Cost-Aware TSFM Forecasting

Problem#

Cities that most need hyper-local air-quality and weather forecasts often have the shortest data records and the smallest compute budgets. The usual options are a tiny searched architecture or transfer from data-rich cities. Foundation models are usually assumed to be too heavy, an assumption that is rarely tested.

What the repository contains#

• Three deployment strategies, one harness: a Green-NAS tiny specialist and a LightGBM direct model, transfer learning from data-rich cities, and a zero-shot foundation model (Chronos), plus a seasonal-naive floor.
• Two domains across 29 cities: hourly PM2.5 (UCI Beijing multi-site and an OpenAQ city panel) and 2 m temperature (Open-Meteo).
• Cost accounting: inference latency, measured energy and cost per 1,000 forecasts, combined into a cost-adjusted decision rule.
• A claim ledger: every number in the manuscript is traced to an artifact and a command in paper/RESULTS_LEDGER.md and checked by an audit script.

Key results#

• On PM2.5, the zero-shot model and the tuned specialist are statistically indistinguishable (MASE 0.662 vs 0.692 with covariates knowable at forecast time).
• The specialist’s apparent temperature edge (MASE 0.533 vs 0.792) shrinks to a small, non-significant gap (0.745) once its covariates are limited to values knowable at forecast time. The original edge was largely a perfect-foresight artifact.
• Under frequent retraining, the zero-shot model uses about ten times less measured energy.

Reproduce#

pip install numpy pandas lightgbm matplotlib shap
pip install torch chronos-forecasting
python src/run_forecast.py single --source pm25 \
  --data-path data/beijing_pm25/PRSA_Data_20130301-20170228 \
  --column Aotizhongxin --with-chronos --shap --out-prefix pm25_fm

The full regime and cross-city commands are in the repository README.