The signal and its own shadow
A price-prediction signal that looks strong in a backtest is, at the moment it is traded, also a new participant in the market. The very act of acting on the signal moves the price the signal is trying to predict. This is not a footnote; it is the central reason that paper alpha and executable alpha are different quantities.
Three places the gap opens
- Market impact. A backtest assumes you can transact at the printed price. A real order of any meaningful size pays to get filled, and the cost rises with the size of the order relative to the available liquidity.
- Crowding. A pattern that is profitable and known will be traded by more than one participant. As crowding increases, the pattern’s edge compresses, sometimes to zero, sometimes past it.
- Execution lag. A signal generated at the close and traded at the next open is not the same signal. The interval between decision and execution is where a measurable share of paper alpha evaporates.
How we read claims about executable alpha
We look for evidence that the reported numbers were produced after impact, crowding and lag — not before. A claim that a model “would have” made a return is a claim about a counterfactual in which the model did not exist; the moment the model trades, that counterfactual is gone.
The limit, plainly
The distance between a pattern in data and a profit in an account is the distance between a model and a market. The model can describe the pattern honestly and still have nothing to say about whether the pattern can be captured after the costs of capturing it. That distance is the limit, and it does not shrink with more data.