User guide¶
CaNN is built around a small set of user-facing concepts.
Core concepts¶
TrainedModel: a trained surrogate + theModelConfigit was trained with.ModelConfig: parameter bounds and the ordering of model inputs/outputs.MarketSurface: observed market data to calibrate to.CalibrationConfig: optimizer and batching settings for calibration.TrainingConfig: training hyperparameters for building a surrogate.
Data layout (important)
The columns of MarketSurface.market_inputs must match the ordering specified by
trained_model.model_config.market_inputs.
If you pass the right values in the wrong order, calibration will still run but produce nonsense.
Typical workflow¶
Forward pass (offline learning)¶
Train a surrogate from synthetic or historical data (see Examples).
Save the trained weights as an artifact for later reuse.
Backward pass (online calibration)¶
Load a pretrained surrogate (or one you trained offline).
Build a
MarketSurfacewithmarket_inputsandobserved.Run
calibrate_model.Validate the fit by predicting the surface using the calibrated parameters.
Where to look next¶
Examples: see Examples.
Architecture / implementation details: see Design.
API reference: see API reference.