Decomposing Where a Signal's Turnover Comes From

Two signals can have identical information coefficients and wildly different costs, because turnover — how much of the portfolio has to trade at each rebalance — isn't determined by skill alone. Before trying to cut turnover, it pays to know which of three sources is producing it, because each has a different fix.

The three sources

Genuine signal change. The underlying data actually moved — earnings were revised, a price crossed a moving average, a fundamental ratio shifted. This turnover is buying information and is, in principle, worth paying for.

Rank churn from noise. Even if the underlying signal barely moved, small measurement noise can flip the rank of two names that were nearly tied, especially in a densely packed part of the cross-section. This turnover buys nothing — it's an artifact of converting a continuous, noisy number into a discrete rank or a portfolio weight.

Mechanical rebalancing. Turnover generated purely by the rebalance schedule and portfolio construction rules — resetting to target weights, rounding to lot sizes, forced trades from a risk constraint binding. This turnover has nothing to do with the signal's information content at all.

A worked example

A signal's raw daily turnover runs at 8% of the book. The researcher re-runs the backtest with the score bucketed into 10 deciles instead of continuous ranks, holding everything else fixed: turnover falls to 5.5%. That 2.5-point gap is a reasonable estimate of pure rank-noise churn — trades that happened only because a name crossed a rank threshold by a hair, not because anything meaningful changed. Separately, forcing a strict rebalance-to-target every day (versus letting weights drift within a band) adds another 1 point of mechanical turnover in a side test that varies only the construction rule, holding the signal fixed.

That leaves roughly 4.5% as turnover attributable to genuine signal movement — the only part actually worth paying costs for. The fix for the noise component is coarser ranking or a no-trade band; the fix for the mechanical component is loosening the rebalance-to-target rule; neither fix touches the signal itself.