The half-life of alpha

July 19, 2026 · finance · ml · draft — sample essay, to be replaced

Every signal is a depreciating asset. The interesting question is not whether it decays but whether your machinery turns over faster than your information does.

All models are wrong, but some are useful.

George E. P. Box

A quantitative signal is a claim about the future that the market is in the business of falsifying. The moment a forecast is formed it begins to leak into prices — through your own trading, through everyone else who found the same regularity, through the slow arbitrage of disclosure. What remains is conventionally measured as the decay of the information coefficient: the correlation between forecast and realized return, tracked as the forecast ages.1

The decay is not uniform across families. Flow-driven signals are spent in days; estimate revisions persist for weeks; slower cross-sectional effects like momentum survive for months. The differences are large enough that they should drive system design, not tuning.

Signal decay by family

Information remaining as a forecast ages, three representative families, 2016–2026

% OF PEAK SIGNAL REMAINING02550751000d30d60d90d120dtypical rebalance windowprice momentumearnings revisionsorder-flow imbalance

source: simulated data — sample essay

Three consequences follow. First, a signal's half-life bounds its useful rebalance frequency from below: rebalancing a momentum book daily buys transaction costs and nothing else. Second, half-life bounds capacity from above — a fast signal can only be monetized in the size the market will absorb inside its window. Third, and least appreciated, the half-life must be measured after your own participation, because execution is itself a disclosure.2

What the families look like

familyhorizonhalf-lifecapacitygross ic
order-flow imbalanceintraday–days8dlow0.081
earnings revisionsweeks40dmedium0.043
price momentummonths118dhigh0.027

The table understates how different the engineering is. A fast book is an execution problem wearing a research costume; a slow book is a tax and drift problem wearing one. Treating them as one system with one rebalance cadence quietly averages away the edge of both.

The cost of trading your own decay

Net of costs, the relationship between rebalance frequency and realized Sharpe is a hump, not a slope. Trade too rarely and you hold stale forecasts; trade too often and you pay the spread to refresh information the market already had.

The rebalance hump

Net Sharpe of a blended book by rebalance frequency, after modeled costs

NET SHARPE AFTER COSTS0.00.40.81.20.38daily0.86weekly1.12monthly0.94quarterly0.61annual

source: simulated data — sample essay

The peak sits wherever your slowest surviving signal says it should — which is why the honest version of this chart has to be recomputed every time the signal mix changes, and why it belongs in the deployment checklist rather than in a slide from two years ago.


None of this is exotic. It is bookkeeping — the same discipline as reconciliation, applied to information instead of positions. The systems that survive are the ones that treat alpha as inventory with an expiry date, and build the plumbing to mark it to market.

  1. Throughout, decay is measured on the forecast, not the portfolio: the distinction matters once positions are path-dependent.

  2. This is the quiet argument for capacity-aware backtests: the decay curve you estimate without impact is the curve of a book that doesn’t exist.