Traffic Incident Forecasting Eval

---
name: traffic-incident-forecasting-eval
description: Evaluates spatiotemporal models' ability to localize traffic collision events in time and space, and to forecast network-level congestion and emissions. It probes multi-horizon forecasting accuracy and spatial-temporal coherence under simulated disruption scenarios. Use when the user wants to benchmark on NYC Broadway corridor, or asks about evaluating this task. Reports containment_performance.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2509.25515
  bibtex_key: [REDACTED]
  confidence: medium
---

# traffic-incident-forecasting-eval

> Spatiotemporal Forecasting of Incidents and Congestion with Implications for Sustainable Traffic Control — Kinchen et al. (2025) (arXiv:2509.25515, 2025)

## What this evaluates

Evaluates spatiotemporal models' ability to localize traffic collision events in time and space, and to forecast network-level congestion and emissions. It probes multi-horizon forecasting accuracy and spatial-temporal coherence under simulated disruption scenarios.

## Datasets

- **NYC Broadway corridor** — total ?; splits: test (-1)

## Metrics

- `containment_performance` **(primary)** — range: other
  - Counts the number of ground-truth collision events that fall within the model's predicted temporal and spatial intervals. Reported as stacked histograms of actual events captured within predicted ranges.
- `network_ce_tti_accuracy` — range: other
  - Qualitative and visual assessment of predicted vs. ground-truth Carbon Emissions (CE) and Travel Time Index (TTI) across network links and time steps, evaluating magnitude reproduction and spatial coherence.

## Input / output format

**Input**: Historical and current spatiotemporal traffic state data (vehicle trajectories, speeds, network topology) used to forecast collision events (BiLSTM) or network-level congestion and emissions (DCRNN).

**Output**: For BiLSTM: predicted intervals for collision time $t$ and spatial coordinates $(x, y)$. For DCRNN: predicted per-link CE emissions and TTI values across time horizons.

## Scoring recipe

```python
# Containment Performance
covered = 0
for event in ground_truth:
    if event.t in pred_t_interval and (event.x, event.y) in pred_xy_region:
        covered += 1
containment_score = covered / len(ground_truth)

# Network CE/TTI Accuracy
# Evaluated via visual/magnitude comparison of predicted vs ground-truth link-level time series
# No explicit scalar formula provided; assessed by spatial coherence and peak alignment
```

## Common pitfalls

- Evaluation relies heavily on visual inspection of histograms and plots rather than standardized numerical metrics like RMSE or MAE.
- Containment intervals are reported as stacked histograms of counts, making precise threshold-based accuracy extraction difficult without raw data.
- Network-level metrics are compared qualitatively for spatial distribution and magnitude trends, lacking a unified scalar score.

## Evidence (verbatim from paper)

> Figs. 3-6 present the containment performance of the BiLSTM in predicting collision time $t$ and spatial coordinates $(x, y) . The stacked histograms report the number of actual events captured within the predicted intervals across different collision types.

## Citation

```bibtex
@misc{kinchen2025spatiotemporal,
  title={Spatiotemporal Forecasting of Incidents and Congestion with Implications for Sustainable Traffic Control},
  author={Kinchen et al. (2025)},
  year={2025},
  note={arXiv:2509.25515}
}
```

- arXiv: 2509.25515