| Machine-learning forecasting | forecast_forecasting | Forecast one series or a panel with LightGBM, XGBoost, CatBoost, random forests, ridge, lasso and 40 other regressors on lags, rolling/seasonal/expanding/EWM statistics, calendar and exogenous features; recursive or direct multi-step; conformal prediction intervals; in-sample fitted values; feature importances. | 3 |
| Backtesting | forecast_backtest | Rolling-origin cross-validation of machine-learning forecasters: several windows, refit schedule, expanding or fixed-size training, interval coverage, accuracy metrics by model, series and window; LightGBM boosting-round search with early stopping. | 2 |
| Panel preparation | forecast_data | Validate a long table of series (duplicates, missing periods, missing values), insert missing periods on each series’ calendar and fill them, and generate synthetic panels to try the forecasting tools. | 2 |
| Forecast evaluation | forecast_evaluation | Score forecasts against actuals: MAE, RMSE, MAPE, sMAPE, WAPE, MASE, RMSSE, bias, quantile and multi-quantile losses, CRPS, interval coverage, Winkler score and calibration, per series, window or aggregated (weighted); Pareto-optimal models; errors by step ahead; residual rectification. | 2 |
| Feature engineering | forecast_features | The feature matrix the models train on (lags, rolling/seasonal/expanding/EWM statistics, calendar, static and dynamic exogenous columns, per-step targets), lags and rolling statistics of exogenous series, Fourier seasonality, trend and calendar terms with their future values, and the future rows a forecast needs. | 4 |
| Model update and transfer | forecast_transfer | Reuse a trained forecaster: append new observations without retraining and forecast from the new end; forecast series the models never saw (a new ticker, store or market) with intervals recalibrated, rescaled or density-ratio reweighted for them; covariate-shift weights between two feature sets. | 3 |
| Automatic model search | forecast_tuning | Search each model’s hyperparameters jointly with lags, rolling/EWM statistics, calendar features and target transforms by backtest, with a bounded number of trials; retrain the best and forecast. | 1 |
