Turnover Penalties in Model Training

A model trained to minimize prediction error at every single time step has no idea that changing its mind from one day to the next costs real money in trading fees and market impact. Two models can have nearly identical predictive accuracy on paper, yet one produces a smoothly evolving score that barely needs rebalancing while the other flips its prediction back and forth on noise, generating far more trades to capture the same underlying signal. The second model looks fine in an accuracy table and terrible once transaction costs are subtracted from its live returns.

Building the cost of change into the objective

A standard training objective minimizes prediction error alone, something like mean squared error between predicted and realized returns. A turnover-penalized objective adds a second term that punishes the model for changing its output too much between consecutive periods:

$$\mathcal{L} = \underbrace{\sum_t (\hat{y}_t - y_t)^2}_{\text{prediction error}} \;+\; \lambda \underbrace{\sum_t |\hat{y}_t - \hat{y}_{t-1}|}_{\text{turnover penalty}} .$$

In plain English: the model is still rewarded for predicting well, but it now also pays a price, controlled by $\lambda$, every time its prediction jumps around from one period to the next — so among two models with similar accuracy, training will prefer the one whose predictions evolve more smoothly, because that one incurs a smaller penalty term overall.

Worked example: comparing an unpenalized and a turnover-penalized model

Suppose an unpenalized model's daily score for one stock over a week reads 0.8, −0.3, 0.6, −0.5, 0.7 — swinging in sign almost every day, forcing a position to flip direction repeatedly and pay the bid-ask spread and market impact each time. Adding a turnover penalty with $\lambda = 0.3$ to the training objective, and refitting on the same data, might produce a smoother sequence for the same days: 0.5, 0.3, 0.45, 0.2, 0.5 — same directional view throughout (consistently positive), never crossing zero, requiring only small position adjustments rather than full reversals. If the raw prediction accuracy (correlation with realized returns) only drops from, say, 0.041 to 0.038 between the two models, but the penalized model's annual turnover-related costs fall by several percentage points, the penalized model's net performance — the number that actually matters for a live strategy — comes out ahead despite its slightly "worse" raw accuracy score.

What this means in practice

The penalty weight $\lambda$ is itself a hyperparameter that should be tuned against net-of-cost, not gross, performance — too small and the model still churns, too large and it becomes sluggish, failing to update even when a real change in the underlying relationship justifies a genuine shift in the prediction. This approach is closely related to, and often used alongside, converting model scores into positions with an explicit no-trade band: the penalty smooths the model's raw output, while a no-trade band separately filters small changes at the position-sizing stage. Using both is common; using neither is the most frequent mistake in signals that look strong on paper and disappoint once traded.