Three candidate models
Chosen for transparency over novelty.
- Same hour last week
- Mean of the same hour-of-week
- Ridge regression on daily/weekly cycles and a one-week lag
Forecasting
Forecasts that show their working.
Every run compares simple, explainable models on the most recent week of real data, then uses the most accurate one for the next seven days.
Honest scoring
Models are scored on a time-ordered holdout (never shuffled) using MAE, RMSE, WAPE and MAPE.
Uncertainty bands
Prediction intervals come from the empirical spread of holdout errors, so the band reflects how wrong the model has actually been.
Model registry
Saved runs record training window, selected model, scores and code version for later review.
What this does not do
- No weather or occupancy inputs yet, so unusual weeks will forecast poorly.
- At least three weeks of hourly data are required.