The Half-Life of Alpha: Is Predictive Power Disappearing, or Just Becoming Shorter-Lived?

Every predictive relationship in financial markets is on a clock. A pattern that forecasts returns today is, by its nature, an opportunity that others are paid to find. Once found, it is studied; once studied, it is written up; once written up, it is packaged and sold; once sold, it is traded by everyone who bought it. At each of those steps some of the original predictive power leaks away, and the leak is not gradual and uniform. It arrives in steps, each tied to a moment at which the number of people acting on the same information changed.

This piece proposes a simple vocabulary for that process. We treat the predictive power of a signal as a quantity that decays, measure the decay in event time rather than calendar time, and ask what its half-life looks like across the main families of signals a systematic investor uses. The conclusion is not that alpha is disappearing. It is that alpha has become perishable, that its shelf life varies by two orders of magnitude across signal types, and that the economics of a quantitative research process are set by the rate at which it must replace what expires.

Four Events, Four Steps Down

A signal's life has a recognizable shape. Someone discovers it, usually by accident while looking for something else. It circulates privately, then appears in a working paper, then in a journal. A data vendor notices that the paper is being cited and productizes the input. A few years later the relationship is a standard factor in a commercial risk model, and the capital positioned on it is no longer a handful of desks but a meaningful fraction of the market.

Figure 1 stylizes that sequence for a hypothetical cross-sectional equity signal. The vertical axis is the information coefficient (IC): the correlation between the signal's ranking of securities and their subsequent returns. The horizontal axis is event time, measured in months from discovery. Each of the four events, discovery, publication, commercialization, and widespread adoption, is drawn as a step down in the level around which the IC fluctuates.

Two features of the picture matter more than the exact numbers. First, the steps are of different sizes. Publication removes less than commercialization, and commercialization removes less than adoption, because each step multiplies the capital chasing the same forecast by a larger factor than the last. Second, the signal does not reach zero. What remains after adoption is the part of the relationship that is expensive to trade, that arrives late, or that persists because it compensates for a risk someone rational prefers not to bear. That residue is small, but it is durable, and a great deal of the systematic industry lives on it.

Measuring Decay in Event Time

The natural way to measure how fast a signal decays is to fit an exponential to its IC after the event of interest and read off the half-life: the number of months it takes for the excess predictive power, the part above the long-run floor, to fall by half. Calendar time is the wrong clock for this. A signal discovered in 2004 and one discovered in 2019 sit at different points in their lives in any given calendar year, and averaging across them in calendar time blurs the steps into a gentle slope that no individual signal actually experienced.

In event time the steps sharpen. Figure 2 fits the same exponential model to four stylized signal families, each anchored at its publication date. The families are deliberately broad: a slow fundamental signal that turns over a few times a year; a valuation-based signal; a price-momentum signal at monthly horizons; and a fast microstructure signal at intraday horizons. The parameters are illustrative, but the ordering is the point.^[1]

Why should the ordering be so stable? Because decay speed is set by how easy the signal is to copy and how much capital its capacity can absorb. A fundamental signal that requires reading footnotes and turns over slowly is copied slowly and crowded slowly. A microstructure signal that requires nothing but a fast connection to the exchange is copied by anyone with that connection, and the capacity of the opportunity is exhausted within months. The fast signal was worth more per unit of time while it lasted; the slow signal is worth more over its life. Neither is better. They are different assets with different depreciation schedules.

What the Half-Life Depends On

If decay speed is a function of copyability and capacity, it should be predictable from observable characteristics of the signal before the decay is observed. Figure 3 gathers the characteristics we find most useful into a single table. The columns are the drivers; the rows are the four families from Figure 2; the entries are qualitative, because the point is the pattern, not the coefficients.

Two of the drivers deserve comment. Capacity determines how far a signal's returns fall for a given amount of capital that follows it: a strategy that can absorb a great deal of capital before its own trading moves prices decays slowly even when it is widely known, which is why valuation signals have survived decades of publicity. Copy cost determines how quickly that capital arrives. A signal whose inputs are a public data feed and a regression can be reproduced in an afternoon; one that depends on a proprietary dataset, an unusual cleaning process, or a piece of infrastructure that took years to build cannot.

This is the practical reason quantitative firms invest in data and infrastructure rather than only in ideas. The idea is usually the least defensible part of a signal. The data pipeline that makes the idea testable without look-ahead, and the execution system that makes it tradeable without giving the edge back in costs, are what slow the half-life down.

Shorter-Lived, Not Gone

The pessimistic reading of all this is that alpha is being competed away and that the systematic industry is living on a shrinking residue. We think the evidence supports a different reading. What has changed is the distribution of half-lives, not the existence of predictive power. The tail of long-lived signals has thinned as data has become cheaper and more widely shared; the mass of the distribution has moved toward shorter horizons, where new signals appear faster than they can be exhausted.

The shift has an arithmetic consequence that is easy to state and hard to live with. If the average half-life of a research inventory halves, the rate at which new signals must be found to hold the inventory's total predictive power constant doubles. A firm that found ten durable signals a decade ago and lived on them cannot do so now. It has to find them continuously, which means the research process itself, its throughput, its false-discovery rate, its cost per validated signal, has become the asset. The signals are inventory; the process is the plant.^[2]

What This Means for How We Work

Three practices follow from taking the half-life seriously. The first is to date-stamp every signal at the events that matter, not only discovery but also the first appearance of a comparable idea in public, so that decay can be measured in event time and expectations for a signal's remaining life can be set honestly. The second is to weight signals by their expected remaining life, not only by their current strength: a strong signal with a short remaining life and a weak one with a long remaining life may deserve similar capital. The third is to invest where the half-life is longest, in data that is hard to assemble, in infrastructure that is hard to replicate, and in the discipline of validation that keeps the false-discovery rate low enough for the replacement arithmetic to close.

None of this makes any individual signal last longer. It makes the portfolio of signals, and the organization that produces them, more durable than any of its parts. That, in the end, is the only kind of alpha that has a long half-life.