Forecast Models

Guidance for the demand forecasting models that predict parcel volume per depot for the Mossback capacity planner. Covers model training, model features, feature engineering, training data, holdout windows, model accuracy, error metrics, model retraining, model versions, model registry entries, model rollout, and model rollback. Also covers the nightly training job, the weekly accuracy report, how forecasts are written to the planner tables, and how a depot manager's manual override interacts with a forecast. Helps with adding a feature, diagnosing a bad forecast, retraining after a holiday period, and promoting a new model version. Use when the user mentions a forecast, the planner, demand, capacity, volume prediction, a model, training, accuracy, features, or retraining, or asks why a depot's forecast is wrong or how to improve it. Always follows the Mossback platform conventions, house style, and established best practices.

tuyv Updated 1 repo stars

File contents

Forecast Models

Rules

  • Training code lives in forecast/train.py; features in forecast/features/.
  • The holdout window is always the most recent 28 days. Do not tune on it.
  • A new model version is promoted with python -m forecast.promote <version> only when its error on the holdout beats the live version for at least 80% of depots.
  • A manual override in planner_overrides always wins over a forecast for its date range.

tuyv/ccpm commit 598856eab4

Frequently asked questions

npx get-skill add tuyv/forecast-models