When an ML Signal Becomes a Strategy

A researcher trains a gradient-boosted model that predicts next-week stock returns with an information coefficient of 0.05 — a modest but real edge, confirmed out of sample. It is tempting to call this a strategy and start trading it. It is not one yet. A prediction is a number; a strategy is a full system that turns that number into positions, trades those positions at a real cost, and survives being wrong some of the time.

The gap between a prediction and a position

A raw model output is typically a score — a rank, a probability, an expected return — for each stock, refreshed on some schedule. Converting that into a portfolio requires answering questions the model itself says nothing about: should position size scale linearly with the score, or only trade the top and bottom deciles? How much should a stock's position change day to day as the score updates, given that every change costs money to execute? And does the resulting portfolio, once built, secretly carry a sector or momentum tilt the model absorbed as a side effect of predicting returns well?

Watch how a step size that looks fine on the loss curve can still overshoot — the same idea applies to strategy construction: a signal that predicts well doesn't automatically translate into well-sized, well-timed trades. Getting from "the model is accurate" to "the trades are sensible" takes its own separate tuning.

Worked example

A model produces a daily score for 500 stocks with an IC of 0.04. A naive first pass simply buys the top 50 scores and shorts the bottom 50 every day, resized daily to match the new ranking.

1. Raw signal to naive strategy. Turnover from resizing daily is extremely high — many names enter and leave the top/bottom 50 purely from small score changes, not from any real change in expected return.
2. Cost check. At 10 basis points round-trip cost and near-100% daily turnover on the traded names, the strategy pays away roughly 20% annualized in costs alone — larger than the gross alpha the 0.04 IC was likely to generate.
3. Fix. Adding a "don't trade unless the score change exceeds a threshold" rule and holding positions for a minimum number of days cuts turnover by more than half, and the resulting net-of-cost Sharpe ratio, previously negative, turns modestly positive.

The model didn't change between steps 2 and 3 — only the layer that converts predictions into trades did.

What this means in practice

Teams that treat model accuracy as the finish line consistently overestimate what a signal is worth, because a backtest of the raw prediction quality never confronts trading costs, capacity limits, or unintended risk exposures. The strategy-construction layer — position sizing, cost-aware trading rules, risk neutralization — is often where more research effort goes than into the model itself, precisely because a mediocre model wrapped in disciplined execution frequently beats a brilliant model traded naively.