Skip to content

Sensitivity

deepbullwhip.sensitivity.compute_sensitivity(model, X_test, perturbation=0.01, lag_column=0)

Finite-difference forecast sensitivity lambda_f.

Parameters:

Name Type Description Default
model object with .predict(X) method
required
X_test (array, shape(n_samples, n_features))
required
perturbation float

Relative perturbation size.

 0.01
lag_column int

Column index to perturb (default 0 = lag_1).

 0

Returns:

Type Description
(mean_sensitivity, std_sensitivity)
Source code in deepbullwhip/sensitivity.py 
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
def compute_sensitivity(
    model,
    X_test: np.ndarray,
    perturbation: float = 0.01,
    lag_column: int = 0,
) -> tuple[float, float]:
    """Finite-difference forecast sensitivity lambda_f.

    Parameters
    ----------
    model : object with .predict(X) method
    X_test : array, shape (n_samples, n_features)
    perturbation : float
        Relative perturbation size.
    lag_column : int
        Column index to perturb (default 0 = lag_1).

    Returns
    -------
    (mean_sensitivity, std_sensitivity)
    """
    base_pred = model.predict(X_test)
    X_perturbed = X_test.copy()
    delta = perturbation * np.abs(X_test[:, lag_column]).mean()
    if delta == 0:
        return 0.0, 0.0
    X_perturbed[:, lag_column] += delta
    perturbed_pred = model.predict(X_perturbed)
    sensitivities = np.abs(perturbed_pred - base_pred) / delta
    return float(sensitivities.mean()), float(sensitivities.std())