Repository navigation
v0.3.0 — Weather Prediction + Lorenz Attractor
What's New
CRNG validated against 10 years of real weather data and tested on Lorenz strange attractors.
Weather Prediction Experiment
Data: 10 years daily temperature (2014–2023), ERA5 reanalysis (ECMWF), 3 cities.
Key Result — Kurtosis Match 3/3:
| City | Real K | PRNG K | CRNG K |
|---|---|---|---|
| São Paulo | 6.39 | 2.99 | 6.45 |
| New York | 4.71 | 2.99 | 5.79 |
| London | 4.98 | 2.99 | 5.57 |
PRNG always produces K ≈ 3.0 (Gaussian). CRNG captures real fat-tailed climate variability.
Distribution scorecard: CRNG wins 5/6 metrics (83%)
Forecast MAE: CRNG wins 4/5 horizons in São Paulo (fattest tails). Mixed in seasonal climates.
Lorenz Attractor Experiment
CRNG preserves chaotic dynamics better than PRNG:
- Vol clustering: CRNG 0.996 ≈ Deterministic 0.997 >> PRNG 0.990
- Regime changes: CRNG 289 < PRNG 338 (fewer artificial transitions)
Data Sources (verifiable)
- Open-Meteo Historical API (ERA5, CC BY 4.0)
- NOAA GHCN-Daily
- NOAA ISD
Reproduce
pip install crng
python experiments/weather_prediction.py
python experiments/lorenz_attractor.py