beita6969 / scienceclaw-benchmark-for35

Monthly seasonal series forecasting (Monash Tourism Monthly)

Task. Input: histories of monthly series (about 90-330 observations, own units), horizon 24, seasonal period 12, possibly a pool of other series of the same collection. Deliverable: float array (n_series, 24), the 24 months after each history, finite, >= 0, in the series' units.

Quality. Mean MASE over series, lower is better: mean absolute error over the horizon divided by the in-sample seasonal-naive error (lag 12). Reference: seasonal naive (repeat the last 12 months); its published mean MASE on the full Tourism Monthly set is about 1.63.

Library (check scienceclaw_tools(operation=show, target=forecast)):

• scilib.forecast.fit_predict(histories, horizon, period, train=..., phase=...): median of damped ETS, Theta, STL+ETS, airline SARIMA and global ridge / extra-trees window models, clipped at 0. Also forecast_panel, combine, global_window, global_lgbm, rolling-origin backtest_panel + combination_mase, metrics mase, panel_mase, smape, rmsse.
• Pretrained, zero-shot, GPU or remote worker (guard with scilib.tsfm.available(); assets chronos_2, chronos_bolt_base): forecast.pretrained_forecast(histories, 24) (Chronos-2 median, last 120 values, log1p for non-negative series); fit_predict(..., pretrained=w) blends it with weight w; tsfm.forecast gives quantiles.
• Operators: seasonal_panel_forecast_ensemble, statistical_forecast_single_method, forecast_backtest_method_ranking, chronos_median_forecast_nonnegative_panel.

Routes. Baselines: seasonal naive, per-season mean, damped ETS, Theta. Default: the fit_predict ensemble, optionally blended with a pretrained forecast. Choose by backtest_panel (hold out the last 24 months of each history).

Rules.

• Respect temporal causality: nothing after a series' forecast origin may inform its forecast. When backtesting, cut every pool series at the same origin (train_cut="auto" does this when the pool contains the histories).
• Build the deliverable from the series to be forecast, not from truncated backtest copies of them.
• Pass phase / train_phase (month - 1 of the first observation) when series start in different months.
• Mean MASE over few series is noisy; a 2-3 % backtest gain is within noise.
• Chronos pretraining corpora include Monash-archive-type data, so tourism series may have been seen. Disclose this overlap.