forecast_* tool: LightGBM, XGBoost, CatBoost, random forests, ridge, lasso and 40 other regressors trained on lags, rolling, seasonal, expanding and EWM statistics, calendar features and exogenous columns; recursive or direct multi-step forecasts with conformal prediction intervals; rolling-origin backtests; automatic hyperparameter and feature search; model update and transfer to new series; and forecast accuracy metrics. It runs on Fincept’s servers; your agent sends numbers or data references and gets back a result envelope with tables and charts.
Fincept Forecast tools are free. They count toward the rate limit only; a
$fincept reference inside a call is charged like a direct call to that tool.Modules
Tools are grouped in modules. Each module is a toolset namedforecast_<module>, so you can list it or search within it. The Fincept Forecast reference lists every module and tool.
forecast_catalog returns the live list of modules and tools from inside your agent.
Inputs
The id, time and target columns are
unique_id, ds and y unless id_col, time_col and target_col name others. Columns listed in static_features are constant within each series; every other extra column is dynamic and needs future values.
Limits
A computation that runs past its time limit answers
timeout; a busy engine answers busy and the call can be retried a few seconds later.
Outputs
Forecasting tools return the forecast of every series and model with-lo-<level> and -hi-<level> interval bands, charts and feature importances. Backtests return every window’s forecasts and the chosen metrics per model, window and series; forecast_auto_ml adds each model’s best hyperparameters, features and trial losses. forecast_evaluate returns scores per series and per model, the Pareto-optimal models and errors by step ahead.
Example
Prompt
machine learning forecast, finds forecast_cross_validation and forecast_ml_forecast, and passes the closes as a $fincept reference to market_get_candles.