Turnover Control in ML Strategies

A model refreshed daily produces a new score for every stock every day, and each small wiggle in that score — even one driven by noise rather than real information — technically implies a slightly different optimal position. Trading every wiggle would generate enormous turnover for very little benefit, since much of the day-to-day change in a noisy score isn't new information at all. Turnover control is the set of practical rules that stop a strategy from chasing noise.

Why naive trading on every score update is expensive

Consider a score that moves by a small random amount most days, with an occasional larger move that reflects real new information. If a strategy resizes positions to match the score exactly every day, most of the resulting trades are chasing the small random moves — pure noise, at real cost, for essentially zero expected benefit — while paying the same cost for the rare, informative moves that were worth trading. The fix isn't to ignore the score; it's to only act on changes large enough, or persistent enough, to be worth the cost of trading.

A mean-reverting path wiggles constantly around a slower-moving level — trading every wiggle instead of the level itself is exactly the mistake turnover control is designed to prevent.

Worked example

A model's daily score for a given stock has a standard deviation of noise-driven daily change of about 0.02 (on a scale where a full position corresponds to a score of 1), but real information shifts the underlying level by roughly 0.15 every few weeks on average.

1. No turnover control. Resizing daily to match the score exactly generates trading on essentially every 0.02 wiggle, adding up to far more total turnover over a month than the handful of genuine 0.15 shifts would require on their own.
2. Threshold rule. Only rebalancing a name when its score has moved more than 0.05 from the last traded level cuts out almost all the noise-driven trades while still catching the real 0.15 shifts within a few days of them occurring.
3. Result. Backtested turnover falls by more than half, trading costs fall proportionally, and because the filtered-out trades were mostly noise anyway, the strategy's gross return is barely affected — so net-of-cost performance improves substantially.

What this means in practice

Common turnover-control techniques include trading only when a score crosses a threshold, smoothing the score with a moving average before it drives trades, imposing a minimum holding period once a position is opened, and capping how much any single position can change in one rebalance. The right combination depends on how fast the underlying signal decays — a signal with a short half-life needs to react quickly and tolerates less smoothing; a slow-decaying signal can be smoothed heavily with little cost to the captured alpha.