Overfitting
Overfitting is when a model learns noise specific to its training data instead of the underlying signal, producing strong in-sample results that collapse on new data. It is closely related to Walk-Forward Backtesting, Ensemble Model and Calibrated Confidence, each defined in the hedgewing.ai glossary.
Overfitting is the central failure mode of machine learning applied to markets. With enough parameters, a model can memorise the idiosyncratic wiggles of a training set — noise that will never repeat — and score beautifully in-sample while failing completely on live data.
Defences include regularisation, simpler architectures, ensembling diverse models, and most importantly rigorous out-of-sample testing. The signal-to-noise ratio in financial returns is extremely low, which makes overfitting easy and devastating.
hedgewing.ai mitigates overfitting by stacking four structurally different architectures (LSTM, GRU, Temporal CNN, Transformer), calibrating confidence on held-out data, and gating every weight update behind walk-forward validation.
FAQ
What is Overfitting?
Overfitting is when a model learns noise specific to its training data instead of the underlying signal, producing strong in-sample results that collapse on new data.
Which terms are related to Overfitting?
Overfitting is closely related to Walk-Forward Backtesting, Ensemble Model and Calibrated Confidence. Each has its own entry in the hedgewing.ai glossary.
Is this definition of Overfitting investment advice?
No. This glossary entry explains what Overfitting means. It is educational reference material, not personalised investment advice and not a recommendation to buy or sell any security.
Create a free Hedgewing account to try hedgewing.ai yourself.