Refit Frequency for Cross-Sectional Models

A cross-sectional model ranks hundreds of stocks every day using relationships it learned from history — value beats growth in some periods, momentum works in others, and the mix of what matters keeps drifting. The model itself doesn't update on its own. Someone has to decide: every day? Every month? Every quarter? Refit too rarely and the model keeps applying last year's relationships to this year's market. Refit too often and you're re-estimating a noisy signal from more noise, changing the model's view for reasons that have nothing to do with a real regime shift.

The tradeoff in plain terms

Think of the model's coefficients as a recipe learned from a cookbook of past trades. If the underlying cuisine changes — say, momentum stops paying and quality starts — an old recipe keeps producing the wrong dish. A fresh refit tastes the latest ingredients and adjusts. But if you re-taste and rewrite the recipe every single day, you're mostly reacting to that day's noisy sample rather than to any real shift in the underlying cuisine, and the recipe itself becomes unstable — jumping around and forcing costly changes to the portfolio for no real improvement in forecast quality.

Formally, a refit uses a training window of $T$ periods to estimate parameters $\hat\beta$, and those parameters are then applied out-of-sample until the next refit date. The staleness of the model at any point is how many periods have passed since the last refit; staleness increases the mismatch between $\hat\beta$ and the true, currently-operating relationship. In plain English: every day since the last refit, the model is very slightly more out of date, and the cost of that slippage has to be weighed against the cost — turnover, instability, overfitting to a short recent window — of refitting again.

Worked example: comparing two refit schedules

Suppose a cross-sectional ranking model is trained on 5 years of monthly data ($T=60$ months) and back-tested with two schedules: refit every month, versus refit every quarter. Refitting monthly uses the most current relationships each time, but each refit re-estimates from a training window that only grows by one new month relative to the last, so $\hat\beta$ barely changes and the model burns transaction costs re-ranking stocks on the margin of noise. Refitting quarterly uses three fresh months of data each time; suppose the resulting information coefficient (IC) — a correlation between predicted rank and realized return — averages 0.045 with quarterly refits versus 0.041 with monthly refits, but monthly refits' extra turnover costs roughly 15 basis points a year more than the very small IC gain is worth. Backtesting a small grid of refit frequencies (monthly, quarterly, semi-annual) and comparing net-of-cost Sharpe ratio, not raw IC, is how this decision actually gets made in practice — not by picking the schedule that looks "most responsive" on paper.

What this means in practice

Most quant shops settle on a monthly or quarterly refit for cross-sectional equity models, not a daily one, because relationships between fundamentals or factors and returns don't actually change day to day — only the noise in the estimate does. The right test isn't "does a faster refit raise in-sample fit" (it almost always will), but whether it raises out-of-sample, net-of-cost performance across a walk-forward backtest, since faster refitting adds both turnover and the risk of chasing a spurious pattern that reverses next period.