Monitoring Live Alpha Model Decay

Every alpha model's edge fades once it is live — other traders find the same pattern, the market microstructure that produced it shifts, or the world simply changes. The danger is not that decay happens; it's that it's usually noticed only after the P&L has already bled for months, because P&L is noisy and a bad month looks identical to normal variance until it isn't. Monitoring live alpha model decay means tracking a signal's quality, not just its P&L, continuously enough to catch the fade while it is still small.

The analogy: a car's tire tread, not just its stopping distance

You don't wait for a car to fail to stop in time to learn its tires are worn — you check tread depth directly, because tread wears down gradually and measurably long before a genuine stopping failure occurs. A trading signal's "tread" is its statistical relationship to the returns it's meant to predict: how strongly it ranks stocks, how consistent that ranking is week to week. Watching that relationship decay is catching worn tread; waiting for a bad month of P&L is waiting for the skid.

Worked example 1: tracking rolling information coefficient

The information coefficient (IC) is the correlation between a signal's predicted ranking and the return that actually followed. In the model's backtest, the average monthly IC was $0.06$. Tracking it live on a rolling 6-month window:

$$\text{IC}_{\text{months 1–6}} = 0.055, \quad \text{IC}_{\text{months 7–12}} = 0.041, \quad \text{IC}_{\text{months 13–18}} = 0.019$$

Each window is individually noisy — a single month's IC can be negative even for a genuinely good signal — but the rolling average trending from $0.055$ down to $0.019$ over three windows is a decay pattern, not noise, because it is monotonic across increasingly long, overlapping samples rather than one bad blip.

Worked example 2: a decay half-life estimate

Fitting an exponential decay to the rolling IC series above, $\text{IC}(t) = \text{IC}_0 \, e^{-\lambda t}$, with $t$ in months and $\text{IC}_0 = 0.055$:

$$0.019 = 0.055 \, e^{-\lambda \times 15} \implies \lambda = \frac{\ln(0.055/0.019)}{15} \approx 0.071 \text{ per month}$$

The half-life is $\ln(2)/\lambda \approx 9.8$ months — meaning if the current trend continues, the signal's predictive power will have halved again roughly every 10 months from here. That number turns "the signal feels weaker lately" into a concrete, plannable estimate of how much runway is left before the model needs replacing or retiring.

The decay curve fit in worked example 2 is a special case of exponential decay generally. Adjust the shape below to see how the same family of curves can decay slowly (a long half-life, plenty of runway left) or quickly (a short half-life, a signal near the end of its useful life).

What this means in practice

Alpha research desks track rolling IC (and related turnover, hit-rate, and factor-exposure metrics) on every live signal on a fixed cadence, with pre-agreed thresholds for when a declining trend triggers a review rather than a shrug. A model flagged for decay is not automatically killed — the review checks whether the decay is genuine signal fade or a data pipeline issue, a regime the signal is known to struggle in, or crowding from other traders — before deciding to retrain, resize, or retire it.