PaperlayerView exact revision on arXiv

Source-linked AI summary

How Good Are Time-Series Foundation Models for Pedestrian Crowd Count Forecasting? A Cross-Dataset Comparative Study

Theivaprakasham Hari, Ziteng Li, Yanan Xin, Winnie Daamen, Serge Hoogendoorn

arXiv:2609.16415v1cs.LGcs.AI

TL;DR

The study asks whether time-series foundation models reliably outperform classical methods in pedestrian sensing, where event deployments provide little history and data regimes differ. It benchmarks seven univariate methods across event-driven and seasonal datasets, finding regime-dependent performance: long-context foundation models excel at seasonal long horizons, while simpler methods remain competitive with limited event history.

  • Problem

    Evidence remains limited on whether time-series foundation models reliably improve pedestrian-count forecasting across data regimes, despite the importance of forecasting for pedestrian operations.

    Source: §I · p. 1

  • Method

    The study benchmarks seven univariate forecasting methods across limited-history event data and multi-year seasonal data.

    Source: §V · p. 7

  • Results

    No model family dominates: Seasonal Naive and boosted trees remain competitive with limited event history, while long-context FMs lead at Melbourne weekly and monthly horizons; extending TimesFM context from 1 week to 16,000 hours reduces MAE about 16–49%.

    Source: §V · p. 8

  • Takeaways & Limitations

    Choose models according to available history and distributional shift rather than defaulting to the most complex architecture.

    Source: §V · p. 8

    The univariate evaluation cannot capture schedule-driven crowd surges; the authors identify exogenous event features as future work.

    Source: §V · p. 8

Abstract

from arXiv · show

Pedestrian-count forecasting supports pedestrian-oriented Intelligent Transportation Systems (ITS), including crowd monitoring, pedestrian-traffic staffing and routing, and proactive risk mitigation during surges. Recent time-series foundation models (FMs) report strong zero-shot accuracy on heterogeneous forecasting benchmarks, but it remains unclear whether these gains transfer reliably to pedestrian sensing deployments. We benchmark seven univariate forecasting approaches spanning four paradigms: Seasonal Naive, gradient-boosted trees (LightGBM, CatBoost), deep learning models (N-HiTS, PatchTST), and two pretrained FMs (TimesFM, Chronos-2). Experiments cover two complementary regimes: (i) a five-day special event dataset SAIL2025 at 3-minute resolution with limited in-domain history; and (ii) Melbourne pedestrian sensors as a multi-year hourly dataset (2010--2017) with strong seasonality. We compare the MAE and RMSE results per sensor across datasets and multiple forecast horizons. Results show three consistent findings. First, with limited historical data, Seasonal Naive remains a strong baseline for long-horizon forecasting on high-volume sensors, while trained models can degrade when the next day differs substantially from prior days. Second, boosted trees can be competitive on lower-volume sensors but exhibit higher sensitivity on high-volume sensors under event-driven shift. Third, FMs excel in the seasonal and data-rich regime under long-context configuration. The findings highlight the importance of choosing pedestrian forecasting models based on both the underlying data conditions and the forecasting horizon.

I. INTRODUCTION

Pedestrian forecasting supports operational planning but faces limited history, day-to-day shift, and short-term volatility. Existing evidence does not establish foundation models’ reliable gains across pedestrian data regimes or their deployment costs.

I. INTRODUCTION

  • The study addresses the lack of a unified, reproducible benchmark comparing forecasting accuracy and deployment considerations across pedestrian data regimes.It evaluates when time-series foundation models are useful and when simpler methods remain preferable.Source: §I · pp. 1–2
  • TimesFM and Chronos-2 are compared with Seasonal Naive, LightGBM, CatBoost, PatchTST, and N-HiTS.Source: §I · p. 2
  • The five-day SAIL2025 dataset at 3-minute resolution supports an expanding walk-forward assessment of zero-shot foundation-model forecasting under limited-history event conditions.Source: §I · p. 2
  • On Melbourne, maximum foundation-model contexts—8,192 tokens for Chronos-2 and 16,000 time steps for TimesFM—are tested for horizons up to 720 hours.Source: §I · p. 2
  • Deployment costs are profiled on identical CPU-only hardware, covering training time, inference latency, memory footprint, and model size.Source: §I · p. 2

II. RELATED WORK

Prior work covers pedestrian-count methods, multi-horizon deep forecasting, and time-series foundation models; the review highlights limited evidence across datasets and deployment regimes.

A. Pedestrian Count Forecasting

  • Pedestrian-count studies often draw conclusions from a single dataset or narrow model set, limiting assessment of whether gains generalize across event-driven and multi-year regimes.Source: A. Pedestrian Count Forecasting · p. 2

B. Multi-Horizon Deep Forecasting Models

  • Multi-horizon architectures use attention, decomposition, patching, or MLP designs, but their benefits remain uncertain under short, shift-heavy event windows.Examples include Temporal Fusion Transformers, Informer, Autoformer, PatchTST, and TiDE [18].Source: B. Multi-Horizon Deep Forecasting Models · p. 2

C. Time-Series Foundation Models

Time-series foundation models use large-scale pretraining to support zero- or few-shot forecasting, but their reliability on pedestrian data remains insufficiently documented. The paper defines a univariate forecasting task and varies context length for foundation models.

C. Time-Series Foundation Models

  • Time-series foundation models learn transferable temporal representations through heterogeneous pretraining and aim to forecast with minimal task-specific adaptation.The review describes TimesFM and Chronos-2 [11] among models whose pedestrian-sensor evaluation remains limited.Source: C. Time-Series Foundation Models · p. 2

III. METHODOLOGY AND EXPERIMENTS

  • The methodology specifies the datasets, forecasting task, model families, and experimental configurations.Source: §III · p. 3

A. Problem Formulation

  • For each sensor, the task uses a context window of L observations to forecast the next H values; foundation-model context length is varied to measure its effect on accuracy.Source: A. Problem Formulation · p. 3

B. Datasets

The study treats each sensor as an independent univariate series and compares two deployment regimes: multi-year seasonal Melbourne data and short event-driven SAIL2025 data.

B. Datasets

  • Each sensor channel is modeled independently, matching standalone sensor operation without requiring spatial metadata.Cross-sensor dependence is outside the study’s scope.Source: B. Datasets · p. 3

1) Melbourne Pedestrian Dataset:

  • Melbourne provides 16 sensors with continuous hourly observations from 2010 to 2017, strong daily and weekly seasonality, and mean counts ranging from 147 at T8 to 1,508 at T4.The data support long-horizon evaluation up to 720 hours; Fig. 1 shows a 14-day hourly sample for sensor T4.Source: 1) Melbourne Pedestrian Dataset · p. 3, Figure 1 · p. 3

2) SAIL2025 Event Dataset:

  • SAIL2025 contains 11 sensors recorded every 3 minutes over five event days, with opening and closing days differing from intermediate days.Fig. 1 shows five consecutive days for sensor GASA-03.Source: 2) SAIL2025 Event Dataset · p. 3, Figure 1 · p. 3

C. Forecasting Models

The section describes the benchmark’s forecasting models, from seasonal repetition to boosted trees and deep-learning baselines. Tree models use a fixed univariate setup, while deep models use standard untuned architectures.

C. Forecasting Models

  • Seasonal Naive repeats the most recent seasonal cycle, using a one-day period for SAIL2025 and a one-week period for Melbourne.The seasonal periods are 480 timestamps and 168 hours, respectively.Source: C. Forecasting Models · p. 3

2) Gradient-Boosted Trees:

  • LightGBM and CatBoost fit an independent regressor for each forecast step using lagged counts, without exogenous or calendar features.Both use the same fixed configuration across datasets; the authors note that errors may be conservative versus dataset-specific tuning.Source: 2) Gradient-Boosted Trees · p. 3

3) Deep Learning Baselines:

  • N-HiTS and PatchTST use standard untuned architectures, with training performed independently per sensor and capped at 500 gradient updates.Training uses early stopping with patience of five epochs and the Adam optimizer.Source: 3) Deep Learning Baselines · pp. 3–4

4) Foundation Models:

TimesFM and Chronos-2 are evaluated without fine-tuning, across event-driven limited-history and seasonal data-rich regimes. The setups vary training history, forecast horizon, and foundation-model context length.

4) Foundation Models:

  • TimesFM and Chronos-2 are evaluated strictly in zero-shot mode, without parameter updates or fine-tuning.Inference runs on a CPU laptop with 16 GB of RAM and no GPU acceleration.Source: 4) Foundation Models · p. 4

D. Experimental Setup 1: Event and limited-history regime (SAIL2025)

  • SAIL2025 uses four expanding walk-forward tests, training on one to four event days and forecasting 2, 6, 12, or 24 hours ahead.Foundation-model contexts are 6 or 24 hours; trained baselines use only data available at each step.Source: D. Experimental Setup 1: Event and… · p. 4

E. Experimental Setup 2: Seasonal and data-rich regime (Melbourne)

  • Melbourne uses chronological 2010–2015 training, 2016 validation, and held-out 2017 testing, with horizons of one day, one week, or one month.Foundation models are compared at 168-hour context and their maximum supported context: 16,000 hours for TimesFM and 8,192 for Chronos-2.Source: E. Experimental Setup 2: Seasonal and… · p. 4

F. Evaluation Metrics

Evaluation reports MAE and RMSE separately for each sensor. The paper also profiles model fitting and serving costs, then presents results by dataset followed by cross-regime discussion.

F. Evaluation Metrics

  • The evaluation reports Mean Absolute Error and Root Mean Squared Error as its accuracy metrics.Sensor-level reporting avoids collapsing sensors with different scales and dynamics into one global score.Source: F. Evaluation Metrics · p. 4

G. Computational Cost Protocol

  • Computational profiling records fitting time, memory, serialized model size, and median single-forecast latency for representative configurations.Latency is measured through each library’s serving interface over 20 origins after two warm-up calls, with each model profiled in a separate process.Source: G. Computational Cost Protocol · p. 4

IV. RESULTS AND DISCUSSION

  • Results are reported for each dataset in turn, followed by a cross-regime comparison and discussion of model selection and deployment implications.Source: §IV · p. 5

A. SAIL2025 Event Dataset

In the limited-history event regime, results vary by sensor volume and horizon: Seasonal Naive is strong for long-horizon high-volume forecasts, while boosted trees lead on several lower-volume results. Qualitative forecasts also show that univariate histories miss an unprecedented event surge, whereas long context helps foundation models recover seasonal patterns in Melbourne.

1) Qualitative Analysis:

  • None of the seven models captures the fireworks-driven late-evening surge on the high-volume SAIL2025 sensor.The forecasts range from Seasonal Naive’s coarse declining trend to over-smoothed FM forecasts; the surge has no training-set precedent.Source: 1) Qualitative Analysis · p. 5, Figure 2 · p. 5

2) Quantitative Results:

  • Seasonal Naive achieves the lowest high-volume SAIL2025 24-hour MAE in E4 at 89.3, while CatBoost leads at all four horizons on the medium-volume sensor.Chronos-2 has the best 2-hour MAE in E1 across high-, medium-, and low-volume sensors: 35.9, 21.5, and 11.2; its advantage diminishes as local data accumulates.Source: 2) Quantitative Results · p. 5

1) Qualitative Analysis:

  • On Melbourne’s high-volume T4 sensor, extending FM context visibly recovers forecast peak magnitudes and daily rhythm relative to short-context forecasts.TimesFM uses 16,000 time steps and Chronos-2 uses 8,192 in the long-context comparison.Source: 1) Qualitative Analysis · p. 5, Figure 3 · p. 5

2) Quantitative Results:

Long context improves foundation-model accuracy in seasonal Melbourne data, while deployment costs vary sharply across models. SAIL results are reported across walk-forward training-data growth, and cost measurements show different trade-offs in latency, memory, and retraining.

C. Computational Cost and Deployment Latency

2) Quantitative Results:

  • Table III tracks SAIL2025 MAE across one-to-four-day training sets with a 6-hour input context; E1 trained baselines appear only at 2- and 6-hour horizons.Zero-shot FM results are also shown at 12 and 24 hours where available.Source: Table III · p. 6
  • TimesFM with 16,000-hour context reduces MAE by approximately 16–49% versus 168 hours and achieves best or second-best MAE at 168- and 720-hour horizons.Long-context Chronos-2 improves less, but wins MAE on medium- and low-volume sensors at daily and weekly horizons.Source: 2) Quantitative Results · p. 5, Table IV · p. 6

C. Computational Cost and Deployment Latency

  • TimesFM takes 910 ms per forecast on SAIL and 913 ms on Melbourne, with 961 MB and 973 MB of inference memory and a 925 MB checkpoint.It has the largest storage footprint among evaluated models; on high-volume SAIL, that expense yields no accuracy benefit because Seasonal Naive has the lowest 24-hour MAE in E4.Source: C. Computational Cost and Deployment… · pp. 6–7
  • LightGBM and CatBoost require 299–1,392 ms per forecast and use up to 19.5 MB on SAIL and 46.7 MB on Melbourne among trained baselines.Their per-horizon regressors make training and inference scale linearly with horizon length.Source: C. Computational Cost and Deployment… · p. 7
  • CatBoost training on Melbourne takes 781 seconds and uses 1.28 GB during serving, constraining concurrent sensors on a 16 GB machine.Zero-shot FMs avoid repeated in-domain training, but not necessarily high memory use.Source: C. Computational Cost and Deployment… · p. 7
  • At studied network sizes, even the slowest models fit within forecast cycles, but retraining costs roughly 675 seconds per sensor and horizon, or about two hours per SAIL walk-forward step.Seasonal Naive and zero-shot FMs avoid this retraining cost; N-HiTS trains more economically than PatchTST despite having more parameters.Source: C. Computational Cost and Deployment… · p. 7

D. Cross-Regime Comparison and Discussion

Model rankings depend on history availability and distributional shift: limited-history event data favors different approaches from data-rich seasonal data. The mechanisms and deployment recommendations vary accordingly.

D. Cross-Regime Comparison and Discussion

  • In event-driven data, rankings vary by sensor and horizon; seasonal data favors long-context FMs at weekly and monthly horizons, while supervised models remain competitive daily.No model family is universally optimal; available history and distributional shift determine which approach is most effective.Source: D. Cross-Regime Comparison and… · p. 7
  • Seasonal Naive is competitive on high-volume SAIL sensors when consecutive event days have similar profiles, while one-to-four days of training may not capture event variance.Trees generalize on stable medium- and low-volume profiles but become unstable when later high-volume surges differ from training data.Source: D. Cross-Regime Comparison and… · p. 7
  • FMs offer cold-start value, but their seasonal-regime advantage depends on long context; the paper recommends matching model choice to history, horizon, and shift.Its proposed tiering uses zero-shot FMs initially, Seasonal Naive for limited-history long horizons, lightweight trees after local data accumulates, and long-context FMs for seasonal planning.Source: D. Cross-Regime Comparison and… · p. 7

V. CONCLUSIONS

The benchmark finds no universally dominant model family: performance depends on data regime, available history, and forecast horizon. Its univariate evaluation also leaves schedule-driven surges beyond the models’ capabilities.

V. CONCLUSIONS

  • Across limited-history events and multi-year seasonal data, no model family dominates; long-context TimesFM reduces Melbourne MAE by approximately 16–49% versus a one-week context.Seasonal Naive and boosted trees remain competitive in limited-history events, while FMs provide cold-start value that diminishes as local data accumulates.Source: §V · pp. 7–8
  • The cost comparison measures memory, latency, and model size, and zero-shot FMs incur no in-domain training cost.Source: Table V · p. 8
  • Because the evaluation is univariate, the models cannot capture schedule-driven crowd surges; operational forecasting needs exogenous event features.Source: §V · p. 8

Related papers

Loading 2609.16415v1…