Source code for quantdata.dysanos.tangent_levy.physical_measure
"""Optional low-dimensional physical-measure drift estimation."""
from __future__ import annotations
import math
from dataclasses import dataclass
import numpy as np
from cdxcore.err import verify_inp
from .constraints import hjm_drift, musiela_drift
from .dynamics import DynamicsFit
[docs]
@dataclass(frozen=True)
class PhysicalMeasureFit:
"""Regularized constant factor market price of risk."""
theta: np.ndarray
ridge: float
residual_rmse: float
target_drift_norm: float
explained_drift_fraction: float
[docs]
def fit_constant_market_price_of_risk(
forward_density_history: np.ndarray,
dynamics: DynamicsFit,
*,
ridge: float = 1.0e-4,
consecutive_mask: np.ndarray | None = None,
) -> PhysicalMeasureFit:
"""Fit ``alpha_P = alpha_Q + sum beta_n theta_n`` to historical mean drift.
This function estimates a physical layer; it does not alter or re-estimate
the HJM-consistent risk-neutral drift stored by the core model.
"""
history = np.asarray(forward_density_history, dtype=np.float64)
verify_inp(
history.ndim == 3
and history.shape[1:] == dynamics.beta.shape[1:]
and history.shape[0] >= 3
and np.all(np.isfinite(history)),
lambda: (
"history must be finite with shape (D,T,X) matching dynamics, got "
f"{history.shape}"
),
)
verify_inp(
math.isfinite(ridge) and ridge >= 0.0,
lambda: f"'ridge' must be finite and nonnegative, got {ridge!r}",
)
changes = np.diff(history, axis=0)
state_before_change = history[:-1]
if consecutive_mask is not None:
consecutive = np.asarray(consecutive_mask, dtype=bool)
verify_inp(
consecutive.shape == (changes.shape[0],),
lambda: (
"'consecutive_mask' must have one value per date transition, got "
f"{consecutive.shape}"
),
)
changes = changes[consecutive]
state_before_change = state_before_change[consecutive]
verify_inp(
changes.shape[0] >= 2,
lambda: (
"physical-drift estimation requires at least two consecutive changes, "
f"got {changes.shape[0]}"
),
)
empirical_drift = np.mean(changes, axis=0) / dynamics.dt
mean_state = np.mean(state_before_change, axis=0)
active_mask = (
np.ones(mean_state.shape, dtype=bool)
if dynamics.active_mask is None
else np.asarray(dynamics.active_mask, dtype=bool)
)
alpha = hjm_drift(
tau=dynamics.tau,
jump_x=dynamics.jump_x,
beta=dynamics.beta,
active_mask=active_mask,
)
q_drift = musiela_drift(
tau=dynamics.tau,
forward_density=mean_state,
hjm_alpha=alpha,
)
target = (empirical_drift - q_drift)[active_mask]
factor_matrix = dynamics.beta[:, active_mask].T
gram = factor_matrix.T @ factor_matrix
theta = np.linalg.solve(
gram + ridge * np.eye(dynamics.num_factors),
factor_matrix.T @ target,
)
residual = target - factor_matrix @ theta
target_norm = float(np.linalg.norm(target))
explained = (
1.0 - float(np.dot(residual, residual)) / float(np.dot(target, target))
if target_norm > 0.0
else 0.0
)
return PhysicalMeasureFit(
theta=theta,
ridge=float(ridge),
residual_rmse=float(np.sqrt(np.mean(np.square(residual)))),
target_drift_norm=target_norm,
explained_drift_fraction=explained,
)