Alpha Decay

Alpha decay is the entropy of quant trading: every edge erodes, and the only questions are how fast and why. A signal has two clocks running against it, a fast clock (the information decays into prices within days or weeks, setting how urgently you must trade) and a slow clock (the edge is arbitraged away over years as others discover it). Managing both is the difference between a strategy that compounds and one that quietly dies.

The fast clock: signal half-life

A predictive signal does not pay off instantly and then vanish; its information content decays as the market incorporates it. Model the predictive power, say the Signal Construction information coefficient at horizon $h$, as exponential decay:

$\text{IC}(h) = \text{IC}_0\, e^{-\lambda h},$

with half-life $t_{1/2} = \ln 2 / \lambda$. A post-earnings-drift signal might have a half-life of weeks; a fast order-flow signal, minutes. The half-life dictates urgency: a signal that decays in three days is worthless if your positions turn over monthly, because the alpha is gone before you trade. Equivalently, if you hold a position with decaying alpha, the marginal alpha you still earn shrinks each day, so there is an optimal holding period past which the position is no longer worth its risk and cost.

The slow clock: crowding and post-publication decay

The second, longer decay is competitive. As more capital learns a signal, trading on it moves prices toward fair value and compresses the very mispricing it exploited, Market Efficiency (The EMH) reasserting itself. McLean and Pontiff measured this directly for 97 published anomalies and found returns fall by roughly 26% out-of-sample (statistical shrinkage, the backtest was partly luck) and by a further ~32% post-publication (genuine arbitrage, investors trade the newly-public signal). Their decomposition is the cleanest evidence that publicity itself destroys alpha, and it maps onto the Factor Investing risk-vs-anomaly distinction: true risk premia should not decay after publication (you cannot arbitrage a risk), while behavioral anomalies should, and mostly do. Crowding also raises correlation among arbitrageurs, so decayed factors are more prone to synchronized deleveraging crashes (the 2007 quant quake).

The turnover-decay-cost tradeoff

Here is the central tension. A fast-decaying signal wants to be traded quickly to capture its edge before it dies, but fast trading means high turnover, and turnover costs money (spread, impact; see Transaction Costs). The net alpha of a signal is

$\text{Net alpha} = \underbrace{\text{gross alpha captured}}_{\text{rises with trading speed}} - \underbrace{\text{cost} \times \text{turnover}}_{\text{rises with trading speed}}.$

Trade too slowly and the alpha decays away uncaptured; trade too fast and costs eat it. The optimum is an interior trading rate, not a binary "in or out." Gârleanu and Pedersen solved this formally: with quadratic costs and mean-reverting signals, the optimal policy is to trade partway toward the target each period, a proportional adjustment whose speed depends on the ratio of alpha decay to trading cost. The intuition is exactly the fast/slow tension:

$\text{trade rate} \propto \sqrt{\frac{\text{signal mean-reversion (decay) speed}}{\text{transaction cost}}}.$

A fast signal is worth chasing hard; a slow signal in a high-cost name should be tracked lazily. This is why the same raw idea is a great strategy in liquid large-caps and a loser in micro-caps, the cost term flips the sign of net alpha.

Worked example: does the signal survive costs?

A signal has gross alpha of 8%/year but requires 400% annual one-way turnover (fully rotating the book every quarter) to capture it. Round-trip costs are 20 bps. With 400% one-way turnover the book does 800% of one-way trades per year across entries and exits, so cost drag is $8.0 \times 0.20\% = 1.6\%$/year. Net alpha $\approx 8\% - 1.6\% = 6.4\%$, still healthy. Now suppose crowding halves the gross alpha to 4% while turnover (and thus 1.6% cost) is unchanged: net alpha collapses to 2.4%, and after fees and slippage the strategy may be dead. The lesson: costs are roughly fixed while alpha decays, so a signal's net-of-cost life is much shorter than its gross-of-cost life, the strategy dies long before the gross edge reaches zero.

Failure modes

• Backtest alpha that was never real. The out-of-sample 26% haircut is the average; an overfit signal (Overfitting) can lose all of it. Deflate backtested Sharpe for the number of trials.
• Ignoring the fast clock. A beautiful monthly backtest of a three-day signal is a fiction, you cannot hold long enough to realize it, or you pay ruinous turnover to try.
• Crowding you cannot see. You rarely observe others entering your trade until the drawdown; watch valuation spreads, factor correlations, and your own slippage for early warning.
• Re-optimizing on the decay. Chasing a fading signal by refitting parameters usually just fits noise and accelerates the loss.

In interviews

Separate the two decays cleanly: the fast decay (signal half-life, how quickly information flows into price, setting how urgently you trade) and the slow decay (crowding/post-publication, how the edge is arbitraged away, quantified by McLean-Pontiff's ~26% out-of-sample + ~32% post-publication shrinkage). The analytical centerpiece is the turnover-decay-cost tradeoff: fast signals must be traded fast, but turnover costs money, so net alpha is maximized at an interior trading rate, cite Gârleanu-Pedersen's "trade partway to the target" result. A strong closer: because costs stay roughly constant as gross alpha decays, the net edge disappears well before the gross edge does, which is why live strategies die suddenly. See Transaction Costs and Signal Construction for the two sides of the tradeoff.