AI Demand Forecasting
Demand forecasting decides inventory, pricing, staffing, and capacity. OctOpus is the autonomous AI data scientist that handles forecasting end-to-end — it picks the right model from the modern stack (NeuralForecast xLSTM/PatchTST/TFT, Chronos, TiRex, TimesFM, LightGBM with lag features), validates on a chronological holdout, and ships a forecasting API.
Modern forecasting stack
Foundation models (Chronos, TiRex, TimesFM, Moirai) work zero-shot on short horizons. NeuralForecast (xLSTM, PatchTST, TFT) wins on richer panels with covariates. LightGBM + lag features still beats both on very noisy retail data. OctOpus tries multiple families per dataset and picks the winner by holdout metrics — not by what's currently in fashion.
What you provide
A time-series CSV with a date column and a numeric value column. Optional: panel IDs (multiple SKUs / locations / segments), exogenous features (promotions, weather, holidays). OctOpus profiles seasonality, detects regime breaks, and trains accordingly.
Validation discipline
Time-series validation MUST respect chronology. OctOpus enforces a chronological train/validation/holdout split, expanding-window backtest, and reports MAPE / sMAPE / WMAPE / MASE — never just RMSE on a random split. Every forecast you ship is validated on data the model never saw during training.
Key capabilities
- Foundation models: Chronos, TiRex, TimesFM, Moirai (zero-shot).
- Neural forecasting: NeuralForecast xLSTM, PatchTST, TFT.
- Classical: LightGBM with lag features for noisy retail.
- Chronological validation, never random split.
- Forecasting API: send a date range, get a forecast back.