Live Information Coefficient Monitoring

A cross-sectional equity signal backtests at an average monthly information coefficient of 0.05 — a modest, entirely plausible edge for a mid-frequency signal on 2,000 names. It launches, and for the first four months live IC averages 0.048, indistinguishable from the backtest. Nobody is tracking it with a proper statistical band, just an eyeballed monthly print, so the fifth month's IC of 0.01 reads as "a quiet month," the sixth month's 0.005 reads the same way, and by the time the P&L has clearly turned, the signal has been dead for two months and the desk is only now asking why.

What the information coefficient is, and why it needs a band

The information coefficient is the cross-sectional correlation, each period, between the signal's forecast and the realized forward return:

$$IC_t = \text{corr}\big(\hat{r}_{i,t}, r_{i,t+1}\big) \quad \text{across all names } i \text{ at time } t.$$

In words: on any given day or month, how well did the signal's ranking of names line up with what actually happened next. A single period's IC is noisy — even a genuinely good signal bounces between, say, −0.05 and +0.15 month to month purely from cross-sectional noise. What matters for monitoring is not any one period's IC but whether a rolling average has drifted outside the band that noise alone would produce around the signal's known historical mean.

Worked example: setting the control band

Backtested mean monthly IC: 0.05, monthly SD across the backtest 0.09 (typical — cross-sectional ICs are noisy even for real signals). For a rolling 6-month average, SE shrinks by $\sqrt{6}$: $0.09/\sqrt{6}\approx0.037$. Two-SE lower control limit: $0.05-2(0.037)\approx-0.024$.

Live rolling 6-month IC: month 4 average 0.048 (fine), month 6 average 0.021 (inside band), month 8 average −0.008 (worth a flag), month 9 average −0.031 (below limit). The check would have flagged this by month 8 or 9, well before an eyeballed monthly print made it obvious around month 11.

Worked example: false alarm from a regime, not decay

A different signal's rolling IC drops from 0.06 to 0.01 over two months, tripping the same rule. Stratified by sector, excluding energy names (which went through a macro-driven repricing that quarter), the remaining universe's rolling IC is still 0.052 — essentially unchanged. The alarm was real in the blended number but attributable to one sector's regime, and the fix is a temporary exclusion, not retiring the signal.

The banding logic above is exactly the reasoning behind the plot's own confidence interval: a rolling mean of noisy monthly draws has a predictable spread under "nothing changed," and a genuine break shows up as an excursion outside that spread — not as any single low reading, which noise alone produces routinely.

What this means in practice

Compute IC every period, maintain a rolling average with a control band derived from the signal's own historical IC volatility, and stratify any flagged drift by sector, size bucket, or region before concluding the signal itself has broken. Pair this with CUSUM Monitoring of Strategy PnL, which catches the same kind of drift from the P&L side and rarely trips at exactly the same moment — agreement between the two is strong evidence, disagreement tells you where to look first.