When a Good Forecast Is Untradeable
A signal can be statistically real and still be worthless to trade, because the cost of acting on it, the speed it decays at, or the size it works at rules out any way to actually capture it.
A researcher can build a model that predicts next-hour returns with genuine, statistically significant skill and still end up with nothing tradeable. The gap between "this forecast is real" and "this forecast is worth trading" comes down to three things a backtest's raw accuracy number does not show: how fast the signal's edge decays after it appears, how much it costs in spread and market impact to act on it, and how much capital can actually be deployed against it before the trading itself erodes the edge.
A signal that predicts a 10 basis point move but decays within two minutes may require trading so quickly and so often that commissions and slippage consume the entire edge before it's captured. A signal that only works reliably on a handful of thinly traded small-cap names might be real but simply too small to matter at any meaningful fund size — this is the capacity problem. And a signal that requires knowing a data point before it is actually available in production, even by a few seconds, was never tradeable in the first place; it was a subtle form of look-ahead bias hiding inside a real historical correlation.
Statistical significance answers "is this pattern real," but tradeability is a separate question about speed, cost and size — a forecast can pass every statistical test and still be economically worthless once real-world frictions are honestly priced in.
The discipline of checking a promising signal against realistic transaction costs and true available liquidity, before committing research time to it, is what separates ideas that reach production from ideas that only ever looked good on paper.
Further reading
- Grinold & Kahn, Active Portfolio Management