### Table 5. Diebold-Mariano Statistics of Predictive Accuracy

"... In PAGE 22: ... In particular, Table 3 reports the relative rankings and Table 4 contains the criteria values. Finally, Table5 reports the pairwise model comparison statistics based on Diebold-Mariano predictive accuracy tests for all the commodities and their nearbies. Only the results from one step ahead and five step ahead forecasts are reported here, but the results for two to four step ahead forecasts are available upon request.... In PAGE 23: ... Tables 3.1 and 3.2 report the rankings of all models and the values univariate model selection criteria based on one-step amd five-step ahead forecasts. Table5 reports the results from DM test statistics. In all of these tables, only the results from the most recent nearby and most distant nearby futures contracts are reported, for the sake of brevity.... In PAGE 24: ... In other words, all models are useful for predicting the direction of price changes. Entries in Table5 are the Diebold-Mariano statistics. At 10% significance level, all DM statistics suggest accepting the null hypothesis (i.... ..."

### Table 7: Diebold and Mariano Test for VolatilityForecast Performance

"... In PAGE 37: ... Since we are inter- ested in detecting whether ONI manages to reduce persistence in predicted volatility,we will suggest, in the spirit of Granger and Pesaran #281996#29 or of Christo#0Bersen and Diebold #281996#29, a simple nonlinear function g#28#01#29 which penalizes overprediction more than underprediction, namely g#28e i;t #29= 8 #3E #3E #3E #3C #3E #3E #3E : je i;t j if e i;t #15 0 e 2 i;t if e i;t #3C 0 #28recall that forecast errors are smaller than 1 in modulus in our case#29. 6 The results are arranged in Table7 where the sign of the test statistics signals a better forecast of the model in the column relative to GARCH if 6 Note that this function is di#0Berent from what was considered in the illustrative example of Diebold and Mariano, who use in their paper symmetric functions such as the absolute value and the square of forecast errors, therefore testing for the signi#0Ccance of the di#0Berence between Mean Absolute Errors and between Mean Squared Errors.... ..."

### Table 2: The Diebold-Mariano test statistic: whole sample

"... In PAGE 10: ...The results are shown in Table2 . As can be seen, we reject the hypothesis of equal expected squared error (i.... In PAGE 10: ... Therefore, these results reinforce our earlier conclusion from Table 1. [ Table2 here] As in Gardner and Perraudin (1993) and Henry and Weidmann (1995), we have checked the possible effects on the previous results following German reunification. To this end, we have divided the sample in two parts, the breaking point being 29 November 1990 [as in Henry and Weidmann (1995)].... ..."

### Table 4: The Diebold-Mariano test statistic: before and after German reunification

### Table 4: AMSPE of the Deterministic Predictor: Additional Elements with re- spect to Brown and Mariano (1989)

"... In PAGE 17: ... Table 2) and asymptotic mean square prediction error (cf. Table4 ) to zero. By the same token, the procedure suggests the adoption of a smaller a smaller estimated covariance matrix ~ will reduce the AMSPE through the quadratic form terms (cf.... ..."

### Table 5: AMSPE of the Monte Carlo Predictor: Additional Elements with respect to Brown and Mariano (1989)

### Table 5 Diebold-Mariano tests: model benchmark Gaussian GARCH - MS comparisons k=1 k=5 k=20

2005

### Table 1: Size and power of weighted modi ed Diebold-Mariano test statistics of equal forecast accuracy based on squared loss

"... In PAGE 9: ... 3.2 Monte Carlo results Table1 displays results from the rst experiment, with the rst panel referring to size and the second to sixth panels to power. Only results for = 0 are shown to save space.... In PAGE 17: ...acy based on a squared loss function. The DGP is the AR(1) process yt = yt 1 + quot;t. P one-step ahead forecasts are obtained from AR and TAR models, where the parameters in these models are estimated using a rolling scheme with a moving window of R observations. See Table1 for a description of the test statistics. The null is tested at a nominal 5% sig- ni cance level against the one-sided alternative that the AR model renders more accurate forecasts.... In PAGE 18: ...tc., indicate the average percentage of observations in the lower-upper regimes of the DGP. P one-step ahead forecasts are obtained from AR and TAR models, where the parameters in these models are estimated using a rolling scheme with a moving window of R observations. See Table1 for a description of the test statistics. The null is tested at a nominal 5% signi cance level against the one-sided alternative that the TAR model renders more accurate forecasts.... In PAGE 19: ... P one-step ahead forecasts are obtained from AR and TAR models, where the parameters in these models are estimated using a rolling scheme with a moving window of R observations. See Table1 for a description of the test statistics. The null is tested at a nominal 5% signi cance level against the one-sided alternative that the TAR model renders more accurate forecasts.... ..."

### Table 7: Number of series where model A (column) is better than model B (row) according to the modified Diebold-Mariano test at a 0.05 level (nRMSE test).

2001

"... In PAGE 14: ...05 level are shown in Tables 7 and 8 (See Appendix C for details). The results in Table7 concern to the nRMSE test and the ones in Table 8 concern to the MAE test.... ..."

Cited by 2

### Table 8: Number of series where model A (column) is better than model B (row) according to the modified Diebold-Mariano test at a 0.05 level (MAE test).

2001

"... In PAGE 14: ...05 level are shown in Tables 7 and 8 (See Appendix C for details). The results in Table 7 concern to the nRMSE test and the ones in Table8 concern to the MAE test.... ..."

Cited by 2