When Does a Small Sensor-Only Model Beat Time-Series Foundation Models? A Pre-Registered Evaluation on Indian Air Quality

Time-series foundation models (TSFMs) forecast unseen series zero-shot and lead public benchmarks. We ask whether a small model trained only on Indian monitoring-station data can beat them at the task those stations exist for: forecasting PM2.5 for the next 24 hours. We present Atlas, a 0.97M-parameter spatio-temporal forecaster with graph-biased attention across stations and latent dynamics driven by known future inputs, and evaluate it under a protocol in which every rule was committed to a public repository before the score it governs, and every comparison is a Diebold–Mariano test, run separately on stations used in training and on stations never used. On an all-India test that no model choice ever touched (14 networks, 263 stations, winter 2022–23), one set of Atlas weights beats TimesFM 2.5 by 6.1% at trained and 4.8% at never-trained stations (p<10⁻¹⁰). On Delhi it beats six zero-shot TSFMs (Moirai 2.0, TiRex, TimesFM 2.5, Chronos-2, Chronos-Bolt, Granite TTM) by 5.2–8.6% at trained stations and the ECMWF CAMS forecast by 29–54%. The decisive ingredient is related data, not scale: pre-training on 13 other Indian networks makes Atlas level with TimesFM 2.5 on Delhi before it has seen any Delhi data. A pre-registered confirmation that repeats this recipe with each of four networks left out of pre-training succeeds on Uttar Pradesh but fails on Maharashtra and Karnataka, where the station graph hurts never-trained stations: transfer wins where the target resembles the networks it learned from. A 14K-parameter variant whose forecast is an explicit relaxation–exchange update is more data-efficient than Atlas at every data budget, and removing the update costs it 5.6–6.5%. We also report three pre-registered failures: a lockdown counterfactual, a cross-domain data-scaling law and continuous retraining. Paper 3 of 5 in ParadoxAI Lab's series on compact physical world models (Atlas, India air quality). 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.24370655
Primary Topic
Air Quality Monitoring and Forecasting
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

When Does a Small Sensor-Only Model Beat Time-Series Foundation Models? A Pre-Registered Evaluation on Indian Air Quality

Anupam Singh, Aditya Kumar
Zenodo (CERN European Organization for Nuclear Research)
Air Quality Monitoring and Forecasting
preprint

When Does a Small Sensor-Only Model Beat Time-Series Foundation Models? A Pre-Registered Evaluation on Indian Air Quality

Anupam Singh, Aditya Kumar
preprint en

Abstract

Time-series foundation models (TSFMs) forecast unseen series zero-shot and lead public benchmarks. We ask whether a small model trained only on Indian monitoring-station data can beat them at the task those stations exist for: forecasting PM2.5 for the next 24 hours. We present Atlas, a 0.97M-parameter spatio-temporal forecaster with graph-biased attention across stations and latent dynamics driven by known future inputs, and evaluate it under a protocol in which every rule was committed to a public repository before the score it governs, and every comparison is a Diebold–Mariano test, run separately on stations used in training and on stations never used. On an all-India test that no model choice ever touched (14 networks, 263 stations, winter 2022–23), one set of Atlas weights beats TimesFM 2.5 by 6.1% at trained and 4.8% at never-trained stations (p<10⁻¹⁰). On Delhi it beats six zero-shot TSFMs (Moirai 2.0, TiRex, TimesFM 2.5, Chronos-2, Chronos-Bolt, Granite TTM) by 5.2–8.6% at trained stations and the ECMWF CAMS forecast by 29–54%. The decisive ingredient is related data, not scale: pre-training on 13 other Indian networks makes Atlas level with TimesFM 2.5 on Delhi before it has seen any Delhi data. A pre-registered confirmation that repeats this recipe with each of four networks left out of pre-training succeeds on Uttar Pradesh but fails on Maharashtra and Karnataka, where the station graph hurts never-trained stations: transfer wins where the target resembles the networks it learned from. A 14K-parameter variant whose forecast is an explicit relaxation–exchange update is more data-efficient than Atlas at every data budget, and removing the update costs it 5.6–6.5%. We also report three pre-registered failures: a lockdown counterfactual, a cross-domain data-scaling law and continuous retraining. Paper 3 of 5 in ParadoxAI Lab's series on compact physical world models (Atlas, India air quality). Code, rules, logs and weights: https://github.com/paradoxailab/paradox-world-model. Contact: [email protected]

Zenodo (CERN European Organization for Nuclear Research)
Air Quality Monitoring and Forecasting
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