Model Retraining Cadence

A model trained once on five years of history and then left untouched for the next three years is quietly betting that nothing important about the market has changed. That bet is usually wrong: relationships between features and returns drift as market structure, participants, and regimes shift. The obvious fix — retrain constantly, on the freshest data available — has its own cost: retraining introduces fresh estimation noise every time, can whipsaw a live portfolio's positions and turnover, and risks fitting to whatever happened to be true in the last few noisy weeks. Retraining cadence is the decision of how often to refit, and it sits directly between these two failure modes.

The tradeoff in one picture

Think of model error as having two sources that move in opposite directions as cadence changes. Retrain rarely and staleness error grows — the model's learned relationships increasingly reflect a market regime that no longer exists. Retrain frequently and estimation error grows — each refit is based on a smaller, noisier recent window, so the model's parameters bounce around more than the underlying truth actually does. The right cadence sits where the sum of the two is smallest, and that point depends on how fast the true relationship changes (a slow macro factor tolerates infrequent retraining; a fast microstructure signal doesn't) and how much data each refit needs to be estimated reliably.

Worked example: choosing a cadence via walk-forward testing

A team building a monthly equity-selection model tests three cadences by walk-forward backtest, refitting and measuring out-of-sample rank correlation each time:

Cadence	Avg. out-of-sample IC	Turnover cost impact
Retrain weekly	0.021	High — frequent small parameter shifts
Retrain monthly	0.034	Moderate
Retrain quarterly	0.019	Low, but IC visibly decays late in each quarter

Monthly retraining wins here: weekly refits are too noisy given the model's modest data requirements, and quarterly refits let real drift accumulate before the next update, visible as declining IC in the third month of each quarter. The team also checks that the model's inputs aren't near-identical each week, since retraining a model on nearly the same data it just saw adds noise without adding information.

What this means in practice

Set cadence with a walk-forward test that measures out-of-sample performance across candidate refit frequencies, the same way you'd tune any other hyperparameter — don't just default to "daily" or "whatever's easiest to schedule." Faster-decaying signals (order-flow, short-horizon microstructure) generally need faster retraining than slow ones (macro, valuation), and cadence should be revisited whenever a strategy's realized performance starts drifting from its backtest, since that's often the first sign staleness has crept past what the current schedule can handle.