01 Production-minded ML / Python
142 years of observations, seven local models and a second predictor built from 12 numerical weather systems, historical analogues and residual ML.
Next-day mean temperature
daily records across the active archive
local candidates under one protocol
future days held out from training
numerical weather systems in hybrid mode
02 Observations / 1884—2026
Daily mean, min/max envelope and precipitation from the production CSV. Change the window and inspect any day.
Fields already produced by data.py
03 Evaluation / MAE
Every candidate faces the same chronological holdout. Lowest MAE on the final 20% wins.
No future observation enters training. Time order stays intact.
04 Hybrid forecast / live API
Numerical forecasts set the baseline. Historical analogues add local context, then ridge ML learns the remaining error without touching the recent audit.
80% empirical interval20.20 — 21.83°C
Previous model runs are aligned with observed station temperatures.
The residual target is learned only from earlier dates.
Alpha and residual share are selected here.
The final quality check never changes model settings.
05 Learned weights / local model
The compact six-feature model exposes an interpretable coefficient set. Right is positive, left is negative and zero remains visible.
06 Architecture / two paths
Both predictors start with the same long local history. One ends in a saved file; the other reaches calendar tomorrow through live forecasts.
data.pypandas reads 142 years, fills wind gaps, then creates lags, rolling windows and three seasonal harmonics.
52,062 → 51,650preprocessing.pyThe first 80% becomes train; the final 20% remains an untouched future holdout.
order preservedtraining.pySeven models train and score through the same orchestration layer.
7 candidatesgradient_model.pyThe 14-feature model uses NumPy matrix operations, a custom loss loop and early stopping.
vectorizedapi_weather_model.pyTwelve numerical systems are bias-corrected and weighted by recent observed error.
12 providersapi_weather_model.pyHistorical analogues and 66 API-derived signals predict what the baseline still misses.
ridge MLreporting.pyLocal mode saves a trace to forecast.txt; hybrid mode prints its interval and backtest.
Deterministic, offline and tied to the next date in the dataset.
Targets calendar tomorrow and falls back to local analogues if APIs fail.
07 Latest saved local forecast
This is the latest result persisted by main.py.
It remains separate from the live hybrid run above.
01Next day forecast
02-----------------
03Model: Advanced weather gradient
04Last known date: 2026-07-16 00:00:00
05Forecast date: 2026-07-17
06Last temp avg: 21.8
07Predicted next temp avg: 21.41
08 Why the result is testable
This is a verifiable project snapshot, not a promise of permanent accuracy.
Lags and rolling means use past days only through shift(1).
No shuffling: evaluation reflects how the model performs later in time.
MAE expresses all seven local approaches as the same average miss in degrees.
Hybrid parameters are chosen on validation; recent audit only measures the result.
Average model miss, measured in degrees
Loss function used during gradient training
Ridge ML learns only what the adaptive API baseline still misses
09 Beta / next cycle
The current build now documents both real predictors. The next work is operational: repeatable outputs, tests and history.
Data → features → models → evaluation → forecast.
Metrics, charts, weights and architecture in one interface.
12 forecasts + analogues + adaptive baseline + residual ridge.
Persist local and hybrid metrics so the site can update automatically.
Compare every prediction with reality and track error over time.
Cover the pipeline and cache historical API responses.