Early Stopping
Early stopping treats training time as a capacity knob: halt when a validation metric stops improving so the model does not wander into overfit.
Definition
Monitor a held-out loss or task metric; stop after patience epochs of no improvement and restore the best checkpoint. It is implicit regularization: the optimizer has not reached the interpolating minimum yet.
Why it matters
It is the cheapest regularizer and the one most often skipped when people ‘just train longer.’ On small financial panels it is often the only thing standing between you and a memorized IS Sharpe.
Case
An MLP on 8 years of daily factors: val IC peaks at epoch 12 and dies by 40. Early stopping keeps epoch 12. Without it the deck looks smarter and the live book does not.
How to read it
The validation set must be locked. If you early-stop on the test set, you no longer have a test set.