Calibrator

Least-squares calibration over a recorded objective.

You supply a recording that reads the parameters by name (rec.input("alpha", ...)), builds the sum of squared residuals against the market, and calls rec.output(sumSq) once. The objective is compiled to one kernel; every iteration is a setInput + one adjoint sweep, so the gradient is exact and costs one extra sweep regardless of the parameter count. A box-projected L-BFGS drives it.

Calibrator.Result r = Calibrator.of(rec -> {
        SDouble alpha = rec.input("alpha", 0.2);
        SDouble rho   = rec.input("rho",  0.0);
        SDouble nu    = rec.input("nu",   0.3);
        SDouble beta  = rec.constant(0.5);
        SDouble sse = rec.constant(0.0);
        for (Quote q : quotes) {
          SDouble model = SabrHagan.blackVol(rec, alpha, beta, rho, nu, F, q.strike(), T);
          SDouble d = model.sub(q.vol());
          sse = sse.add(d.mul(d));
        }
        rec.output(sse);
      })
      .parameter("alpha", 0.2, 1e-4, 2.0)
      .parameter("rho",   0.0, -0.999, 0.999)
      .parameter("nu",    0.3, 1e-4, 5.0)
      .solve();

View source ↗Full module reference on javadoc.io ↗

Methods

static Calibrator of(Consumer<AadRecorder> objective)

Minimise a recorded scalar objective (usually a sum of squared residuals) with L-BFGS.

static Calibrator leastSquares(MultiOutput.Measures residuals)

Fit a recorded residual vector with Levenberg-Marquardt. The body records the parameters by name and returns the named residuals model_i - market_i; the Jacobian each iteration comes from one MultiOutput evaluation (1 + N adjoint sweeps).

Calibrator scenarios(long n)

Scenario count for a Monte-Carlo objective; the default 1 is a deterministic one.

Calibrator seed(long s)

Calibrator parameter(String name, double initial, double lo, double hi)

Calibrator maxIterations(int n)

Calibrator tolerance(double t)

Calibrator on(String engine)

Result solve()