🌤 PolyWeather

Multi-model ensemble postprocessing for daily maximum temperature forecasts at 10 globally distributed airport stations. Uses LightGBM regression with ensemble member shifting to produce calibrated bucket probability predictions.

This repository contains the code and dataset for the NLP Course final project (Spring 2026).

How It Works

11 NWP Models (Previous Runs API)             ICON Ensemble (39 members)
  ECMWF, ICON, GFS, JMA, GEM,                  raw probability distribution
  UKMO, KNMI, DMI, CMA, MeteoFrance,                       │
  NCEP GraphCast                                           │
         │                                                 │
         ▼                                                 │
   LightGBM Regressor                                      │
   (8,339 samples, 27 features)                            │
   → calibrated T_max prediction                           │
         │                                                 │
         └─────────── shift ensemble mean ─────────────────┘
                              │
                              ▼
              P(bucket) = shifted members in bucket / N

Quick Start

git clone <repo-url>
cd polyweather
pip install -r requirements.txt

# Collect data and train model from scratch
python retrain.py --from-scratch

# Or use the included pretrained model
python bot.py --test-forecast

# Single scan / continuous run
python bot.py --once
python bot.py --interval 60

Repository Structure

File	What it does
cities.py	City configs — coordinates, stations, timezones
forecast.py	Ensemble fetch + LightGBM calibration
predictor.py	LightGBM inference — fetches 11-model features, predicts T_max
markets.py	Polymarket market scanner
strategy.py	Edge detection, position sizing, exits
executor.py	Order execution
bot.py	Main scan loop
retrain.py	Automated data collection + model retraining

Dataset

The dataset is included in data/:

File	Description	Rows
data/model_forecasts_day1.csv	24-hour-ahead forecasts from 11 NWP models + 7 context variables	8,340
data/wu_historical.csv	Daily T_max from Weather Underground (ground truth)	11,138
data/clim_normals_era5.csv	ERA5 1991-2020 climatological normals per city/day	3,660
data/models/global_model_tuned.txt	Trained LightGBM model (native format)	—
data/models/feature_cols.pkl	Feature column list	—
data/models/best_params.json	Optimal hyperparameters (Optuna, 80 trials)	—

Coverage: 10 airport stations, Feb 5, 2024 — May 18, 2026 (~28 months).

Data collection: Run python retrain.py --from-scratch to reproduce the dataset from public APIs.

Model

LightGBM regression trained on 27 features:

• NWP forecasts (11): ECMWF IFS, GFS, ICON, JMA, GEM, MeteoFrance, UKMO, KNMI, DMI, CMA GRAPES, NCEP GraphCast
• Context (7): Cloud cover, precipitation, radiation, humidity, wind speed, pressure, dew point (ECMWF IFS, 24h ahead)
• Seasonal (3): doy_sin, doy_cos, city_id
• History (2): WU T_max d-2 and d-3
• Ensemble stats (4): mean, std, min, max across NWP models

Hyperparameters (Optuna tuned): 313 estimators, lr=0.029, max_depth=6, num_leaves=35.

Honest walk-forward CV bucket accuracy: 37.9% (1°C discretization).

Per-city Performance

City	Bucket Accuracy	MAE (°C)
Wellington	43.8%	0.75
Helsinki	42.2%	0.92
Ankara	41.8%	0.81
Singapore	41.6%	0.77
Tel Aviv	41.4%	0.78
Buenos Aires	35.6%	0.99
Seoul	34.6%	1.00
Toronto	33.3%	1.13
Tokyo	32.8%	0.98
São Paulo	32.0%	1.04

Data Sources

• NWP forecasts: Open-Meteo Previous Runs API — temperature_2m_previous_day1 variable for 24h-ahead forecasts
• Ground truth: Weather Company V1 API — daily T_max at airport ICAO stations
• Climatology: Open-Meteo Climate API — ERA5 1991-2020 normals
• Markets: Polymarket Gamma API + CLOB /price endpoint (downstream application)

Notes

This project explores statistical postprocessing of ensemble weather forecasts — a well-established field in meteorology. The downstream application is automated trading on Polymarket temperature prediction markets, but the methodology applies to any setting requiring calibrated short-term temperature forecasts (renewable energy, agriculture, climate risk assessment).