`dm_diebold_mariano` – Diebold-Mariano (1995) equal-predictive-ability test with Newey-West HAC SE.#

Back to equal_predictive_test axis | Back to L6 | Browse all options

Operational op under axis equal_predictive_test, sub-layer L6_A_equal_predictive, layer l6. Standalone callable: mf.functions.dm_test.

Function signature#

mf.functions.dm_test(
    loss_a: np.ndarray,
    loss_b: np.ndarray,
) -> DMTestResult

Parameters#

name	type	default	constraint	description
`loss_a`	`np.ndarray`	—	—	Per-period losses for model A (e.g. squared errors).
`loss_b`	`np.ndarray`	—	—	Per-period losses for model B.

Returns#

DMTestResult — frozen dataclass with fit results.

Attribute	Type	Description
`stat`	`float or None`	DM test statistic
`pvalue`	`float or None`	Two-sided p-value
`decision`	`bool`	Reject H0 at 5%
`n_obs`	`int`	Observations used
`hln_correction`	`bool`	HLN correction applied

Behavior#

Pairwise test of equal expected loss between two forecasts. Implements DM with HLN small-sample correction (Harvey-Leybourne-Newbold 1997) and a configurable HAC kernel (newey_west default, andrews / parzen available). Two-sided alternative tests equality of MSE / MAE losses.

When to use

Pairwise comparison of two non-nested forecasts.

When NOT to use

Nested-model comparisons – use Clark-West (L6.B) instead.

In recipe context#

Set params.equal_predictive_test = "dm_diebold_mariano" in the relevant layer to activate this op within a recipe:

# Layer L6 recipe fragment
params:
  equal_predictive_test: dm_diebold_mariano

References#

macroforecast design Part 3, L6: ‘tests must report (statistic, p-value, kernel, lag) and respect HAC dependence-correction.’
Diebold & Mariano (1995) ‘Comparing Predictive Accuracy’, JBES 13(3): 253-263.
Harvey, Leybourne & Newbold (1997) ‘Testing the equality of prediction mean squared errors’, IJF 13(2): 281-291.

dm_diebold_mariano – Diebold-Mariano (1995) equal-predictive-ability test with Newey-West HAC SE.#