Imported from useriswild7099/ADHARA- (.agents/skills/forecast-model/SKILL.md). Install upstream with npx skills add useriswild7099/ADHARA- --skill forecast-model. Copyright stays with the author.

Forecast model conventions (MOIL / SIH26009)

Pick the model based on how much real history you have

• Fewer than ~24 months of real monthly data → don't trust a neural model. Use statsforecast (AutoETS/AutoARIMA) as the primary model.
• 24+ months of real data with rainfall/downtime as extra columns → neuralforecast (start with N-BEATS or TFT) is worth trying, but ALWAYS keep the statsforecast run as a comparison baseline (see train_forecast.py — it already runs both and prints the error of each).
• Data is real but very sparse/gappy (e.g. only a handful of points per mine) → try chronos-forecasting or timesfm zero-shot instead of training anything from scratch. Note: Nixtla's own benchmarking has found Chronos can be noticeably slower and sometimes less accurate than a well-tuned classical model, so don't assume the fancier model wins — always compare against the simple baseline before choosing it.

Never skip

• The shortfall-risk threshold logic and the "likely cause" rule (rainfall vs. downtime vs. unclear) already in train_forecast.py. This is what turns a raw forecast into something a mine manager can act on — it matters more to judges than which model produced the number.

Output contract

Save forecast_results.csv with columns: month, actual, regression_forecast (or chosen model's forecast), shortfall_risk, likely_cause app.py reads this file directly — keep these column names stable.