Class AlphaEnsemble

java.lang.Object
com.quantfinlib.microstructure.AlphaEnsemble
public final class AlphaEnsemble extends Object
IC-weighted alpha ensemble — the layer above the individual signals. A desk rarely trades one alpha; it blends several ( SignalEngine.alpha, OnlineAlphaLearner.normalizedPrediction, a lead-lag echo, a bespoke signal), and the blending question is the same honesty problem the learner solves for its weights: how much should each component be trusted, on evidence it could not have memorized?

The ensemble runs one prequential IC per component: each interval, onObservation(double[], double) scores the component values snapshotted at the PREVIOUS call against the return just realized (values now would contain the move — the same nowcast trap OnlineAlphaLearner closes), then snapshots the current values for the next round. combined(double[]) weights each component by max(0, IC) and the weights ARE the sizing — deliberately NOT renormalized to sum to 1, because renormalizing would let a lone component with IC 0.01 emit at full strength: a barely-trusted blend must be a barely-sized signal. A component that has not demonstrated live predictive power gets zero weight, and while the track record spans less than one IC memory the ensemble emits 0 outright. Output is clamped to [-1, 1]: drop-in for BenchmarkExecutor.MarketState.alpha.

Same caveat as every learned signal here: the live IC is a tripwire, not a validation — run any blend you intend to trade through the alpha package's walk-forward machinery. Components must be dimensionless (~[-1, 1]); non-finite inputs are handled per component (see onObservation(double[], double)). Zero allocation per event (caller-owned arrays), single writer, one instance per symbol. The IC evidence persists via persist.Checkpoint — restored trust is earned trust.

  • Constructor Summary

    Constructors
    Constructor
    Description
    AlphaEnsemble(int components)
    1% IC weight (≈ a few-hundred-observation memory).
    AlphaEnsemble(int components, double icAlpha)
     

  • Method Summary

    Modifier and Type
    Method
    Description
    double
    combined(double[] values)
    The blended alpha in [-1, 1]: clamp(Σ max(0, ICᶜ) × valueᶜ).
    double
    componentIC(int c)
    The prequential (out-of-sample) IC of one component — the trust diagnostic per signal. 0 before enough variance exists.
    int
    components()
     
    void
    onObservation(double[] values, double realizedReturn)
    One interval: the current component values and the return realized since the previous call.
    void
    readState(DataInput in)
    Restores the IC evidence; the snapshot (intraday alignment state) resets.
    long
    samples()
     
    void
    writeState(DataOutput out)
    Persists the per-component IC evidence — see persist.Checkpoint.

    Methods inherited from class java.lang.Object

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

  • Constructor Details

    • AlphaEnsemble

      public AlphaEnsemble(int components, double icAlpha)
      Parameters:
      components - number of component signals (fixed order — the caller owns the mapping)
      icAlpha - EWMA weight of the IC statistics, e.g. 0.01

    • AlphaEnsemble

      public AlphaEnsemble(int components)
      1% IC weight (≈ a few-hundred-observation memory).

  • Method Details

    • onObservation

      public void onObservation(double[] values, double realizedReturn)
      One interval: the current component values and the return realized since the previous call. Scores the PREVIOUS snapshot against the return (honest alignment), then snapshots values. The first call only snapshots. Non-finite handling is per-component: a NaN component skips ITS scoring (each component's moments stay conditioned on exactly the returns its covariance saw) while finite siblings still score — but an observation where NOTHING scored (NaN return, or every snapshot value non-finite) does not count toward the track record: the gate must never open on evidence that scored nothing.

    • combined

      public double combined(double[] values)
      The blended alpha in [-1, 1]: clamp(Σ max(0, ICᶜ) × valueᶜ). The IC weights are the SIZE of the signal, not just its mix (see the class doc for why they are not renormalized). 0 while the track record spans less than one IC memory or no component has a positive IC — an unproven blend is silent, exactly like the learner it sits above.

    • componentIC

      public double componentIC(int c)
      The prequential (out-of-sample) IC of one component — the trust diagnostic per signal. 0 before enough variance exists.

    • components

      public int components()

    • samples

      public long samples()

    • writeState

      public void writeState(DataOutput out) throws IOException
      Persists the per-component IC evidence — see persist.Checkpoint.
      Throws:
      IOException

    • readState

      public void readState(DataInput in) throws IOException
      Restores the IC evidence; the snapshot (intraday alignment state) resets. Throws on a component-count or version mismatch.
      Throws:
      IOException