Rolling ICIR: measuring the persistence of signal ranking

A signal can record a positive information coefficientInformation coefficient (IC)The cross-sectional correlation between a signal score and a subsequent return. Rank IC uses ranked values and measures whether the signal orders securities correctly.Open glossary entry → on one date by chance. The allocation problem is concerned with a more demanding question: has the signal ranked subsequent returns consistently over a recent sample?

Rolling ICIR summarizes that recent history. It rewards a positive average IC and penalizes variation in IC through time.

Definition

Let \(\operatorname{IC}_{i,t}\) be the cross-sectional information coefficient for signal \(i\) at observation time \(t\). For a trailing window of \(T\) observations, define

\[\overline{\operatorname{IC}}_{i,t} = \frac{1}{T} \sum_{s=t-T+1}^{t} \operatorname{IC}_{i,s}\]

and the sample standard deviation

\[s_{i,t} = \sqrt{ \frac{1}{T-1} \sum_{s=t-T+1}^{t} \left( \operatorname{IC}_{i,s} - \overline{\operatorname{IC}}_{i,t} \right)^2 }.\]

The annualized rolling ICIR is

\[\operatorname{ICIR}_{i,t} = \sqrt{A}\, \frac{ \overline{\operatorname{IC}}_{i,t} }{ s_{i,t} },\]

where \(A\) is the annualization convention. StrategyNet uses \(A=252\) for daily IC observations.

What the ratio adds

Mean IC alone does not distinguish between a stable signal and an erratic one. Suppose two signals both have a mean daily IC of \(0.02\):

Signal A receives the higher ICIR because its ranking performance varies less through time. The statistic does not claim that Signal A will earn a particular return. It states that its observed mean IC has been more stable relative to the variability of its IC series.

Why the estimate is rolling

A full-history ICIR can combine periods in which a signal represented different economic conditions. A trailing window allows the estimate to respond when:

• a signal becomes crowded;
• its underlying economic relationship weakens;
• the investment universe changes;
• a market regime changes the relevant forecast horizon; or
• data quality or coverage changes.

The cost of responsiveness is sampling variation. A 20-day estimate can change substantially when one observation enters or leaves the window. A 252-day estimate changes more slowly but may retain evidence from a regime that is no longer relevant.

Window length is therefore part of the model specification. It should be tested out of sample rather than selected from the period used to report performance.

Application in allocation

For each FMP candidate, the current allocator maps rolling ICIR to an expected-return score:

\[\alpha_{i,t} = c\, \operatorname{ICIR}_{i,t},\]

with \(c=0.01\) in the current service. This places FMPs on a common allocation scale before covariance and portfolio constraints are applied.

The mapping has a practical consequence. A sleeve with strong recent returns but inconsistent ranking evidence does not automatically receive a large expected return. The score is tied to the stability of the underlying cross-sectional forecast rather than to the last realized portfolio return.

The score is then used in the nominal objective

\[\underset{w}{\operatorname{maximize}} \quad \alpha_t^{\mathsf T}w - \frac{\lambda}{2} w^{\mathsf T}\Sigma_t w,\]

or in the corresponding worst-case and CVaR extensions described in robust portfolio optimization.

Interpretation limits

Rolling ICIR is not a Sharpe ratio. The numerator is a mean cross-sectional correlation, not a portfolio return, and the denominator is the variability of that correlation through time.

The conventional \(\sqrt{A}\) scaling also does not make ICIR a formal independent-observation test statistic. Serial dependence, overlapping return labels, changing universes, and common market shocks can all reduce the effective sample size.

A complete report should state:

• IC convention: rank or Pearson;
• forecast horizon and signal lag;
• trailing window length;
• annualization convention;
• minimum cross-sectional sample size;
• handling of missing values and ties; and
• whether return labels overlap.

Practical diagnostics

A rolling ICIR series should be reviewed alongside its components. At minimum, plot:

1. daily IC;
2. rolling mean IC;
3. rolling IC standard deviation;
4. rolling ICIR; and
5. the number of securities in each cross-section.

A rising ICIR caused by a stable positive mean is different from one caused by a temporarily collapsing denominator. Inspecting the components prevents the ratio from obscuring that distinction.