ForecastComparison.compute#

ForecastComparison.compute(*, horizon=1)[source]#

Run the unconditional and conditional tests.

Forms \(d_t\), refuses a numerically constant differential, and computes the Bartlett long-run variance through horizon - 1 lags for the Diebold-Mariano ratio; scales that ratio by the HLN factor \(\sqrt{(T + 1 - 2h + h(h-1)/T)/T}\) for the \(t_{T-1}\) reading; and builds the Giacomini-White instruments – a constant and \(d_{t-h}\) – for the conditional Wald statistic. Deterministic: nothing here draws.

Parameters:

horizon (int) – The forecast horizon behind the losses; the differential’s variance sums autocovariances through horizon - 1 lags with Bartlett weights, since an h-step error is mechanically an MA(h - 1).

Return type:

ForecastComparisonResult

Returns:

The ForecastComparisonResult.

Raises:

  • SpecificationError – If the horizon is not positive or exceeds what the window can support, or the two series are numerically identical.

  • NumericalError – If the conditional moment matrix degenerates.

Example

>>> import numpy as np
>>> rng = np.random.default_rng(0)
>>> a, b = rng.standard_normal(100) ** 2, rng.standard_normal(100) ** 2
>>> verdict = ForecastComparison(a, b).compute(horizon=4)
>>> verdict.horizon, verdict.nobs
(4, 100)
>>> bool(abs(verdict.hln_statistic) < abs(verdict.dm_statistic))
True
>>> ForecastComparison(a, a).compute()
Traceback (most recent call last):
    ...
cultivars.exceptions.SpecificationError: the two loss series are numerically ...