Choosing a Signal Horizon

A signal has to be tested against a return measured over some window — next day, next week, next month — and that choice is often made almost by default, borrowed from whatever the last project used. It shouldn't be. The horizon is a hypothesis about how quickly the market absorbs the kind of information the signal captures, and testing at the wrong horizon can make a real signal look worthless, or a coincidental one look powerful.

Horizon as a claim about the mechanism

A signal built from breaking news should predict returns over minutes to a few days — if the information takes a month to get into price, either the market is astonishingly slow or the "signal" is not really about the news. A signal built from a structural, slow-moving mispricing (say, a spin-off that institutional index funds haven't yet been allowed to buy) might not show any edge over a week but a clear one over two or three months, because the mechanism it is capturing operates on that timescale. Testing a fast-mechanism signal at a slow horizon, or vice versa, is testing the wrong hypothesis and will produce a null result that says nothing about whether the underlying idea is right.

The all-too-common mistake

Researchers sometimes scan a signal against ten or twenty candidate horizons and report whichever one had the best in-sample information coefficient. This is horizon-mining, a close cousin of p-hacking, and it produces results that look sharp and specific but rarely replicate out of sample, because the "best" horizon in the training window is partly luck. The horizon should be picked before looking at performance, based on a stated belief about how fast the mechanism operates, and then tested — not searched for after the fact.

How the signal is meant to work	Horizon that makes sense to test
Reaction to breaking news / order flow	Minutes to a few days
Earnings surprise, analyst revision	Days to a few weeks
Slow structural mispricing (spin-offs, index flows)	Weeks to a few months
Valuation / fundamental mean reversion	Months to a year or more