Class JumpRobustVolatility

java.lang.Object
com.quantfinlib.microstructure.JumpRobustVolatility
public final class JumpRobustVolatility extends Object
Jump-robust streaming volatility. A squared-return estimator (the one inside SignalEngine) cannot tell a news gap from diffusion: one headline print enters as r² and reads as a volatility regime shift for the estimator's whole memory. Bipower variation (Barndorff-Nielsen invalid input: '&' Shephard, 2004) fixes this with a beautifully simple trick: use the product of CONSECUTIVE absolute returns, (π/2)·|rₜ|·|rₜ₋₁|, instead of r². Diffusion moves both factors together, so the product estimates the same σ²; a single jump inflates only ONE factor of two neighboring products instead of one whole squared term — its weight in the estimate collapses.

Both estimators run side by side on time-decayed rates per second: volPerSqrtSecond() is the jump-robust (bipower) volatility — the one to feed VolatilityCurve and any model that should read regimes, not headlines — while rawVolPerSqrtSecond() is the squared-return volatility, and jumpFraction() is the share of raw variance the robust estimator attributes to jumps (1 − BV/RV, clamped to [0,1]).

Gap discipline: a non-finite return or non-positive Δt drops the sample AND resets the consecutive-return pairing — multiplying across a feed gap would pair returns that were never neighbors. The first return after a gap therefore updates only the raw estimator; the bipower leg resumes one sample later. Irregular sampling is handled exactly: the two-return product is normalized by √(Δtₜ·Δtₜ₋₁) (each |r| scales with √ of ITS OWN interval), so event-time feeds — where activity accelerates precisely when volatility bursts — do not bias the estimator. Zero allocation per event, single writer, one instance per symbol.

  • Constructor Summary

    Constructors
    Constructor
    Description
    JumpRobustVolatility()
    10-second half-life.
    JumpRobustVolatility(long halfLifeNanos)
     

  • Method Summary

    Modifier and Type
    Method
    Description
    double
    jumpFraction()
    The share of raw variance attributed to jumps: clamp(1 − bipower/raw, 0, 1).
    void
    onReturn(double ret, long dtNanos)
    One return observation: the relative mid change over the elapsed dtNanos.
    double
    rawVolPerSqrtSecond()
    The plain squared-return volatility (jumps and all), per √second.
    long
    samples()
     
    double
    volPerSqrtSecond()
    The jump-robust volatility, as return per √second — the diffusion component, with jumps down-weighted. 0 until two consecutive valid returns exist.

    Methods inherited from class java.lang.Object

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

  • Constructor Details

    • JumpRobustVolatility

      public JumpRobustVolatility(long halfLifeNanos)
      Parameters:
      halfLifeNanos - decay half-life, e.g. 10s = 10_000_000_000L

    • JumpRobustVolatility

      public JumpRobustVolatility()
      10-second half-life.

  • Method Details

    • onReturn

      public void onReturn(double ret, long dtNanos)
      One return observation: the relative mid change over the elapsed dtNanos. Non-finite returns or non-positive gaps drop the sample and break the pairing (see class doc).

    • volPerSqrtSecond

      public double volPerSqrtSecond()
      The jump-robust volatility, as return per √second — the diffusion component, with jumps down-weighted. 0 until two consecutive valid returns exist.

    • rawVolPerSqrtSecond

      public double rawVolPerSqrtSecond()
      The plain squared-return volatility (jumps and all), per √second.

    • jumpFraction

      public double jumpFraction()
      The share of raw variance attributed to jumps: clamp(1 − bipower/raw, 0, 1). Near 0 in pure diffusion, spikes after a discontinuous move, and decays back as the jump washes out of the raw estimator's memory. 0 while either estimator is unlearned.

    • samples

      public long samples()