What Is Alpha Decay? The Mathematical Reason Your Strategy Stops Working

The intuition

A trading strategy generates "alpha" when it predicts something the market hasn't priced in yet. Buy before the price rises, short before it falls, pocket the difference. Simple enough in a backtest, where we're the only one acting on the signal.

In a live market, we're not alone. The moment we trade on a signal, we move the price, however slightly, in the direction that signal predicted. We have contributed to the very price discovery that makes the signal valuable. If enough capital does this — whether from scaling up, or from other managers who found the same signal independently — the market becomes more efficient with respect to that signal. The predictive gap narrows. The alpha decays.

This isn't a flaw in a particular strategy. It's a structural feature of how markets work. And it means every alpha source has a half-life.

The mathematics

Let the alpha of a signal at time $t$ be $\alpha(t)$, defined as the expected excess return from trading it. Under a simple exponential decay model:

$\alpha(t) = \alpha_0 \cdot e^{-\lambda t}$

where $\alpha_0$ is the initial alpha at discovery and $\lambda > 0$ is the decay rate. The half-life of the signal — the time at which expected alpha falls to half its original value — is $t_{1/2} = \ln(2) / \lambda$.

A faster-discovered signal, or one traded by many participants, has a larger $\lambda$. A signal in an illiquid or under-researched market may have a very small $\lambda$ and remain exploitable for years.

The exponential model is a simplification. In practice, decay is often front-loaded: alpha erodes quickly once a strategy attracts attention, then stabilizes at a lower but nonzero level as only the fastest or most efficient implementations remain profitable. Researchers have modeled this with regime-switching processes and with logistic decay curves that flatten rather than approach zero asymptotically.

Capacity and the crowding problem

Alpha decay accelerates with capital. If a strategy manages $Q$ dollars and generates returns $r(Q)$, the marginal return typically decreases as $Q$ rises — more capital means larger trades, more market impact, and faster signal erosion. Formally, for a strategy with market impact function $I(Q)$:

$r(Q) = \alpha - I(Q)$

where $I(Q)$ is increasing and convex. The optimal capital allocation $Q^*$ is where marginal return equals zero:

$\frac{dr}{dQ} = -\frac{dI}{dQ} = 0 \implies Q^* = I'^{-1}(0)$

This gives the strategy's "alpha capacity" — the maximum assets under management before the strategy becomes unprofitable. Above $Q^*$, the fund is paying more in market impact than it earns in signal.

In the ML era, capacity constraints have become more acute. Statistical learning methods applied to public datasets tend to discover the same signals at roughly the same time across different firms. If ten quantitative funds independently train models on the same alternative data source and arrive at similar factor exposures, the effective capital competing on that signal is the sum of all ten funds' allocations — even though no individual manager knows the others are there.

Recent research has begun formalizing this as "certified alpha capacity" thresholds: analytically derived bounds on deployment size and time horizon before measurable decay, rather than waiting to observe the decay empirically in live P&L.

So what does this actually change about how we build strategies?

The short answer: it shifts the question we ask before deploying anything.

A naive backtest workflow asks "did this signal work historically?" Alpha decay means that's the wrong question. The right questions are: at what capital scale does this strategy remain profitable, and for how long? Those are different questions, and they have different answers.

It also changes how we think about portfolio construction. If individual signals have finite half-lives, a fund that bets on a fixed set of exposures is slowly running down a depleting resource. What actually sustains performance over time is a pipeline — continuous research generating new signals as existing ones mature and lose their edge.

Machine learning makes this harder, not easier. A model trained on public data is likely to find the same signals as other models trained on the same data. The crowding isn't intentional or visible — no manager knows how many others are sitting in the same position — but it's systematic. This is part of why the gap between a strategy's backtest performance and its live performance often isn't just overfitting. The backtest modeled the signal in isolation. The live market didn't.

Limitations

The exponential decay model is clean but not always accurate. Some signals decay in bursts rather than smoothly — holding stable until a trigger event (a major fund launch, a widely circulated paper) after which they deteriorate sharply. The capacity model assumes a well-defined market impact function, which is genuinely difficult to estimate for strategies spanning multiple instruments or complex execution schedules. And crowding is inherently unobservable from outside: we can't see how many other funds are running similar strategies until the stress event arrives and they all exit at once.

The framework is a way of thinking, not a precise measurement tool. Used well, it stops us from mistaking a good backtest for a durable strategy.