Trained Small, Scored Fairly: A Sub-Million-Parameter Graph Model Against Published Specialists and Zero-Shot Foundation Models on Physical Network Forecasting

Time-series foundation models now forecast any series zero-shot, while graph specialists are still trained per network, and the two are rarely compared under one protocol. We compare them on physical sensor networks with a fixed, open benchmark: six traffic networks (PEMS03/04/07/08, METR-LA, PEMS-BAY) and two monsoon rainfall regions, four tracks (standard context, open context, 30% dead sensors, zero-shot transfer), one scorer for point error, pinball loss and interval score, saved forecasts and paired Diebold–Mariano and block-bootstrap tests. We introduce PWM-Dyn, a 0.3–0.4M-parameter graph-diffusion model with explicit missing-data inputs, quantile outputs and split-conformal calibration, and compare it with STID, Graph WaveNet and STAEformer trained with their official code on identical windows under a fixed compute budget, and with zero-shot TimesFM 2.5 and 3.0. Configurations were chosen on validation and the reported models use fresh seeds; four networks were never used during development. Under a 12-step context a three-seed PWM-Dyn ensemble has the lowest MAE on five of six traffic networks (0.3–4.2% below the best trained opponent on the four new ones), is not significantly different from STID on three (PEMS08, PEMS04, PEMS07), and trails a budget-capped STAEformer on PEMS08. Against TimesFM 3.0 it has lower MAE and interval score on seven of eight tasks (five of six traffic networks; the MAE difference is significant on four; rainfall entries from the previous model version) and loses on PEMS03. With 30% of sensors dark it beats per-sensor climatology on seven of eight tasks, where specialists without a missing-data mechanism fail. An ablation traces most of an earlier accuracy loss on PEMS04 to a climatology input. Code, benchmark and forecasts: https://github.com/paradoxailab/paradox-world-model. Paper 1 of 5 in ParadoxAI Lab's series on compact physical world models (PWM-Dyn and ParadoxBench). Code, rules, logs and weights: https://github.com/paradoxailab/paradox-world-model. Contact: [email protected]

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-19
DOI
https://doi.org/10.5281/zenodo.24370610
Primary Topic
Traffic Prediction and Management Techniques
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Trained Small, Scored Fairly: A Sub-Million-Parameter Graph Model Against Published Specialists and Zero-Shot Foundation Models on Physical Network Forecasting

Anupam Singh, Aditya Kumar
Zenodo (CERN European Organization for Nuclear Research)
Traffic Prediction and Management Techniques
preprint

Trained Small, Scored Fairly: A Sub-Million-Parameter Graph Model Against Published Specialists and Zero-Shot Foundation Models on Physical Network Forecasting

Anupam Singh, Aditya Kumar
preprint en

Abstract

Time-series foundation models now forecast any series zero-shot, while graph specialists are still trained per network, and the two are rarely compared under one protocol. We compare them on physical sensor networks with a fixed, open benchmark: six traffic networks (PEMS03/04/07/08, METR-LA, PEMS-BAY) and two monsoon rainfall regions, four tracks (standard context, open context, 30% dead sensors, zero-shot transfer), one scorer for point error, pinball loss and interval score, saved forecasts and paired Diebold–Mariano and block-bootstrap tests. We introduce PWM-Dyn, a 0.3–0.4M-parameter graph-diffusion model with explicit missing-data inputs, quantile outputs and split-conformal calibration, and compare it with STID, Graph WaveNet and STAEformer trained with their official code on identical windows under a fixed compute budget, and with zero-shot TimesFM 2.5 and 3.0. Configurations were chosen on validation and the reported models use fresh seeds; four networks were never used during development. Under a 12-step context a three-seed PWM-Dyn ensemble has the lowest MAE on five of six traffic networks (0.3–4.2% below the best trained opponent on the four new ones), is not significantly different from STID on three (PEMS08, PEMS04, PEMS07), and trails a budget-capped STAEformer on PEMS08. Against TimesFM 3.0 it has lower MAE and interval score on seven of eight tasks (five of six traffic networks; the MAE difference is significant on four; rainfall entries from the previous model version) and loses on PEMS03. With 30% of sensors dark it beats per-sensor climatology on seven of eight tasks, where specialists without a missing-data mechanism fail. An ablation traces most of an earlier accuracy loss on PEMS04 to a climatology input. Code, benchmark and forecasts: https://github.com/paradoxailab/paradox-world-model. Paper 1 of 5 in ParadoxAI Lab's series on compact physical world models (PWM-Dyn and ParadoxBench). Code, rules, logs and weights: https://github.com/paradoxailab/paradox-world-model. Contact: [email protected]

Zenodo (CERN European Organization for Nuclear Research)
Traffic Prediction and Management Techniques
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon