Class SignalEvaluator

java.lang.Object

com.quantfinlib.alpha.SignalEvaluator

public final class SignalEvaluator
extends Object

Signal evaluation — the metrics that decide whether a factor is worth constructing a portfolio from, computed before any backtest so weak signals die cheaply:

• IC (information coefficient) — Spearman rank correlation between scores at t and forward returns over (t, t+horizon], per evaluation date. Rank (not Pearson) because factor scores have arbitrary units and fat tails; rank IC is invariant to any monotone transform of the signal.
• IR — mean(IC) / std(IC): signal strength per unit of signal inconsistency, the standard Grinold-Kahn quality number (0.05 mean IC with steady sign beats 0.10 that flips).
• t-stat — mean(IC) / (std(IC)/√n): is the mean IC distinguishable from zero given how many dates we observed.
• Hit rate — fraction of (symbol, date) pairs where the score sign called the forward-return sign.
• Turnover — half the L1 change in normalized weights between consecutive evaluation dates: how much trading the signal demands, the denominator of "does the alpha survive costs".
• Factor exposure — mean cross-sectional rank correlation against another factor: a "new" signal that is 0.9 rank-correlated with momentum is momentum.

Evaluation dates step by horizon so forward-return windows don't overlap — overlapping windows inflate the t-stat through serial correlation, the classic way factor research fools itself.

Nested Class Summary

Nested Classes

Modifier and Type	Class	Description
static final record	SignalEvaluator.QuantileReport	Mean forward return per score quantile — the picture behind the IC: meanReturns()[0] is the average forward return of the lowest-scored names, the last entry of the highest-scored, and SignalEvaluator.QuantileReport.spread() is the top-minus-bottom long/short return per period.
static final record	SignalEvaluator.Report	The evaluation scorecard; SignalEvaluator.Report.format() renders it for humans.

Method Summary

Modifier and Type	Method	Description
static SignalEvaluator.Report	evaluate(AlphaContext ctx, AlphaFactor factor, int startIndex, int horizon)	Evaluates a factor over [startIndex, ctx.bars() − horizon), stepping by horizon (non-overlapping forward windows).
static double	factorExposure(AlphaContext ctx, AlphaFactor a, AlphaFactor b, int startIndex, int step)	Mean cross-sectional rank correlation between two factors' scores — how much of factor B is already inside factor A.
static SignalEvaluator.QuantileReport	quantileReturns(AlphaContext ctx, AlphaFactor factor, int startIndex, int horizon, int quantiles)	Buckets each evaluation date's cross-section into quantiles score-ranked groups and averages the forward returns per group, over the same non-overlapping date grid as evaluate(AlphaContext, AlphaFactor, int, int): dates step by horizon from startIndex, a NaN score or NaN forward return drops that (symbol, date) pair, and a date with fewer complete pairs than quantiles contributes to no bucket at all.

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Method Details

evaluate

public static SignalEvaluator.Report evaluate(AlphaContext ctx,
 AlphaFactor factor,
 int startIndex,
 int horizon)

Evaluates a factor over [startIndex, ctx.bars() − horizon), stepping by horizon (non-overlapping forward windows).

Parameters:

horizon - forward-return horizon in bars (also the step)

quantileReturns

public static SignalEvaluator.QuantileReport quantileReturns(AlphaContext ctx,
 AlphaFactor factor,
 int startIndex,
 int horizon,
 int quantiles)

Buckets each evaluation date's cross-section into quantiles score-ranked groups and averages the forward returns per group, over the same non-overlapping date grid as evaluate(AlphaContext, AlphaFactor, int, int): dates step by horizon from startIndex, a NaN score or NaN forward return drops that (symbol, date) pair, and a date with fewer complete pairs than quantiles contributes to no bucket at all. Bucketing is by ascending score rank (rank * quantiles / n), so groups are as equal-sized as the cross-section allows; ties are split by input order at the boundary — quantile membership, unlike the rank IC, is not fully tie-invariant, stated.

Parameters:

quantiles - number of buckets, ≥ 2 (5 = quintiles, 10 = deciles)

factorExposure

public static double factorExposure(AlphaContext ctx,
 AlphaFactor a,
 AlphaFactor b,
 int startIndex,
 int step)

Mean cross-sectional rank correlation between two factors' scores — how much of factor B is already inside factor A. Above ~0.7 the "new" factor adds little beyond the old one.