Monitoring Signal Decay in Production

A backtest measures a signal's predictive power once, on history that has already happened, and reports one clean number. Nothing about that number tells you whether the signal still works next Tuesday. Signals decay the way medicine loses potency after its expiration date — not because the chemistry changes overnight, but because the market conditions the signal was tuned to (a pricing inefficiency, a behavioral bias, a piece of information nobody else was using) get arbitraged away, gradually, as more people find it. The label on the bottle told you the strength on the day it was bottled. It says nothing about the strength today.

The dangerous part is that decay is usually gradual and PnL is noisy, so the two stories — "the signal is quietly dying" and "we're just having a bad month" — look identical for a while. Signal decay monitoring is the discipline of measuring the signal's own predictive power directly, on a rolling basis, so you can tell those two stories apart before the PnL forces the answer on you.

The measurement, one symbol at a time

The core quantity is the information coefficient, $IC$ — the correlation, usually rank correlation, between what the signal predicted and what actually happened, measured over some window of predictions. An $IC$ of 0.05 means a weak but real edge; an $IC$ near 0 means the signal is doing no better than guessing. Computed once, over the whole backtest, this is a single summary number. Computed on a rolling window — say, the trailing eight weeks, recomputed every week — it becomes a time series, $IC_t$, that you can watch move.

If a signal is genuinely decaying rather than just noisy, $IC_t$ tends to fall off roughly exponentially, which is worth naming because it lets you put a number on "how fast":

$$IC_t \;\approx\; IC_0 \, e^{-t/\tau},$$

read as: today's information coefficient equals the original information coefficient shrunk by a factor that grows with elapsed time $t$, at a speed set by the half-life $\tau$ (tau) — a smaller $\tau$ means faster decay. This is the same shape as radioactive decay or drug metabolism, and it's a useful shape to fit because a constant percentage loss per period, rather than a constant absolute loss, is what "getting arbitraged away" typically looks like.

Worked example 1: telling decay from a noisy patch

A signal was designed with $IC_0 = 0.06$ in backtest. Live, the rolling eight-week $IC$ prints 0.055, 0.058, 0.041, 0.052, 0.035, 0.019, 0.024, 0.011 over eight consecutive checks. A single low reading — week 3 at 0.041 — is well within normal noise for an IC computed on a few hundred predictions; its standard error is roughly $1/\sqrt{N}$, so with $N=300$ predictions per window that's about 0.058, comfortably larger than the swing you're seeing. But by week 8 the reading (0.011) is a large fraction of that standard error below the trend, and seven of eight readings have fallen in a row. That's not one bad week — it's a trend, and a trend in a shrinking direction is exactly what decay looks like.

Worked example 2: fitting the half-life

Take two of those points to estimate how much runway is left. At $t = 0$ (design time), $IC_0 = 0.06$. At $t = 8$ weeks, $IC_8 = 0.011$. Solving $IC_8 = IC_0 e^{-8/\tau}$ for $\tau$:

$$\tau = \frac{-8}{\ln(0.011/0.06)} = \frac{-8}{\ln(0.183)} = \frac{-8}{-1.70} \approx 4.7 \text{ weeks}.$$

A half-life near five weeks means the signal is losing about half its remaining strength every five weeks — at that rate it's down to roughly 0.006 by week 13 and effectively indistinguishable from noise by week 18. That number turns a vague worry ("it feels weaker") into a decision with a deadline: either re-derive the signal, tighten the universe it's applied to, or plan to retire the strategy on a known schedule rather than finding out from the PnL.

Doing it properly

Choose the rolling window length to balance two costs: too short and every reading is mostly noise; too long and a real decay takes months to become visible. A window covering a few hundred predictions is a reasonable floor for the standard-error arithmetic above to mean anything. Track decay separately for different slices of the universe — a signal often decays first in the most liquid, most-watched names, where competition arrives quickest, and keeps working longer in the corners nobody else is looking at, so a blended, universe-wide $IC$ can hide a decay that's already severe in the segment you actually trade the most size in.

Feed the rolling $IC$ into the same alerting machinery as the rest of the book — see A Monitoring Stack for Live Strategies — so a slow structural decline gets flagged with the same discipline as a broken data feed, rather than being left to the eye test. It also pays to log the half-life estimate itself alongside the raw $IC$ series: a $\tau$ that keeps shrinking on each re-fit is a second, independent confirmation that the decay is accelerating rather than settling into a new, lower-but-stable steady state, which is the difference between "reduce size and keep watching" and "retire the strategy on a fixed schedule." Finally, resist refitting the signal on the most recent data the moment $IC_t$ dips — that risks curve-fitting to the same noise the monitor was built to see past, and turns a legitimate decay signal into another source of overfitting.