Short-Horizon ML Alpha Under Cost Constraints

A short-horizon model that predicts the next few minutes of price movement with a small but real edge sounds like a strong result — until the trading costs required to act on that edge at high frequency are subtracted. At short horizons, edges are typically small in magnitude and trades are frequent, so the ratio of expected gain per trade to cost per trade is the number that actually decides whether the strategy is viable, not the raw prediction accuracy alone.

Why the gross-to-net gap is so much larger here

A monthly-horizon strategy might expect a few percent of expected return per position and pay a fraction of a percent in costs — costs are a rounding error. A model predicting price direction over the next minute might have an edge measured in a few basis points, while the bid-ask spread and market impact of trading alone can also run to a few basis points. At that scale, whether the strategy is profitable at all often comes down to modeling execution costs as carefully as the prediction itself, because a small error in either estimate can flip the sign of expected net profit.

$$\text{net edge per trade} = \text{gross edge} - \text{spread cost} - \text{market impact}$$

In words: what's actually captured on each trade is the predicted edge minus what it costs to cross the spread and minus the price impact of the trade's own size — and at short horizons all three terms can be of comparable magnitude.

Market impact typically grows faster than linearly with trade size, which is why this cost term can dominate at short horizons: doubling the size traded more than doubles the impact cost, eating into the edge disproportionately as a strategy tries to scale up.

Worked example

A model predicts 60-second-ahead returns with an average absolute edge of 3 basis points on signals it acts on, trading a liquid stock with a typical bid-ask spread of 2 basis points (round trip) and estimated market impact of 1 basis point for the size traded.

1. Gross edge. 3 basis points expected per trade, based on the model's historical accuracy.
2. Cost per round trip. 2 basis points spread plus 1 basis point impact equals 3 basis points total cost.
3. Net result. Expected net edge is $3 - 3 = 0$ basis points — the strategy is exactly at breakeven before accounting for any estimation error, meaning in practice it is more likely to lose money once real-world costs run even slightly higher than the clean estimate, or the model's edge is slightly weaker live than in the backtest.

This is a common outcome for short-horizon strategies: an apparently real signal turns out to sit right at the cost boundary, and only strategies with either genuinely larger edges or access to cheaper execution (rebates, internalized flow, lower-impact order types) clear it comfortably.

What this means in practice

Firms running short-horizon ML strategies typically only trade the subset of signals whose predicted edge clears a cost threshold with margin, rather than trading every signal the model produces — effectively raising the bar so that estimation error in either the edge or the cost model doesn't flip a marginal trade from profitable to unprofitable. Execution quality (getting filled at good prices, minimizing impact) becomes as much a source of edge as the prediction model itself.