DMARIANO Procedure

@DMARIANO does the Diebold-Mariano comparison of forecasts (which must already be computed) from Diebold and Mariano(1995), or (optionally) the Modified DM test from Harvey, Leybourne, Newbold(1997). @GNEWBOLD is a related procedure with does the Granger-Newbold test.

Note well that the Diebold-Mariano test does not apply when the forecasting procedures are "nested" (one is a special case of the other) because the asymptotics collapse if the more restricted model is correct.

@DMariano( options ) actual f1 f2 start end

Parameters

actual	series of actual data
f1	first set of forecasts
f2	second set of forecasts
start, end	range to analyze (common range of f1, f2 and actual by default)

Options

CRITERION=[MSE]/MAE

Chooses whether the error criterion will be squared or absolute errors

LWINDOW=NEWEYWEST/BARTLETT/[TRUNCATED]/PARZEN/QUADRATIC

Chooses the lag window form (NEWEYWEST and BARTLETT are synonyms).

LAGS=number of lags (for most windows) or bandwidth (for LWINDOW=QUADRATIC)

MODIFIED/[NOMODIFIED]

MODIFIED chooses the small sample correction from Harvey, Leybourne and Newbold.

TITLE=title for report ["Diebold-Mariano Forecast Comparison Test"]

[PRINT]/NOPRINT

Variables Defined

%CDSTAT

Test statistic for f1=f2 vs f2 better than f1. Flipping the sign gives the test statistic for f2=f1 vs f1 better than f2.

Sample Output

The P values will (by construction) always add up to one. A small P value indicates that the forecast on the line will be rejected in favor of the other.

Diebold-Mariano Forecast Comparison Test

Forecasts of SPREAD over 1995:04 to 2008:01

Forecast MAE Test Stat P(DM>x)

FORECAST_AR7 0.35514485 0.0241 0.49039

FORECAST_AR2MA17 0.35478985 -0.0241 0.50961