Skip to content

Constant-parameter GSR process bindings

C, Go and Python expose the existing standalone GsrProcess. This is a one-dimensional Gaussian state process under a forward measure, not a fitted short-rate model, calibration engine, curve or pricing engine.

Construction and domain

  • Python: itofin.processes.GsrProcess(times, vols, reversions, t_measure).
  • Go: session.NewGsrProcess(times, vols, reversions, tMeasure).
  • C: itofin_gsr_process_new copies three pointer/count arrays.
  • Volatility count must equal grid-time count plus one. Reversion count must be one or equal volatility count. All repeated volatility and reversion values must be identical. Varying parameter grids are unsupported.
  • Grid times must be finite, nonnegative and strictly increasing. Volatility is finite and nonnegative. Reversion is finite and may be negative or zero.
  • The forward-measure horizon is finite and nonnegative. Grid boundaries do not constrain horizon updates. Zero horizon and zero volatility are valid.
  • Query times and step endpoints must be within [0,T]. States and Gaussian draws must be finite; step durations must be finite and nonnegative.
  • Date conversion is not supported. Times are process years.

Queries and horizon changes

The facades expose x0, sigma(t), reversion(t), y(t), g(t,end), drift, instantaneous diffusion, conditional expectation, standard deviation, variance and evolution with one standard Gaussian draw. The initial state is zero. The affine kernel requires ordered absolute endpoints, not a duration. Zero reversion uses the exact continuous limit, without a small-reversion cutoff.

Python forward_measure_time() and set_forward_measure_time(horizon), and Go ForwardMeasureTime() and SetForwardMeasureTime(horizon), access and update the horizon through the core notifying setter. A valid update immediately changes drift and transition means. An invalid update preserves the previous horizon. The process has no exposed observer-registration API.

The existing Rust process intentionally differs from QuantLib 1.43 in two edge cases: it does not apply a small-reversion cutoff, and horizon changes update transition means as well as drift. QuantLib 1.43 retains its original transition-mean horizon after the inherited setter changes the reported horizon. These bindings preserve Rust behavior, not that stale-horizon defect.

C uses itofin_gsr_process_query: kinds 0 initial state, 1 horizon, 2 sigma, 3 reversion, 4 y, 5 G, 6 drift, 7 diffusion, 8 expectation, 9 standard deviation, 10 variance and 11 evolve. Kind 5 treats its dt argument as the absolute end time. Unused arguments are ignored. Scalar and handle outputs are written only on success. itofin_gsr_process_set_forward_measure_time updates the horizon.

Ownership and errors

Construction copies inputs, so subsequent caller-array changes have no effect. C handles belong to their creating context and OS thread and must be released. Go methods are serialized by the owning session and support concurrent callers; close both process and session explicitly. Python objects are unsendable and retain their own process state. Native domain failures propagate as binding errors. Unrepresentable nonfinite results are rejected rather than returned.

Examples

from itofin.processes import GsrProcess

p = GsrProcess([], [0.03], [0.0], 10.0)
mean = p.expectation(2.0, 0.1, 0.8)
p.set_forward_measure_time(5.0)
changed_mean = p.expectation(2.0, 0.1, 0.8)
assert mean != changed_mean

The Go example is in sdk/go/examples/gsr_process. These additive bindings do not extend the supported core model or varying-grid scope.