Cross-Validation
Cross-validation is a model-evaluation method that repeatedly trains on part of the data and tests on the rest to estimate how well a model will generalize to unseen data.
Cross-validation is a technique for estimating how well a model will perform on data it has not seen. Rather than relying on a single train/test split, which can be lucky or unlucky, it partitions the data into several folds, repeatedly trains the model on some folds and tests it on the held-out fold, and averages the results. The most common form, k-fold cross-validation, rotates through the data so every observation is used for testing exactly once, producing a more stable and trustworthy performance estimate than a single split.
Standard k-fold cross-validation assumes the data points are interchangeable, which is fine for many problems but dangerous for time series. With financial data, randomly shuffling observations across folds lets the model train on future data and test on the past, injecting look-ahead bias and inflating measured accuracy. For sequential data, time-aware variants are required, such as forward-chaining (expanding-window) validation or purged and embargoed schemes that prevent information from leaking across the boundary between training and testing periods.
For investors, cross-validation is central to honest model selection. It is how you decide between competing models or settings without fooling yourself, and how you tune the strength of regularization. But it is only as valid as its respect for time: a cross-validation procedure that ignores chronology will reward models that cannot actually be traded.
At hedgewing.ai the time-series discipline of cross-validation is taken to its logical conclusion in nightly walk-forward backtesting, which trains on a historical window, tests on the next out-of-sample period, then slides forward and repeats. This chronological evaluation is what lets the four-model ensemble's reported performance and calibrated confidence reflect conditions a live trader would actually have faced.
Related terms
Walk-Forward Backtesting · Look-Ahead Bias · Overfitting · Regularization
Back to the hedgewing.ai glossary · See AI stock forecasts