Use case · Forecasting
Autonomous AI forecasting — from CSV to deployed prediction API.
Drop your time-series data, describe the horizon, get a forecast. OctOpus picks the model family the data calls for — deep forecasting architectures, time-series foundation models, or tree-based models with lags — runs experiments, validates on a real out-of-sample holdout, and deploys a prediction endpoint.
TL;DR. Forecasting is OctOpus's strongest wedge. The agent moves between model families as the evidence warrants — tree-based models with lags and calendar features, modern deep forecasting architectures, time-series foundation models — keeping whichever wins on out-of-sample holdout error.
What OctOpus forecasts well
- Demand and inventory — per-SKU, per-store, per-warehouse. Daily, weekly, monthly horizons.
- Revenue and KPI — per-customer, per-segment, per-product. Multi-step ahead with confidence bounds.
- Energy and load — power consumption, renewable generation, grid balancing.
- Sensor and IoT — environmental, industrial, vehicle telemetry.
- Hydrology and earth science — rainfall-runoff, river stage, flood horizons. Includes physics-informed feature engineering.
- Hierarchical panels — thousands of related series trained as a single global model.
Models the agent rotates through
| Approach | Family | When the agent picks it |
|---|---|---|
| Baseline | Gradient boosting with lag, rolling, calendar features | Almost always — fast, strong, interpretable. |
| Tuned gradient boosting | A different boosting family with automated hyperparameter search | When the baseline has headroom and the data justifies a search. |
| Deep / modern | Modern deep forecasting architectures | Long horizons, multi-series panels, exogenous signals. |
| Foundation | Time-series foundation models | Zero-shot first; cold-start data; benchmark anchor. |
| Stacking | An ensemble over diverse base learners | When the residuals of different families are uncorrelated. |
How a forecasting run looks
- Profile. The agent reads the schema, detects the time column, infers granularity (daily / hourly / weekly), spots multiple series, and flags exogenous candidates.
- Plan. Writes a research spec: target, horizon, validation strategy (rolling-origin or fixed cutoff), and which families to try in which order.
- Run. Writes new training code for every experiment, executes it in a sandbox, reads the metrics.
- Diagnose. When something fails, the agent reads the traceback, writes a targeted fix (dtype, missing dependency, cardinality blow-up), and retries.
- Validate. The winner is scored on a holdout slice the LLM never sees.
- Deploy. A prediction endpoint plus a downloadable bundle with the full training code and the trained model.
What you get back
- Forecast with confidence intervals at the requested horizon.
- Holdout error per series and aggregate (sMAPE, MAPE, MASE, MAE — the agent picks per data shape).
- Calibration plot and residual diagnostics.
- The exact training code the agent wrote — fully inspectable, fully reproducible.
- A deployable prediction API.