availability at issue time Encoding Leakage risk Expected role Notes
Hour of day, day of week Known Cyclic splines or categorical None Strong Use local time zone; 168 hour-of-week categories
Public holidays Known (official calendar) Binary + type + proximity Low Moderate Few holidays in 5 months; pool effects
School term / breaks Known (official calendar) Binary + term type Low Moderate Affects weekday daytime
Scheduled events Known if announced Type, attendance, distance, window Medium Moderate–strong locally Only if announcement precedes (t_0)
Unscheduled events Unknown Surprise flag at evaluation High if used in training as known Local anomaly Do not include future information
Weather forecast Known (forecast) Temp, precip, wind, humidity, cloud, lags Medium Strong Use archived forecasts for evaluation
Weather observations Known only in the past Same variables High if used for future Strong in training Oracle vs forecast gap must be reported
Lagged footfall Known in the past Lags, rolling means, differences High if lag < horizon Strong at short horizons Enforce (t-\ell \le t_0)
Neighboring sensors Known in the past; current for nowcast Spatial lags, graph aggregates High if future values used Moderate For forecast, only past values
POI density Static snapshot Counts/densities by category at 100/250/500 m Low–medium Spatial explanation Use as-of snapshot; future openings leak
Transit schedule / disruptions Known (schedule) / known when announced Frequency, station proximity, disruption flag Medium Moderate GTFS; real-time alerts only if available
Paydays / month end Known Binary None Weak–moderate Country-specific
Tourism season Partially known Seasonal index / school break Low Moderate Confounded with weather in 5 months
Air quality / noise Forecast/observed Continuous Medium Weak Only if data available

Rule. For every feature, record available_at and source in the feature store. A feature is admissible for a forecast issued at (t_0) only if available_at <= t0. This one column prevents most leakage.

3.2 Weather

What the evidence says. Weather affects pedestrian activity, but the magnitude and shape are context-dependent:

  • Rain is usually the largest negative driver; cloud cover, wind, and extreme temperatures also reduce volumes (Attaset et al. 2010).
  • Temperature, humidity, wind, and direct/lagged precipitation matter; effects are stronger on weekends than weekdays and vary by season (Miranda-Moreno & Lahti 2013).
  • A 5°C temperature increase was associated with a 14% increase in pedestrians across nine cities; a shift from snow to dry conditions with a 23% increase; a 5% increase in sunlit area with a 2% increase (de Montigny et al. 2012).
  • Temperature and light cloud cover are positively associated with hourly walking trips; effects vary by season and trip purpose; once a trip starts, weather has little effect on duration (Vanky et al. 2017).
  • Aultman-Hall et al. (2009) find seasonal and weather effects on pedestrian traffic volumes, with heterogeneity across locations.

These are associations from observational studies, not universal causal coefficients. Do not import a literature coefficient into the model; estimate the effect from the local data with a hierarchical prior.

Encoding.

  • Temperature: spline or piecewise linear with a comfort plateau; include a heat/cold extreme indicator. A linear term is often adequate in a narrow 5-month range, but the plateau matters at both ends.
  • Precipitation: amount (mm) + binary indicator + lagged "wet ground" term. Rain timing errors in forecasts are more damaging than temperature errors.
  • Wind: continuous + extreme indicator; interactions with temperature (wind chill).
  • Humidity/cloud: continuous; cloud can be a proxy for pleasant conditions.
  • Snow: binary/amount; important in November–December.
  • Derived comfort index: apparent temperature or a nonlinear combination; only if it has a physical rationale.

Interactions. Rain × weekend, temperature × season, weather × hour-of-day. With five months, estimate only a few domain-motivated interactions; a full tensor product will overfit. Use a hierarchical prior or a GBM to learn interactions implicitly.

Forecast vs observed. This is the most important methodological point in the covariates section.

  • Training on observed weather teaches the model the effect of actual weather.
  • Serving with forecast weather introduces covariate error. Temperature forecasts are relatively accurate at 1–3 days; precipitation timing is not. Footfall is sensitive to rain, so a mis-timed rain forecast can produce a large error.
  • Evaluating on observed weather is an oracle upper bound, not a deployable number. Report it as such.
  • The honest protocol: use archived forecast weather at the same issue time as the forecast (e.g., the forecast valid for (t+h) issued at (t_0)). If an archive is unavailable, simulate forecast error with a calibrated noise model, or report a "climatology" lower bound and an "oracle" upper bound.
  • For 7-day horizons, weather forecast skill decays; do not claim weather-driven accuracy beyond ~3–5 days. Use climate normals or scenario bands for the tail.

The project currently trains on observed weather and uses forecast/current weather in the dashboard and scenarios. [PROJECT] That is acceptable for a demo if the paper says so explicitly and reports the oracle caveat. For the research contribution, add a forecast-weather evaluation or a forecast-error sensitivity analysis.

Scenario analysis. A scenario weather input is a conditional forecast: "if it is 4°C warmer and dry, the model predicts X." It is not a causal effect of weather. Show the scenario within the training support; tree models cannot extrapolate a +10°C winter scenario reliably.

3.3 Holidays, school calendars, and events

Holidays. Official public holidays are known in advance and safe to use. Encode:

  • binary is_holiday;
  • holiday type (public, religious, school, bridge day);
  • proximity (day before/after, long weekend);
  • interaction with day-of-week (a Thursday holiday behaves differently from a Sunday holiday);
  • a hierarchical effect if there are enough holidays; otherwise a single pooled effect.

With five months, the holiday sample is small. Do not fit a separate dummy for every holiday and then interpret each coefficient. Report the holiday effect as a pooled association with a wide interval. The project found that a holiday-only feature added little in the 28-day test [PROJECT]; that is a legitimate result and should be reported honestly.

School calendar. Term dates, breaks, exam periods, and school hours affect weekday daytime footfall, especially near schools and family-oriented retail. Encode school_in_session, school_break, and optionally exam_period. These are known in advance and low-risk. In Germany/Baden-Württemberg, school break dates are published; use the official calendar.

Events. Scheduled events (markets, concerts, sports, conventions) can dominate local footfall. The current project does not yet have a structured event table; event features require collection with an available_at timestamp (announcement time), and if events are unavailable the paper should say so and treat event effects as a limitation rather than inventing them. [PROJECT] Encode:

  • event type and expected attendance (log scale);
  • venue location and distance to each sensor (decaying spatial kernel);
  • time window (start, end, lead/lag);
  • ticket sales or announcement reach if available;
  • interaction with weather (an outdoor market is rain-sensitive).

With few events, use a hierarchical prior or a simple multiplicative uplift with a wide uncertainty band. Do not fit a separate free coefficient per event. For unscheduled events (protests, accidents, flash crowds), the event is not known at (t_0); include a "surprise" flag only at evaluation, never as a training feature that would be unavailable at serving. Event detection from social media or news can provide a nowcast signal, but it is noisy and has its own latency (Sakaki et al. 2010).

Causal caution. Events are not randomized. A post-event lift may be confounded by weather, day-of-week, season, or a concurrent promotion. If the paper wants an event effect, use a difference-in-differences or synthetic control around the event with a clean comparison window, and report the assumptions. With five months and few events, this is exploratory, not definitive.

3.4 POI and transit

POI. Point-of-interest densities within 100/250/500 m capture the attractor mix: retail, food, offices, schools, transport, culture, parks. POI-based functional regions are a standard way to explain urban activity patterns (Yuan et al. 2012). They explain cross-sectional differences between sensors and can improve a global model's spatial generalization. Use as-of snapshots; a POI that opened after the training period is leakage. Avoid categories that are proxies for protected attributes (e.g., places of worship, health services) unless there is a clear planning rationale and a fairness review.

Transit. Station proximity, entry/exit volumes, bus frequency, and service disruptions affect footfall. Scheduled service is known in advance; real-time disruptions are known only when announced. Encode distance to the nearest stop, station entries (if available), and a disruption flag with the announcement time. Transit data can also help transfer across cities: a station-adjacent sensor in one city has an analogous counterpart in another.

POI/transit and spatial confounding. POI density is spatially smooth and correlated with the spatial random effect. If both are included, the POI coefficient can be absorbed by the spatial field (Hodges & Reich 2010). Fix the spatial basis or report the sensitivity. Do not interpret the POI coefficient as a causal "attractor effect" without a design.

3.5 Encoding recipe

A practical feature set for the global GBM:

  • Calendar: hour-of-week (categorical or cyclic), day-of-week, month, is_weekend, is_holiday, holiday type, day-before/after holiday, school term/break, payday, day-of-year (for seasonal trend, with caution).
  • Weather: temperature (spline), precipitation (amount + indicator), wind, humidity, cloud, snow, lagged precipitation (1–3 h), forecast horizon.
  • Events: event count, max attendance (log), distance-weighted attendance, event type, active window.
  • Spatial: sensor ID (categorical), POI densities at 100/250/500 m, distance to transit, street-segment centrality, log catchment area.
  • Lags: footfall at (t-1, t-2, t-4, t-96, t-672) (if available at issue time), rolling means (1 h, 3 h, 1 day, 1 week), same-hour-yesterday/last-week, deviation from the weekly profile.
  • Neighbors (optional): past footfall at neighboring sensors, graph-lagged average; never future values.
  • Missingness: flags for sensor outage, weather missing, event unknown; the model should learn to degrade gracefully.

Scaling. Fit scalers/imputers on the training fold only. For tree models, scaling is not required; for neural models, use robust scaling and clip outliers. Do not per-sensor z-score the target for a heatmap model unless you invert it carefully.

3.6 Leakage risks and rules

Leakage is the dominant threat to credibility in small spatiotemporal studies. Kapoor & Narayanan (2023) document leakage across 17 fields and 294 papers; Kaufman et al. (2012) give a taxonomy. The urban-sensing-specific risks are:

  1. Target leakage. Using future footfall (lags shorter than the horizon, rolling windows that include the forecast period, or a target-derived feature computed over the full period).
  2. Temporal covariate leakage. Using observed weather, observed event impacts, or future POI/transit states at forecast time.
  3. Spatial leakage. Building a graph or a spatial feature from correlations computed over the full period, including the held-out sensor; or using the held-out sensor's values in a spatial lag. This inflates LOSO performance.
  4. Group leakage. The same event or holiday appearing in both train and test; the same day appearing in overlapping windows.
  5. Normalization leakage. Fitting a scaler, imputer, or target encoder on the full dataset.
  6. Pretraining leakage. A foundation model trained on the test period or the same dataset; check the pretraining corpus and use a non-leaking benchmark where possible.
  7. Model-selection leakage. Tuning hyperparameters on the test fold, or reporting the best of many seeds.
  8. Scenario leakage. Using the realized weather/event as the scenario input and calling it a forecast.

Rules.

  • Every feature has an available_at timestamp; assert available_at <= t0 for every training row.
  • Lags must satisfy (t-\ell \le t_0); for direct (h)-step forecasts, the most recent target lag is (t_0), not (t_0+h).
  • Fit all preprocessing inside the training fold; use a pipeline.
  • Build spatial graphs from external geometry or training-only data; never from full-period target correlations.
  • Use a gap of at least the maximum horizon between train/validation and test.
  • Evaluate weather with archived forecasts; if unavailable, report the oracle explicitly.
  • Scheduled events only; unscheduled events are evaluation surprises.
  • Check foundation-model pretraining overlap; report it.
  • Keep a model info sheet with data sources, features, splits, and known leakage risks (Kapoor & Narayanan 2023).

3.7 Causal caution

Weather, holidays, and events are observational. The model estimates conditional associations; it does not identify causal effects. Three implications:

  • Language. Say "associated with," "conditional forecast," or "scenario," not "causes" or "uplift," unless a causal design supports it.
  • Scenario interpretation. A scenario is a conditional prediction under an assumed input, not an intervention effect. Show the assumption and a sensitivity range.
  • Policy use. Staffing decisions can use predictive associations; they should not claim that changing the weather or holding an event causes the predicted change. For event ROI, use a separate causal study.

If the paper wants a causal claim, the minimum design is a difference-in-differences or synthetic control with a clean event window, a comparable control location, and a pre-trend check. With five months and few events, report it as exploratory.


4. Transfer/foundation models

4.1 Four transfer axes

Axis Question Protocol Main risk
Sensor Can the model predict a held-out sensor? LOSO Spatial leakage; sensor-specific bias
Location Can it transfer to another cluster/city? Leave-one-cluster-out Covariate shift; different POI/transit
Task Can a shared representation serve forecast, heatmap, staffing? Multi-task ablation Negative transfer; label mismatch
Time Can it adapt to a new period/regime? Rolling origin; post-shock test Distribution shift; concept drift

Sensor transfer. A global model with a sensor embedding and partial pooling is the simplest form of transfer. Train on 7 sensors, predict the 8th; compare to a per-sensor model. With 8 sensors, LOSO is noisy, so report per-sensor results and the distribution, not just the mean.

Location transfer. Train on one city/cluster, test on another. Use relative indices (devices per baseline) to reduce scale differences; include location-level covariates (POI, transit, climate) to explain heterogeneity. Report the degradation relative to in-domain training. A model that only works in one cluster is still useful for the paper, but the generalization claim must be qualified.

Task transfer. Forecast, heatmap, and staffing share a representation of demand, but they have different losses and validation. A shared encoder can help if the tasks are correlated and the data are sufficient; it can hurt if the tasks conflict (negative transfer). Use task-specific heads and uncertainty weighting (Kendall et al. 2018) or gradient surgery; compare to separate models. With 8 sensors, separate simple models are a strong baseline.

Time transfer. Rolling-origin evaluation is the minimal test. For a true distribution shift (e.g., a new store, a post-holiday regime), use adaptive conformal inference or periodic retraining with a validation gate.

4.2 Global models and partial pooling

A global (panel) model fits one function to all sensors, with sensor identity as a feature or a random effect. It pools the calendar/weather response and borrows strength across sensors. Montero-Manso & Hyndman (2021) show that global models are not more restrictive than local models and can generalize better when series are related. DeepAR is a neural example: one model trained on many related series with a shared likelihood (Salinas et al. 2020).

With 8 sensors, the benefit is real but modest. The hierarchy is:

  1. Complete pooling: one model, no sensor effect. Biased if sensors differ.
  2. Global with sensor intercept: one model, sensor ID as a categorical feature. Simple and strong.
  3. Partial pooling: hierarchical model with sensor intercepts drawn from a common distribution. Better uncertainty, especially for short-history sensors.
  4. Per-sensor models: flexible but data-hungry; only if the sensors are very different and each has enough data.

Recommendation. Use global with sensor intercepts (or partial pooling) as the default; test per-sensor models as a baseline. Report the pooling gain.

4.3 Foundation models for time series

Model Architecture / pretraining Covariates Output Access Reported evidence Caveat
Chronos (Ansari et al. 2024) T5 encoder-decoder, 20M–710M; tokenized values; 42 datasets + synthetic GPs Univariate; no exogenous Probabilistic (quantile samples) Open weights Zero-shot comparable to supervised on 42 datasets No covariates; 15-min high-frequency not the focus
TimesFM (Das et al. 2024) Decoder-only patched Transformer, 200M; ~100B timepoints Univariate; no exogenous Point/quantile Open weights Zero-shot close to supervised on diverse benchmarks No covariates; long-horizon focus
Moirai (Woo et al. 2024) Masked encoder, any-variate attention, mixture distributions; LOTSA 27B observations Multivariate; limited exogenous Probabilistic Open weights Competitive/superior zero-shot vs full-shot Any-variate ≠ causal covariates; needs care
TimeGPT (Garza & Mergenthaler-Canseco 2023) Proprietary Transformer API API-dependent Probabilistic Commercial API Zero-shot claims; limited public detail Data governance; reproducibility; cost
Lag-Llama (Rasul et al. 2023) Decoder-only, lag features; open Univariate Probabilistic Open weights Competitive zero-shot on benchmarks No covariates; smaller community
MOMENT (Goswami et al. 2024) Encoder, multi-task pretraining; open Univariate Forecasting/classification/imputation Open weights Strong limited-supervision results Not a covariate-rich forecaster
TabPFN-TS (Hoo et al. 2025) Tabular PFN on timestamp-derived features Timestamp features Probabilistic Open Matches/outperforms Chronos-Large on 24 datasets Not a temporal FM; small-data tabular prior
TimesFM-ICF (Das et al. 2025) In-context fine-tuning of TimesFM Univariate Probabilistic Research code Few-shot improves over base; matches per-dataset fine-tuning Still univariate; needs related series

4.4 What the evidence actually supports

The original papers are positive:

  • Chronos reports zero-shot performance "comparable and occasionally superior" to models trained on the target dataset across 42 datasets (Ansari et al. 2024).
  • TimesFM reports zero-shot performance close to state-of-the-art supervised models on a diverse benchmark (Das et al. 2024).
  • Moirai reports competitive or superior zero-shot performance vs full-shot models on LOTSA (Woo et al. 2024).

Independent and adversarial evaluations are more mixed:

  • GIFT-Eval (23 datasets, 144k series, 177M points) finds foundation models strong overall, but weak in transport and high-frequency, noisy, lumpy domains; fine-tuning helps long horizons but a gap remains (Aksu et al. 2024).
  • Cloud data: zero-shot FMs are consistently beaten by an online linear model and a seasonal naive baseline; some produce erratic forecasts (Toner et al. 2025).
  • Spectral shift: FMs degrade when the dominant frequency components of the target are underrepresented in pretraining (Wang et al. 2025).
  • Realistic evaluation: with careful leakage control and strong statistical baselines, statistical models and ensembles outperform foundation models by more than 8.2% sCRPS and 20% MASE across 15 large forecast competitions; synthetic pretraining improves an FFM by about 7% (Olivares et al. 2025).
  • Specialized FMs: across genomics, satellite imaging, and time series, simple tuned supervised models match or outperform foundation models; the authors recommend strong baselines and automated supervised workflows (Xu et al. 2025).
  • Not universally negative: some industrial benchmarks find FMs better than traditional baselines, especially compact Chronos variants (Montet et al. 2025). The evidence is domain-dependent, not a verdict for or against.

Synthesis. Foundation models are a genuine advance for zero-shot forecasting on broad, stationary, univariate benchmarks. They are not a substitute for a covariate-aware, spatially structured model on a small, high-frequency, event-driven urban task. The correct posture is: benchmark them, report them, and use them where they add value (cold start, ensemble diversity), not as the default.

4.5 Caveats specific to this task

  1. No exogenous covariates. Most open FMs forecast a univariate series from its own history. They cannot use weather, holidays, or events, which are the main drivers here. A model that ignores rain will lose on rainy days regardless of pretraining.
  2. No cross-sensor pooling. Channel-independent FMs treat each sensor separately; they cannot borrow strength across the 8 sensors or enforce spatial coherence. Moirai's any-variate attention can accept multiple series, but it is not a spatial model and does not know the street network.
  3. High frequency. 15-minute data are noisier and more autocorrelated than the hourly/daily series that dominate pretraining. GIFT-Eval finds FMs weaker on high-frequency, noisy data.
  4. Short history. Five months is enough for zero-shot evaluation but short for fine-tuning a large model. Fine-tuning may overfit; a small adapter or in-context fine-tuning is safer.
  5. Pretraining overlap. The test period or the dataset may appear in pretraining. Check the model card and the pretraining corpus; use GIFT-Eval's non-leaking pretraining data as a reference. Report the risk even if you cannot rule it out.
  6. Count data and calibration. FMs output continuous quantiles; footfall is a non-negative, overdispersed count. Calibrate and clip; check coverage.
  7. Cost, latency, and governance. API models (TimeGPT) send data to a third party; latency may exceed the <2 s serving target; version drift breaks reproducibility. Open weights are preferable for a public-sector paper.
  8. No spatial uncertainty. A univariate predictive interval at a sensor is not a heatmap interval. Do not use FM quantiles as spatial uncertainty.
  • Benchmark, don't replace. Run 2–3 open models (Chronos-Bolt/Chronos, TimesFM, Moirai) zero-shot on the same rolling-origin folds. Report MAE/WAPE/CRPS and coverage. This is a cheap, credible "state of the art" comparison.
  • Fine-tune one model. Fine-tune the best open model on the target series (or use in-context fine-tuning) and compare to zero-shot and to the GBM. Report the compute and the data fraction.
  • Ensemble only if it helps. Add the FM to an ensemble and test whether CRPS improves; do not include it for novelty.
  • Cold start. Use an FM to forecast a new sensor or a new city with little history; compare to a seasonal naive and a pooled GBM. This is the most defensible transfer use case.
  • Be explicit about covariates. State that the FM is a univariate history-only benchmark; the GBM/TFT use weather/holidays/events. The comparison is then fair in coverage, not in information.
  • Governance. Prefer open weights; do not send raw device data to a commercial API; report the model version and date.

4.7 Shared encoder + task heads: when it is justified

A shared spatiotemporal encoder with task heads is the natural architecture when:

  • there are many sensors (roughly 50–100+) and/or multiple cities;
  • there is a large unlabeled corpus for masked pretraining;
  • multiple tasks (forecast, heatmap, anomaly, staffing) share a representation;
  • the encoder can be pretrained on masked reconstruction and fine-tuned per task;
  • there is enough compute and engineering time to tune it.

Multi-task learning can help when tasks share structure, but negative transfer is real and must be tested against separate models (Caruana 1997; Ruder 2017; Kendall et al. 2018). At 8 sensors and 5 months, the encoder has too little data to learn a general representation, and the leakage surface (masking, graph construction, normalization) grows. A statistical shared encoder — spatial basis functions (\phi_k(s)) plus temporal factors (a_k(t)) — is a better fit: it shares structure across sensors, has few parameters, gives uncertainty, and is interpretable. It is also a natural bridge to a neural encoder if the data grow: replace the basis/factor model with a TCN/Transformer encoder and keep the same task heads.


5. Evaluation

5.1 Splits: time, space, events

A credible evaluation for this dataset needs at least three split types.

Chronological train/validation/test with a gap.

[train ................][val ....][gap h][test ....]

The gap must be at least the maximum forecast horizon (e.g., 7 days for a 7-day horizon) so that no training window overlaps the test period. Fit scalers, imputers, and feature selectors inside the training fold.

Rolling origin (walk-forward).

fold 1: [train ......][val][gap][test]
fold 2: [train .........][val][gap][test]
fold 3: [train ............][val][gap][test]
...

With five months, use 4–8 folds of 2–4 weeks, expanding or sliding windows. Report mean ± SD across folds. A single 28-day holdout is not enough: it may contain a holiday, a weather anomaly, or a sensor outage, and it gives no estimate of variability.

Leave-one-sensor-out (LOSO). For each sensor (i), train on the other seven and predict (i). This is the key test for the heatmap claim. Report per-sensor metrics and the distribution, not just the mean. LOSO is also the right test for spatial transfer; random spatial CV would leak.

Leave-one-event-out / leave-one-holiday-out. For rare drivers, hold out one event/holiday at a time and test whether the model generalizes. With few events, this is exploratory; report the uncertainty.

Nested CV for hyperparameters. Tune on the inner folds, evaluate on the outer test; do not tune on the test.

Why not random K-fold? Random splits leak across time (adjacent 15-min windows are nearly identical) and across space (nearby sensors share unobserved drivers). Random CV produces optimistic error estimates; block CV is nearly always more appropriate for structured data (Roberts et al. 2017; Bergmeir & Benítez 2012).

5.2 Metrics

Metric Definition / idea Captures When to use Pitfall
MAE (\frac{1}{n}\sum y-\hat y ) Absolute error in device units
RMSE (\sqrt{\frac{1}{n}\sum(y-\hat y)^2}) Large errors Model comparison Sensitive to outliers
WAPE (\frac{\sum y-\hat y }{\sum y
MASE MAE scaled by the in-sample seasonal-naive MAE Skill vs baseline Cross-series comparison Baseline choice matters
sMAPE Symmetric percentage error Relative error Reporting Unstable near zero
Poisson deviance (2\sum[y\log(y/\hat y)-(y-\hat y)]) Count likelihood fit Count models Requires (\hat y>0)
CRPS (\int (F(x)-\mathbf{1}{y\le x})^2 dx) Full predictive distribution Probabilistic comparison Computationally heavier
Pinball loss ((\tau-\mathbf{1}{y<\hat q})(y-\hat q)) Quantile accuracy Quantile models Not comparable across (\tau) without averaging
Coverage Share of observations inside the interval Calibration Intervals Can be gamed with wide intervals
Interval width Mean width of the interval Sharpness With coverage Narrower is better only if coverage holds
PIT histogram Distribution of (F(y)) Calibration Diagnostics Needs enough observations
Cost Expected under/over-staffing cost Decision value Staffing Requires cost assumptions

Recommendation. Report MAE, WAPE, and MASE for point accuracy; CRPS and pinball for probabilistic accuracy; coverage and width for calibration; and a cost curve for decisions (Hyndman & Koehler 2006; Koenker & Bassett 1978; Matheson & Winkler 1976; Gneiting & Raftery 2007). Do not report only RMSE or only MAPE.

5.3 Probabilistic forecasting and calibration

Quantile loss. For a target quantile (\tau),

[ \rho_\tau(y,\hat q) = \begin{cases} \tau(y-\hat q), & y \ge \hat q,\ (1-\tau)(\hat q-y), & y < \hat q. \end{cases} ]

Training separate quantile models and sorting the outputs is a simple, strong probabilistic baseline. Enforce non-crossing quantiles by rearrangement.

CRPS. A proper score that rewards calibration and sharpness:

[ \text{CRPS}(F,y) = \int_{-\infty}^{\infty}\big(F(x)-\mathbf{1}{y\le x}\big)^2 dx. ]

It is strictly proper; lower is better. For samples, use the empirical CDF or the sorted-sample estimator. Report CRPS per horizon and per sensor.

Calibration diagnostics. PIT histogram, reliability diagram for exceedance probabilities, coverage by horizon and sensor, and interval width (Dawid 1984; Gneiting et al. 2007; Gneiting & Katzfuss 2014). A model can have good CRPS but poor tail calibration; check both.

Conformal prediction. Split conformal gives distribution-free marginal coverage under exchangeability:

[ \hat C(x) = [\hat f(x) - \hat Q_{1-\alpha}(S),; \hat f(x) + \hat Q_{1-\alpha}(S)], ]

where (S) are calibration residuals. For time series, exchangeability is violated; use blocked/online conformal, ACI, or EnbPI (Vovk et al. 2005; Angelopoulos & Bates 2023; Stankevičiūtė et al. 2021; Gibbs & Candès 2021; Xu & Xie 2021; Zaffran et al. 2022). Conformal intervals are attractive for small data because they require no distributional assumption and can wrap any point model. They are not a substitute for a spatial model: they calibrate the sensor-level residual, not the heatmap.

ACI. Adaptive conformal inference updates the miscoverage level online:

[ \alpha_{t+1} = \alpha_t + \gamma(\alpha - \mathbf{1}{y_t \notin \hat C_t}), ]

which maintains long-run coverage under distribution shift (Gibbs & Candès 2021). Use it for monitoring and for the deployed forecast; report the learning rate and the coverage trajectory.

Bayesian posterior predictive. If the hierarchical model is fit with MCMC/INLA, the posterior predictive distribution is the natural probabilistic output. It propagates parameter uncertainty and spatial interpolation uncertainty; it is more coherent than conformal for the heatmap, but computationally heavier. Use it for the spatial layer, and conformal/quantile for the fast sensor-level layer if needed.

Calibration under LOSO. Check coverage at held-out sensors, not only at observed sensors. A model can be calibrated in-sample and overconfident at a new location.

5.4 Decision-level evaluation

A forecast is only useful if it changes a decision. For staffing, the decision is the number of staff (S) for a period with demand (D). With underage cost (c_u) (cost per unit short) and overage cost (c_o) (cost per unit surplus), the expected cost is

[ \mathbb{E}[C(S)] = c_u,\mathbb{E}[(D-S)^+] + c_o,\mathbb{E}[(S-D)^+]. ]

The optimal (S) is the (c_u/(c_u+c_o)) quantile of the predictive distribution. This is the newsvendor result; it connects the forecast quantile to the decision. If understaffing is worse than overstaffing, the optimal quantile is above 0.5.

Cost-loss / value. For a binary decision (e.g., open an overflow queue), the cost-loss ratio determines the optimal probability threshold. The economic value of a forecast is the cost reduction relative to a naive policy, normalized by the value of a perfect forecast (Richardson 2000; Zhu et al. 2002; Wilks 2001). Report a cost curve over decision thresholds, not a single number.

Decision metrics for the paper. The cost parameters are assumptions, not measured labor costs; the project has no POS, sales, queue, or labor-cost data. [PROJECT] Show the assumptions and avoid ROI claims.

  • expected staffing cost per day under the forecast policy vs a flat schedule vs a seasonal-naive policy;
  • coverage: share of hours with enough staff;
  • over/understaffed hours;
  • cost curve: cost vs safety factor / quantile;
  • sensitivity to the cost ratio and to the visitors-per-staff-hour assumption.

Why this matters. A model can improve WAPE and still produce a worse staffing plan if its errors are in the wrong tail. Decision-level evaluation is the honest bridge between forecasting and the product. It also makes the paper's contribution distinctive: most urban forecasting papers stop at MAE.

5.5 Baselines, seeds, and significance

Baselines. Report at least: seasonal naive, daily profile, ETS, SARIMA, Prophet, Poisson/NB GLM, global HistGB. Tune each with the same budget and the same splits. Report the baseline table in the appendix.

Seeds. For stochastic models (neural nets, NGBoost, random forest), run at least 5 seeds (ideally 10) and report mean ± SD. For deterministic models, the fold-to-fold variation is the uncertainty. Do not report the best seed.

Significance.

  • Diebold–Mariano for pairwise forecast comparison (Diebold & Mariano 1995), with a Newey–West/HAC variance and the Harvey et al. (1997) small-sample correction. Use the loss differential series, not the raw errors.
  • Wilcoxon signed-rank across folds/sensors for a nonparametric paired test.
  • Nadeau–Bengio corrected resampled t-test for CV comparisons, because repeated folds are not independent (Nadeau & Bengio 2003).
  • Demšar tests for multiple model comparisons (Friedman + post-hoc Nemenyi) if you compare many models (Demšar 2006).
  • Report effect sizes (WAPE percentage points, cost euros) alongside p-values. With 8 sensors and 4–8 folds, statistical significance is hard; do not overinterpret a p=0.05.

Multiple comparisons. If you test many models/horizons, correct or pre-register the primary comparison. The primary claim should be a single pre-specified comparison (e.g., weather vs no weather at 1 h), not the best of many.

5.6 Leakage checklist

Use this as a pre-submission audit:

  • Every feature has an available_at timestamp and it is (\le t_0).
  • Lags satisfy (t-\ell \le t_0); no rolling window includes the forecast period.
  • Train/validation/test are chronological with a gap (\ge) max horizon.
  • Scalers/imputers/encoders are fit inside the training fold.
  • The spatial graph is built from external geometry or training-only data.
  • LOSO does not use the held-out sensor in any feature, graph, or normalization.
  • Weather evaluation uses archived forecasts, or the oracle is labeled.
  • Events are scheduled/announced before (t_0); unscheduled events are evaluation-only.
  • POI/transit snapshots are as-of (t_0).
  • Foundation-model pretraining overlap is checked/reported.
  • Hyperparameters are tuned on validation, not test.
  • All seeds/folds are reported, not the best.
  • Scenario inputs are labeled conditional, not causal.
  • The model info sheet documents sources, features, splits, and risks.

5.7 Small-data pitfalls

  1. Random CV. Leaks across time and space; use blocked/rolling CV.
  2. Overlapping windows. A 7-day horizon with daily origins creates near-duplicate test samples; use a gap or non-overlapping origins.
  3. Too many hyperparameters. With 8 sensors, tuning 50 configurations on 4 folds overfits the validation; use a small search space and nested CV.
  4. Best-seed reporting. Report all seeds and the distribution.
  5. Strawman baselines. A seasonal naive with the wrong seasonality is not a baseline; tune it.
  6. Annual seasonality claims. Five months cannot identify annual seasonality; do not claim it.
  7. City-wide claims. One cluster is not a city; do not extrapolate the heatmap.
  8. Causal claims from events. Few events, confounded; report associations.
  9. Ignoring sensor outages. Gaps up to 6 days [PROJECT] can bias training; model missingness or weight by coverage.
  10. Ignoring device penetration. A model of devices is not a model of people; say so.

5.8 Reporting template

A compact results table for the paper:

Model Horizon MAE WAPE MASE CRPS 80% coverage Cost (€/day)
Seasonal naive 1 h 1.00
Poisson GLM 1 h
Global HistGB 1 h
Global quantile HistGB + conformal 1 h
TFT (if run) 1 h
Chronos/TimesFM zero-shot 1 h
... 48 h / 7 d

Supplementary: per-sensor LOSO table, calibration figure (PIT/reliability), cost curve, ablation table (weather/holiday/event/POI/spatial pooling), and a coverage-vs-distance figure for the heatmap.

Ablations to report.

  • Weather: none vs temperature vs temperature + rain vs full weather.
  • Holiday: none vs binary vs type/proximity.
  • Event: none vs scheduled event features.
  • POI: none vs static POI.
  • Spatial: per-sensor vs global vs partial pooling; with/without spatial residual field.
  • Probabilistic: point vs quantile vs quantile + conformal.
  • Reconciliation: independent vs MinT.
  • Model class: GBM vs TFT/DeepAR vs FM.

Each ablation should change one thing and report the same metrics, with a significance test for the primary comparison.


6.1 Design principles

  1. Separate estimands. Sensor demand, latent heatmap, forecast, and decision each get their own model, validation, and language. A single end-to-end model can still be used, but its outputs must be evaluated separately.
  2. Structured priors over capacity. With 8 sensors and 5 months, a low-rank spatial field, partial pooling, and a regularized tabular model are more defensible than a high-capacity deep network. Complexity must earn its place on rolling-origin and LOSO evidence.
  3. Leakage-safe by construction. Every feature carries an available_at timestamp; every split has a gap; every graph is built from external geometry or training-only data.
  4. Probabilistic by default. Point forecasts are a derived product. The system produces quantiles or samples, calibrates them, and evaluates coverage and sharpness.
  5. Decision-aware. The forecast is connected to a cost function. The quantile used for staffing is chosen from the underage/overage cost ratio, not from a default 0.5.
  6. Modular and replaceable. The spatial model, forecast model, reconciliation, and decision layer communicate through stable interfaces. A better model can be swapped in without rewriting the pipeline.
  7. Reproducible and auditable. Versioned data, features, models, and forecasts; deterministic seeds; a model card; a leakage checklist; a cached fallback.

6.2 Reference pipeline

┌──────────────────────────────────────────────────────────────────────────┐
│ L0  Data contract & QC                                                   │
│ raw counts · sensor metadata · weather observations/forecasts · holidays  │
│ school calendar · events · POI · transit · UTC + local time · provenance  │
│ missingness flags · privacy flags · data version                          │
└──────────────────────────────────┬───────────────────────────────────────┘
                                   │
┌──────────────────────────────────▼───────────────────────────────────────┐
│ L1  Feature store (versioned, available_at enforced)                     │
│ calendar · weather forecast · holiday · school · event · POI · transit    │
│ lags/rolling · neighbor lags · spatial features · missingness             │
└───────────────┬──────────────────────────────────────┬───────────────────┘
                │                                      │
┌───────────────▼────────────────────┐   ┌─────────────▼───────────────────┐
│ L2  Spatial state model            │   │ L3  Temporal forecast model     │
│ low-rank NB + Matérn/network basis │   │ global quantile HistGB          │
│ posterior field + 80/95% intervals │   │ direct multi-horizon + conformal│
│ reliability mask                   │   │ per-horizon calibration         │
└───────────────┬────────────────────┘   └─────────────┬───────────────────┘
                │                                      │
                └──────────────────┬───────────────────┘
                                   │
┌──────────────────────────────────▼───────────────────────────────────────┐
│ L4  Reconciliation & coherence                                           │
│ MinT across sensors/cluster · non-negativity · sample reconciliation      │
└──────────────────────────────────┬───────────────────────────────────────┘
                                   │
┌──────────────────────────────────▼───────────────────────────────────────┐
│ L5  Decision layer                                                       │
│ staffing quantile · cost-loss · scenario overrides · assumptions          │
│ sensitivity · no causal claims                                           │
└──────────────────────────────────┬───────────────────────────────────────┘
                                   │
┌──────────────────────────────────▼───────────────────────────────────────┐
│ L6  Evaluation harness (offline)                                         │
│ rolling origin · LOSO · seeds · baselines · DM · calibration · cost       │
└──────────────────────────────────┬───────────────────────────────────────┘
                                   │
┌──────────────────────────────────▼───────────────────────────────────────┐
│ L7  Serving & monitoring (online)                                        │
│ batch forecasts · cached fallback · model/data version · coverage/drift   │
└──────────────────────────────────────────────────────────────────────────┘

The L2/L3 split is logical, not necessarily two independent fits. In the recommended factor model, L2 estimates the spatial basis and the observation model, while L3 forecasts the temporal factors; they can be fitted jointly (Bayesian) or iteratively (basis first, then factors). The interface is the latent factor state, which keeps the pipeline modular.

6.3 Component specification

Layer Input Output Method Validation Fallback
L0 Data Raw CSV/API streams Clean panel + metadata Schema validation, UTC/local conversion, dedup, gap flags Row counts, coverage per sensor, outlier checks Reject bad batch; use last good
L1 Features Clean panel + external data Feature matrix with available_at Calendar, weather, holidays, events, POI, lags; as-of joins Leakage assertions; train/serve parity Climatology/median for missing covariates
L2 Spatial Sensor counts + features Posterior heatmap + intervals Low-rank NB + Matérn/network basis; or regression kriging LOSO error, coverage, CRPS, reliability mask IDW/OK baseline; mask unreliable cells
L3 Forecast Feature matrix Quantile forecasts per sensor/horizon Global HistGB (quantile) + conformal; optional TFT/DeepAR Rolling origin, multi-seed, DM, calibration Seasonal naive; cached forecast
L4 Reconciliation Sensor + cluster forecasts Coherent forecasts MinT; non-negative; sample reconciliation Coherence error; CRPS before/after Independent forecasts + warning
L5 Decision Predictive distribution Staffing plan + cost Quantile from cost ratio; shift heuristic Cost curve; coverage; sensitivity Flat schedule; conservative quantile
L6 Evaluation All model outputs Tables, figures, model card Rolling origin, LOSO, seeds, tests Reproducibility checklist —
L7 Serving Baked artifacts JSON/SSE/UI FastAPI + cache + Go fallback Latency, availability, drift Baked forecast; last good

6.4 Shared encoder + task heads vs a simpler hierarchical model

The central architectural choice is whether to build one shared spatiotemporal encoder with task heads, or a simpler hierarchical/panel model. The honest answer is scale-dependent.

Data regime Recommended core Why What to benchmark
8 sensors, 5 months, one cluster (this project) L2 low-rank Bayesian/GP + L3 global quantile GBM Data-efficient; calibrated UQ; interpretable; low leakage surface TFT/DeepAR, Chronos/TimesFM, STGNN ablation
20–50 sensors, 1–2 years, one city Global TCN/Transformer encoder + task heads Enough data to learn shared temporal features; cross-sensor pooling GBM, TFT, per-sensor models
100+ sensors, multiple cities/tasks Pretrained masked spatiotemporal encoder + task heads; FM initialization Transfer and scale; unlabeled pretraining Strong GBM, per-city fine-tune, FM

Why not a shared deep encoder at 8 sensors.

  • The effective sample is small: ~150 days and 8 series, with strong autocorrelation; a deep encoder has more parameters than independent observations.
  • Tabular/tree models remain state of the art on medium-sized data; deep learning's advantage is not established (Grinsztajn et al. 2022; Shwartz-Ziv & Armon 2022).
  • M4/M5 top results are dominated by hybrids/combinations and strong gradient boosting, not monolithic deep nets (Makridakis et al. 2020, 2022).
  • Spatiotemporal GNNs are designed for many nodes; with 8 nodes the graph is nearly complete, and simple baselines are often competitive (Shchur et al. 2018; Corradini et al. 2026).
  • Foundation models are mixed on high-frequency, lumpy, transport-like data and usually lack exogenous covariates (Aksu et al. 2024; Toner et al. 2025; Olivares et al. 2025).
  • A deep encoder adds leakage surface (masking, graph construction, normalization) and makes calibrated spatial uncertainty harder.

What the "shared encoder + heads" idea gets right. It recognizes that forecast, heatmap, and decision share a representation. The statistical version of this is a low-rank spatiotemporal factor model:

[ u(s,t) = \sum_{k=1}^{K}\phi_k(s),a_k(t), \qquad K \approx 3\text{–}6, ]

where (\phi_k(s)) are Matérn/RBF basis functions or graph-diffusion modes, and (a_k(t)) are temporal factors. The observation model is a negative binomial with sensor-specific offsets. The task heads are:

  • Forecast head: predict (a_k(t+h)) with a multivariate model using calendar/weather/event covariates; map to sensors and cells.
  • Heatmap head: evaluate (u(s,t)) on the grid; aggregate to cells; compute posterior intervals.
  • Decision head: consume the predictive distribution; choose the staffing quantile from the cost ratio.

This shares structure across sensors, has few parameters, gives coherent maps, and is interpretable. It is the right "shared encoder" for small data. If the project grows to many sensors, replace the basis/factor model with a TCN/Transformer encoder and keep the same heads and interfaces.

What would change our mind. A small TCN or TFT beats the GBM on rolling-origin and LOSO with 5 seeds and a significant DM test; or a multi-city dataset becomes available; or the task expands to 50+ sensors. Until then, the simpler model is the defensible choice.

6.5 Concrete reference stack

Need Suggested tool Notes
Data wrangling polars/pandas, GeoPandas, OSMnx, H3 UTC + local; network graph; hex/grid support
Spatial model PyMC, Stan, R-INLA, scikit-learn GP INLA for SPDE/BYM2; PyMC/Stan for custom NB models
GAM / baselines mgcv (R), statsmodels Cyclic splines; Poisson/NB GLM
Gradient boosting LightGBM, HistGradientBoostingRegressor Quantile objective; categorical support
Conformal mapie, custom split conformal Blocked/ACI for time series
Neural benchmarks neuralforecast, pytorch-forecasting TFT, DeepAR, N-HiTS, PatchTST
Foundation models chronos-forecasting, timesfm, uni2ts Open weights; zero-shot/fine-tune
Evaluation Custom rolling-origin/LOSO; scipy DM test; coverage; cost curve
Serving FastAPI + joblib + cached JSON Matches the project's Go/FastAPI split [PROJECT]

6.6 What the paper can claim

Claim What would support it Current status How to phrase
"Eight sensors support a ~150 m heatmap" Geometry + LOSO + interval width Design calculation [ESTIMATE]; needs LOSO "We report 100–150 m cells; finer grids are visualization, not inference."
"Weather improves accuracy" Ablation with time-based test Project WAPE 27.8 vs 44.7 [PROJECT] "With observed weather, WAPE improved from 44.7% to 27.8%; forecast-weather evaluation is the deployable number."
"Holiday effect is small" Holiday ablation Project 28-day test [PROJECT] "The holiday-only feature added little in the 28-day test."
"Global model beats per-sensor" Rolling origin + LOSO + seeds To run "The global model improved LOSO WAPE by X points (p=...)."
"GBM beats TFT/FM" Same splits + DM + seeds To run Report either direction; a negative result is useful.
"Foundation model is competitive" Zero-shot benchmark To run "Zero-shot Chronos/TimesFM was competitive on MAE but did not use weather/holidays."
"City-wide heatmap" Dense city sensors Not supported Do not claim.
"Causal event uplift" Clean event study Not supported Do not claim; call it a conditional scenario.
"People counts" Penetration/calibration study Not supported Say "estimated devices."
"Annual seasonality" 12+ months Not supported Say "five months; no annual cycle."

6.7 Minimum experiment plan

A realistic plan for a hackathon/paper, in priority order:

Step Deliverable Time Gate
1. Geometry + variogram + LOSO harness Sensor map, spacing, variogram, LOSO splitter 0.5 day LOSO runs end-to-end
2. Baselines + global HistGB + quantile/conformal Baseline table, quantile forecasts, coverage 1 day Beats seasonal naive on WAPE
3. Spatial state model + heatmap Posterior map + intervals + reliability mask 1 day LOSO coverage near nominal
4. Reconciliation + decision cost Coherent forecasts, cost curve 0.5 day Cost improvement vs flat schedule
5. Foundation model + TFT/DeepAR benchmark Zero-shot/fine-tuned table 1 day Same folds; no leakage
6. Ablations + figures + model card Paper-ready tables/figures, limitations 1 day Every number traceable

If time is short, steps 1–4 are the paper's core; step 5 is a differentiator; step 6 is non-negotiable for credibility.

6.8 Failure modes and fallbacks

Failure Symptom Mitigation Fallback
Sensor outage/gap Missing series, biased map Missingness flags; partial pooling; gap-aware likelihood Drop sensor from map; widen intervals
Weather forecast error Rain mis-timed; large errors Forecast-weather evaluation; scenario bands Climatology; wider intervals
Event surprise Sudden spike unexplained Surprise flag at evaluation; monitoring Wider intervals; human review
Distribution shift Coverage drops over time ACI; drift monitoring; retrain gate Conservative quantile; cached model
Sparse spatial support Wide intervals between sensors Reliability mask; aggregate cells 200–250 m map
Model overfit Good train, poor LOSO Strong baseline; regularization; seeds GBM/seasonal naive
Latency >2 s response Precompute; cache; batch Baked forecast [PROJECT]
Privacy Small-cell disclosure Suppression/DP; aggregation Publish relative index only
Data version mismatch Inconsistent features Versioned artifacts; assertions Last good version

6.9 Generalization to other urban sensing tasks

The architecture generalizes because the interfaces are task-agnostic, even though the models are not:

  • Likelihood: counts (NB/Poisson) for footfall, waste fill events, parking arrivals; continuous (Gaussian/quantile) for air quality, noise, energy; binary (Bernoulli) for occupancy; compositional for mode share.
  • Spatial support: grid for air quality/heat; network for footfall/traffic; polygon for districts; station for transit.
  • Covariates: weather, calendar, events, POI, transit, land use — the same feature store with different sources.
  • Decision: staffing, collection routing, enforcement, energy dispatch, transit frequency. The cost function changes; the quantile/cost-loss logic remains.
  • Evaluation: rolling origin, LOSO/leave-one-site-out, probabilistic calibration, decision cost. The same harness.

The transferable claim is therefore architectural: the same interfaces — data contract, feature store, spatial state model, forecast model, reconciliation, decision, and evaluation — are designed to generalize across sparse urban sensor streams, even when the likelihood, support, covariates, and decision cost change. That is stronger and more honest than "our model generalizes."

6.10 What would make this architecture wrong

  • The cluster is actually dense and 2D, with many more sensors than the 8 in this dataset; then a fine-grid GP/GNN may be justified.
  • The decision only needs the store-level sensor, and the heatmap is decorative; then skip L2 and invest in L3/L5.
  • Events are frequent, scheduled, and well-labeled; then a richer event model and a causal design may pay off.
  • The deployment needs sub-second cell-level updates; then precompute and cache, or simplify the spatial model.
  • A large multi-city dataset becomes available; then a shared neural encoder/foundation model becomes worth testing.

Until one of these is true, the simpler hierarchical/GBM architecture is the right way.


7. Citations

A. Spatial statistics, geostatistics, and interpolation

  1. Tobler, W. R. (1970). A computer movie simulating urban growth in the Detroit region. Economic Geography, 46(sup1), 234–240. https://doi.org/10.2307/143141
  2. Cressie, N. A. C. (1993). Statistics for Spatial Data (rev. ed.). Wiley.
  3. Diggle, P. J., & Ribeiro, P. J. (2007). Model-based Geostatistics. Springer. https://doi.org/10.1007/978-0-387-48536-2
  4. Banerjee, S., Carlin, B. P., & Gelfand, A. E. (2014). Hierarchical Modeling and Analysis for Spatial Data (2nd ed.). CRC Press. https://doi.org/10.1201/b16018
  5. Rasmussen, C. E., & Williams, C. K. I. (2006). Gaussian Processes for Machine Learning. MIT Press.
  6. Rue, H., & Held, L. (2005). Gaussian Markov Random Fields: Theory and Applications. Chapman & Hall/CRC. https://doi.org/10.1201/9780203492024
  7. Rue, H., Martino, S., & Chopin, N. (2009). Approximate Bayesian inference for latent Gaussian models by using integrated nested Laplace approximations. Journal of the Royal Statistical Society: Series B, 71(2), 319–392. https://doi.org/10.1111/j.1467-9868.2008.00700.x
  8. Lindgren, F., Rue, H., & Lindström, J. (2011). An explicit link between Gaussian fields and Gaussian Markov random fields: the stochastic partial differential equation approach. Journal of the Royal Statistical Society: Series B, 73(4), 423–498. https://doi.org/10.1111/j.1467-9868.2011.00777.x
  9. Besag, J., York, J., & Mollié, A. (1991). Bayesian image restoration, with two applications in spatial statistics. Annals of the Institute of Statistical Mathematics, 43(1), 1–20. https://doi.org/10.1007/BF00116466
  10. Riebler, A., Sørbye, S. H., Simpson, D., & Rue, H. (2016). An intuitive Bayesian spatial model for disease mapping that accounts for scaling. Statistical Methods in Medical Research, 25(4), 1145–1165. https://doi.org/10.1177/0962280216660421
  11. Simpson, D., Rue, H., Riebler, A., Martins, T. G., & Sørbye, S. H. (2017). Penalising model component complexity: a principled, practical approach to constructing priors. Statistical Science, 32(1), 1–28. https://doi.org/10.1214/16-STS576
  12. Knorr-Held, L. (2000). Bayesian modelling of inseparable space-time variation in disease risk. Statistics in Medicine, 19(17–18), 2555–2567. https://doi.org/10.1002/1097-0258(20000915/30)19:17/18<2555::AID-SIM587>3.0.CO;2-#
  13. Li, J., & Heap, A. D. (2011). A review of comparative studies of spatial interpolation methods in environmental sciences: performance and impact factors. Ecological Informatics, 6(3–4), 228–241. https://doi.org/10.1016/j.ecoinf.2010.12.003
  14. Li, J., & Heap, A. D. (2014). Spatial interpolation methods applied in the environmental sciences: a review. Environmental Modelling & Software, 53, 173–189. https://doi.org/10.1016/j.envsoft.2013.12.008
  15. Webster, R., & Oliver, M. A. (2007). Geostatistics for Environmental Scientists (2nd ed.). Wiley. https://doi.org/10.1002/9780470517277
  16. Wahba, G. (1990). Spline Models for Observational Data. SIAM. https://doi.org/10.1137/1.9781611970128
  17. Wood, S. N. (2017). Generalized Additive Models: An Introduction with R (2nd ed.). CRC Press. https://doi.org/10.1201/9781315370279
  18. Hengl, T., Nussbaum, M., Wright, M. N., Heuvelink, G. B. M., & Gräler, B. (2018). Random forest as a generic framework for predictive modeling of spatial and spatio-temporal variables. PeerJ, 6, e5518. https://doi.org/10.7717/peerj.5518
  19. Goodchild, M. F., & Lam, N. S.-N. (1980). Areal interpolation: a variant of the traditional spatial problem. Geo-Processing, 1, 297–312.
  20. Gotway, C. A., & Young, L. J. (2002). Combining incompatible spatial data. Journal of the American Statistical Association, 97(458), 632–648. https://doi.org/10.1198/016214502760046953
  21. Openshaw, S. (1984). The Modifiable Areal Unit Problem. CATMOG 38. Geo Books.
  22. Hodges, J. S., & Reich, B. J. (2010). Adding spatially-correlated errors can mess up the fixed effect you love. The American Statistician, 64(4), 325–334. https://doi.org/10.1198/tast.2010.10052
  23. Krause, A., Singh, A., & Guestrin, C. (2008). Near-optimal sensor placements in Gaussian processes: theory, efficient algorithms and empirical studies. Journal of Machine Learning Research, 9, 235–284.
  24. Shuman, D. I., Narang, S. K., Frossard, P., Ortega, A., & Vandergheynst, P. (2013). The emerging field of signal processing on graphs. IEEE Signal Processing Magazine, 30(3), 83–98. https://doi.org/10.1109/MSP.2012.2235192
  25. Ortega, A., Frossard, P., Kovačević, J., Moura, J. M. F., & Vandergheynst, P. (2018). Graph signal processing: overview, challenges, and applications. Proceedings of the IEEE, 106(5), 808–828. https://doi.org/10.1109/JPROC.2018.2820126
  26. Wu, Y., Zhuang, D., Labbe, A., & Sun, L. (2021). Inductive graph neural networks for spatiotemporal kriging. Proceedings of the AAAI Conference on Artificial Intelligence, 35(5), 4478–4485. https://doi.org/10.1609/aaai.v35i5.16575

B. Heatmaps, data fusion, privacy, and urban data

  1. Castanedo, F. (2013). A review of data fusion techniques. The Scientific World Journal, 2013, 704504. https://doi.org/10.1155/2013/704504
  2. Zheng, Y., Capra, L., Wolfson, O., & Yang, H. (2014). Urban computing: concepts, methodologies, and applications. ACM Transactions on Intelligent Systems and Technology, 5(3), 38. https://doi.org/10.1145/2629592
  3. Blondel, V. D., Decuyper, A., & Krings, G. (2015). A survey of results on mobile phone datasets analysis. EPJ Data Science, 4, 10. https://doi.org/10.1140/epjds/s13688-015-0046-0
  4. Carroll, R. J., Ruppert, D., Stefanski, L. A., & Crainiceanu, C. M. (2006). Measurement Error in Nonlinear Models: A Modern Perspective (2nd ed.). CRC Press. https://doi.org/10.1201/9781420010138
  5. Sweeney, L. (2002). k-anonymity: a model for protecting privacy. International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems, 10(5), 557–570. https://doi.org/10.1142/S0218488502001648
  6. Machanavajjhala, A., Kifer, D., Gehrke, J., & Venkitasubramaniam, M. (2007). l-diversity: privacy beyond k-anonymity. ACM Transactions on Knowledge Discovery from Data, 1(1), 3. https://doi.org/10.1145/1217299.1217302
  7. Dwork, C. (2006). Differential privacy. In Automata, Languages and Programming (ICALP 2006), LNCS 4052, 1–12. https://doi.org/10.1007/11787006_1
  8. Andrés, M. E., Bordenabe, N. E., Chatzikokolakis, K., & Palamidessi, C. (2013). Geo-indistinguishability: differential privacy for location-based systems. In Proceedings of the 2013 ACM SIGSAC Conference on Computer & Communications Security (CCS), 901–914. https://doi.org/10.1145/2508859.2516735
  9. Aktay, A., Bavadekar, S., Cossoul, G., Davis, J., Desfontaines, D., Fabrikant, A., et al. (2020). Google COVID-19 Community Mobility Reports: anonymization process description (version 1.1). arXiv:2004.04145.
  10. Yuan, J., Zheng, Y., & Xie, X. (2012). Discovering regions of different functions in a city using human mobility and POIs. In Proceedings of the 18th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 186–194. https://doi.org/10.1145/2339530.2339561
  11. Sakaki, T., Okazaki, M., & Matsuo, Y. (2010). Earthquake shakes Twitter users: real-time event detection by social sensors. In Proceedings of the 19th International Conference on World Wide Web (WWW), 851–860. https://doi.org/10.1145/1772690.1772777

C. Forecasting models

  1. Hyndman, R. J., & Athanasopoulos, G. (2021). Forecasting: Principles and Practice (3rd ed.). OTexts.
  2. Box, G. E. P., Jenkins, G. M., Reinsel, G. C., & Ljung, G. M. (2015). Time Series Analysis: Forecasting and Control (5th ed.). Wiley.
  3. Taylor, S. J., & Letham, B. (2018). Forecasting at scale. The American Statistician, 72(1), 37–45. https://doi.org/10.1080/00031305.2017.1380080
  4. Friedman, J. H. (2001). Greedy function approximation: a gradient boosting machine. Annals of Statistics, 29(5), 1189–1232. https://doi.org/10.1214/aos/1013203451
  5. Chen, T., & Guestrin, C. (2016). XGBoost: a scalable tree boosting system. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 785–794. https://doi.org/10.1145/2939672.2939785
  6. Ke, G., Meng, Q., Finley, T., Wang, T., Chen, W., Ma, W., Ye, Q., & Liu, T.-Y. (2017). LightGBM: a highly efficient gradient boosting decision tree. In Advances in Neural Information Processing Systems 30, 3146–3154.
  7. Duan, T., Anand, A., Ding, D. Y., Thai, K. K., Basu, S., Ng, A., & Schuler, A. (2020). NGBoost: natural gradient boosting for probabilistic prediction. In Proceedings of the 37th International Conference on Machine Learning, PMLR 119, 2690–2700.
  8. Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory. Neural Computation, 9(8), 1735–1780. https://doi.org/10.1162/neco.1997.9.8.1735
  9. Salinas, D., Flunkert, V., Gasthaus, J., & Januschowski, T. (2020). DeepAR: probabilistic forecasting with autoregressive recurrent networks. International Journal of Forecasting, 36(3), 1181–1191. https://doi.org/10.1016/j.ijforecast.2019.07.001
  10. Bai, S., Kolter, J. Z., & Koltun, V. (2018). An empirical evaluation of generic convolutional and recurrent networks for sequence modeling. arXiv:1803.01271.
  11. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., & Polosukhin, I. (2017). Attention is all you need. In Advances in Neural Information Processing Systems 30, 5998–6008.
  12. Zhou, H., Zhang, S., Peng, J., Zhang, S., Li, J., Xiong, H., & Zhang, W. (2021). Informer: beyond efficient transformer for long sequence time-series forecasting. Proceedings of the AAAI Conference on Artificial Intelligence, 35(12), 11106–11115. https://doi.org/10.1609/aaai.v35i12.17325
  13. Zeng, A., Chen, M., Zhang, L., & Xu, Q. (2023). Are transformers effective for time series forecasting? Proceedings of the AAAI Conference on Artificial Intelligence, 37(9), 11121–11128. https://doi.org/10.1609/aaai.v37i9.26317
  14. Nie, Y., Nguyen, N., Sinthong, P., & Kalagnanam, J. (2023). A time series is worth 64 words: long-term forecasting with transformers. In International Conference on Learning Representations (ICLR).
  15. Oreshkin, B. N., Carpov, D., Chapados, N., & Bengio, Y. (2020). N-BEATS: neural basis expansion analysis for interpretable time series forecasting. In International Conference on Learning Representations (ICLR).
  16. Challu, C., Olivares, K. G., Oreshkin, B. N., Garza, F., Mergenthaler-Canseco, M., & Dubrawski, A. (2023). NHITS: neural hierarchical interpolation for time series forecasting. Proceedings of the AAAI Conference on Artificial Intelligence, 37(6), 6989–6997. https://doi.org/10.1609/aaai.v37i6.25854
  17. Lim, B., Arık, S. Ö., Loeff, N., & Pfister, T. (2021). Temporal Fusion Transformers for interpretable multi-horizon time series forecasting. International Journal of Forecasting, 37(4), 1748–1764. https://doi.org/10.1016/j.ijforecast.2021.03.012
  18. Li, Y., Yu, R., Shahabi, C., & Liu, Y. (2018). Diffusion convolutional recurrent neural network: data-driven traffic forecasting. In International Conference on Learning Representations (ICLR).
  19. Yu, B., Yin, H., & Zhu, Z. (2018). Spatio-temporal graph convolutional networks: a deep learning framework for traffic forecasting. In Proceedings of the 27th International Joint Conference on Artificial Intelligence (IJCAI), 3634–3640. https://doi.org/10.24963/ijcai.2018/505
  20. Wu, Z., Pan, S., Long, G., Jiang, J., & Zhang, C. (2019). Graph WaveNet for deep spatial-temporal graph modeling. In Proceedings of the 28th International Joint Conference on Artificial Intelligence (IJCAI), 1907–1913. https://doi.org/10.24963/ijcai.2019/264
  21. Guo, S., Lin, Y., Feng, N., Song, C., & Wan, H. (2019). Attention based spatial-temporal graph convolutional networks for traffic flow forecasting. Proceedings of the AAAI Conference on Artificial Intelligence, 33(01), 922–929. https://doi.org/10.1609/aaai.v33i01.3301922
  22. Rasul, K., Seward, C., Schuster, I., & Vollgraf, R. (2021). Autoregressive denoising diffusion models for multivariate probabilistic time series forecasting. In Proceedings of the 38th International Conference on Machine Learning, PMLR 139, 8857–8868.
  23. Tashiro, Y., Song, J., Song, Y., & Ermon, S. (2021). CSDI: conditional score-based diffusion models for probabilistic time series imputation. In Advances in Neural Information Processing Systems 34.
  24. Makridakis, S., Spiliotis, E., & Assimakopoulos, V. (2020). The M4 competition: 100,000 time series and 61 forecasting methods. International Journal of Forecasting, 36(1), 54–74. https://doi.org/10.1016/j.ijforecast.2019.04.014
  25. Makridakis, S., Spiliotis, E., & Assimakopoulos, V. (2022). M5 accuracy competition: results, findings, and conclusions. International Journal of Forecasting, 38(4), 1346–1364. https://doi.org/10.1016/j.ijforecast.2021.11.013
  26. Grinsztajn, L., Oyallon, E., & Varoquaux, G. (2022). Why do tree-based models still outperform deep learning on typical tabular data? In Advances in Neural Information Processing Systems 35: Datasets and Benchmarks Track.
  27. Shwartz-Ziv, R., & Armon, A. (2022). Tabular data: deep learning is not all you need. Information Fusion, 81, 84–90. https://doi.org/10.1016/j.inffus.2021.11.011
  28. Montero-Manso, P., & Hyndman, R. J. (2021). Principles and algorithms for forecasting groups of time series: locality and globality. International Journal of Forecasting, 37(4), 1632–1653. https://doi.org/10.1016/j.ijforecast.2021.03.004
  29. Hyndman, R. J., Ahmed, R. A., Athanasopoulos, G., & Shang, H. L. (2011). Optimal combination forecasts for hierarchical time series. Computational Statistics & Data Analysis, 55(9), 2579–2589. https://doi.org/10.1016/j.csda.2011.03.006
  30. Wickramasuriya, S. L., Athanasopoulos, G., & Hyndman, R. J. (2019). Optimal forecast reconciliation for hierarchical and grouped time series through trace minimization. Journal of the American Statistical Association, 114(526), 804–819. https://doi.org/10.1080/01621459.2018.1448825
  31. Shchur, O., Mumme, M., Bojchevski, A., & Günnemann, S. (2018). Pitfalls of graph neural network evaluation. arXiv:1811.05868.
  32. Corradini, F., Gerosa, F., Gori, M., Lucheroni, C., Piangerelli, M., & Zannotti, M. (2026). A systematic literature review of spatio-temporal graph neural network models for time series forecasting and classification. Neural Networks, 195, 108269. https://doi.org/10.1016/j.neunet.2025.108269

D. Covariates and urban context

  1. Attaset, V., Schneider, R. J., Arnold, L. S., & Ragland, D. R. (2010). Effects of weather variables on pedestrian volumes in Alameda County, California. Institute of Transportation Studies, UC Berkeley. https://escholarship.org/uc/item/3zn9f4cr
  2. Miranda-Moreno, L. F., & Lahti, A. C. (2013). Temporal trends and the effect of weather on pedestrian volumes: a case study of Montreal, Canada. Transportation Research Part D: Transport and Environment, 22, 54–59. https://doi.org/10.1016/j.trd.2013.02.008
  3. de Montigny, L., Ling, R., & Zacharias, J. (2012). The effects of weather on walking rates in nine cities. Environment and Behavior, 44(6), 821–840. https://doi.org/10.1177/0013916511409033
  4. Vanky, A. P., Verma, S. K., Courtney, T. K., Santi, P., & Ratti, C. (2017). Effect of weather on pedestrian trip count and duration: city-scale evaluations using mobile phone application data. Preventive Medicine Reports, 8, 30–37. https://doi.org/10.1016/j.pmedr.2017.07.002
  5. Aultman-Hall, L., Lane, D., & Lambert, R. R. (2009). Assessing impact of weather and season on pedestrian traffic volumes. Transportation Research Record, 2140(1), 35–43. https://doi.org/10.3141/2140-04
  6. Giannone, D., Reichlin, L., & Small, D. (2008). Nowcasting: the real-time informational content of macroeconomic data. Journal of Monetary Economics, 55(4), 665–676. https://doi.org/10.1016/j.jmoneco.2008.05.010
  7. Evensen, G. (2003). The ensemble Kalman filter: theoretical formulation and practical implementation. Ocean Dynamics, 53(4), 343–367. https://doi.org/10.1007/s10236-003-0036-9
  8. Carrassi, A., Bocquet, M., Bertino, L., & Evensen, G. (2018). Data assimilation in the geosciences: an overview of methods, issues, and perspectives. WIREs Climate Change, 9(5), e535. https://doi.org/10.1002/wcc.535

E. Transfer and foundation models

  1. Ansari, A. F., Stella, L., Turkmen, A. C., Zhang, X., Mercado, P., Shen, H., et al. (2024). Chronos: learning the language of time series. Transactions on Machine Learning Research. arXiv:2403.07815.
  2. Das, A., Kong, W., Sen, R., & Zhou, Y. (2024). A decoder-only foundation model for time-series forecasting. In Proceedings of the 41st International Conference on Machine Learning, PMLR 235, 10148–10167.
  3. Woo, G., Liu, C., Kumar, A., Xiong, C., Savarese, S., & Sahoo, D. (2024). Unified training of universal time series forecasting transformers. In Proceedings of the 41st International Conference on Machine Learning, PMLR 235, 53140–53164.
  4. Garza, A., & Mergenthaler-Canseco, M. (2023). TimeGPT-1. arXiv:2310.03589.
  5. Rasul, K., Ashok, A., Williams, A. R., Ghonia, H., Bhagwatkar, R., Khorasani, A., et al. (2023). Lag-Llama: towards foundation models for time series forecasting. arXiv:2310.08278.
  6. Goswami, M., Szafer, K., Choudhry, A., Cai, Y., Li, S., & Dubrawski, A. (2024). MOMENT: a family of open time-series foundation models. In Proceedings of the 41st International Conference on Machine Learning.
  7. Hoo, S. B., Müller, S., Salinas, D., & Hutter, F. (2025). The tabular foundation model TabPFN outperforms specialized time series forecasting models based on simple features. arXiv:2501.02945.
  8. Das, A., Faw, M., Sen, R., & Zhou, Y. (2025). In-context fine-tuning for time-series foundation models. In Proceedings of the 42nd International Conference on Machine Learning. arXiv:2410.24087.
  9. Aksu, T., Woo, G., Liu, J., Liu, X., Liu, C., Savarese, S., Xiong, C., & Sahoo, D. (2024). GIFT-Eval: a benchmark for general time series forecasting model evaluation. arXiv:2410.10393.
  10. Toner, W., Lee, T. L., Joosen, A., Singh, R., & Asenov, M. (2025). Performance of zero-shot time series foundation models on cloud data. ICLR 2025 Workshop on "I Can't Believe It's Not Better". arXiv:2502.12944.
  11. Wang, T., Ennadir, S., Pertoft, J., Gandler, G. Z., Cao, L., Senane, Z., Katsarou, S., Asadi, S., Karlsson, A., & Smirnov, O. (2025). Frequency matters: when time series foundation models fail under spectral shift. NeurIPS 2025 Workshop on Recent Advances in Time Series Foundation Models. arXiv:2511.05619.
  12. Olivares, K. G., Wolff, M., Konstantinova, T., Ramasubramanian, S., Oreshkin, B., Wilson, A. G., et al. (2025). A more realistic evaluation of cross-frequency transfer learning and foundation forecasting models. NeurIPS 2025 Workshop on Recent Advances in Time Series Foundation Models. arXiv:2509.19465.
  13. Xu, Z., Gupta, R., Cheng, W., Shen, A., Shen, J., Talwalkar, A., & Khodak, M. (2025). Specialized foundation models struggle to beat supervised baselines. In International Conference on Learning Representations (ICLR). arXiv:2411.02796.
  14. Montet, F., Pasquier, B., & Wolf, B. (2025). Benchmarking foundation models for time-series forecasting: zero-shot, few-shot, and full-shot evaluations. Computer Sciences & Mathematics Forum, 11(1), 32. https://doi.org/10.3390/cmsf2025011032
  15. Kendall, A., Gal, Y., & Cipolla, R. (2018). Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 7482–7491. https://doi.org/10.1109/CVPR.2018.00781
  16. Caruana, R. (1997). Multitask learning. Machine Learning, 28(1), 41–75. https://doi.org/10.1023/A:1007379606734
  17. Ruder, S. (2017). An overview of multi-task learning in deep neural networks. arXiv:1706.05098.

F. Probabilistic forecasting and decision evaluation

  1. Koenker, R., & Bassett, G. (1978). Regression quantiles. Econometrica, 46(1), 33–50. https://doi.org/10.2307/1913643
  2. Gneiting, T., & Raftery, A. E. (2007). Strictly proper scoring rules, prediction, and estimation. Journal of the American Statistical Association, 102(477), 359–378. https://doi.org/10.1198/016214506000001437
  3. Gneiting, T., Balabdaoui, F., & Raftery, A. E. (2007). Probabilistic forecasts, calibration and sharpness. Journal of the Royal Statistical Society: Series B, 69(2), 243–268. https://doi.org/10.1111/j.1467-9868.2007.00587.x
  4. Gneiting, T., & Katzfuss, M. (2014). Probabilistic forecasting. Annual Review of Statistics and Its Application, 1, 125–151. https://doi.org/10.1146/annurev-statistics-062713-085831
  5. Matheson, J. E., & Winkler, R. L. (1976). Scoring rules for continuous probability distributions. Management Science, 22(10), 1087–1096. https://doi.org/10.1287/mnsc.22.10.1087
  6. Dawid, A. P. (1984). Statistical theory: the prequential approach. Journal of the Royal Statistical Society: Series A, 147(2), 278–292. https://doi.org/10.2307/2981683
  7. Vovk, V., Gammerman, A., & Shafer, G. (2005). Algorithmic Learning in a Random World. Springer.
  8. Angelopoulos, A. N., & Bates, S. (2023). Conformal prediction: a gentle introduction. Foundations and Trends in Machine Learning, 16(4), 494–591. https://doi.org/10.1561/2200000101
  9. Stankevičiūtė, K., Alaa, A. M., & van der Schaar, M. (2021). Conformal time-series forecasting. In Advances in Neural Information Processing Systems 34.
  10. Gibbs, I., & Candès, E. (2021). Adaptive conformal inference under distribution shift. In Advances in Neural Information Processing Systems 34, 1660–1672.
  11. Xu, C., & Xie, Y. (2021). Conformal prediction interval for dynamic time-series. In Proceedings of the 38th International Conference on Machine Learning, PMLR 139, 11559–11569.
  12. Zaffran, M., Féron, O., Goude, Y., Josse, J., & Dieuleveut, A. (2022). Adaptive conformal predictions for time series. In Proceedings of the 39th International Conference on Machine Learning, PMLR 162, 25834–25866.
  13. Richardson, D. S. (2000). Skill and relative economic value of the ECMWF ensemble prediction system. Quarterly Journal of the Royal Meteorological Society, 126(563), 649–667. https://doi.org/10.1002/qj.49712656313
  14. Zhu, Y., Toth, Z., Wobus, R., Richardson, D., & Mylne, K. (2002). The economic value of ensemble-based weather forecasts. Bulletin of the American Meteorological Society, 83(1), 73–83. https://doi.org/10.1175/1520-0477(2002)083<0073:TEVOEB>2.3.CO;2
  15. Wilks, D. S. (2001). A skill score based on economic value for probability forecasts. Meteorological Applications, 8(2), 209–219. https://doi.org/10.1017/S1350482701002092

G. Evaluation methodology

  1. Hyndman, R. J., & Koehler, A. B. (2006). Another look at measures of forecast accuracy. International Journal of Forecasting, 22(4), 679–688. https://doi.org/10.1016/j.ijforecast.2006.03.001
  2. Bergmeir, C., & Benítez, J. M. (2012). On the use of cross-validation for time series predictor evaluation. Information Sciences, 191, 192–213. https://doi.org/10.1016/j.ins.2011.12.028
  3. Roberts, D. R., Bahn, V., Ciuti, S., Boyce, M. S., Elith, J., Guillera-Arroita, G., et al. (2017). Cross-validation strategies for data with temporal, spatial, hierarchical, or phylogenetic structure. Ecography, 40(8), 913–929. https://doi.org/10.1111/ecog.02881
  4. Diebold, F. X., & Mariano, R. S. (1995). Comparing predictive accuracy. Journal of Business & Economic Statistics, 13(3), 253–263. https://doi.org/10.1080/07350015.1995.10524599
  5. Harvey, D., Leybourne, S., & Newbold, P. (1997). Testing the equality of prediction mean squared errors. International Journal of Forecasting, 13(2), 281–291. https://doi.org/10.1016/S0169-2070(96)00719-4
  6. Demšar, J. (2006). Statistical comparisons of classifiers over multiple data sets. Journal of Machine Learning Research, 7, 1–30.
  7. Nadeau, C., & Bengio, Y. (2003). Inference for the generalization error. Machine Learning, 52(3), 239–281. https://doi.org/10.1023/A:1024068626366
  8. Kapoor, S., & Narayanan, A. (2023). Leakage and the reproducibility crisis in machine-learning-based science. Patterns, 4(9), 100804. https://doi.org/10.1016/j.patter.2023.100804
  9. Kaufman, S., Rosset, S., Perlich, C., & Stitelman, O. (2012). Leakage in data mining: formulation, detection, and avoidance. ACM Transactions on Knowledge Discovery from Data, 6(4), 15. https://doi.org/10.1145/2382577.2382579
  10. Makridakis, S., Spiliotis, E., & Assimakopoulos, V. (2018). Statistical and machine learning forecasting methods: concerns and ways forward. PLoS ONE, 13(3), e0194889. https://doi.org/10.1371/journal.pone.0194889