To select a forecast model among competing models, researchers often use ex-ante prediction experiments over training samples. Following Diebold and Mariano (1995), forecasters routinely evaluate the relative performance of competing models with accuracy tests and may base their selection on test significance on top of comparing forecast errors. With extensive Monte Carlo analysis, we investigated whether this practice favors simpler models over more complex ones, without gains in forecast accuracy. We simulated the autoregressive moving-average model, the self-exciting threshold autoregressive model, and vector autoregression. We considered two variants of the Diebold–Mariano test, the test by Giacomini and White (2006), the F -test by Clark and McCracken (2001), the Akaike information criterion, and a pure training-sample evaluation. The findings showed some accuracy gains for small samples when applying accuracy tests, particularly for the Clark–McCracken and bootstrapped Diebold–Mariano tests. Evidence against this testing procedure dominated, however, and training-sample evaluations without accuracy tests performed best in many cases.

On using predictive-ability tests in the selection of time-series prediction models: A Monte Carlo evaluation

Mauro Costantini;
2021-01-01

Abstract

To select a forecast model among competing models, researchers often use ex-ante prediction experiments over training samples. Following Diebold and Mariano (1995), forecasters routinely evaluate the relative performance of competing models with accuracy tests and may base their selection on test significance on top of comparing forecast errors. With extensive Monte Carlo analysis, we investigated whether this practice favors simpler models over more complex ones, without gains in forecast accuracy. We simulated the autoregressive moving-average model, the self-exciting threshold autoregressive model, and vector autoregression. We considered two variants of the Diebold–Mariano test, the test by Giacomini and White (2006), the F -test by Clark and McCracken (2001), the Akaike information criterion, and a pure training-sample evaluation. The findings showed some accuracy gains for small samples when applying accuracy tests, particularly for the Clark–McCracken and bootstrapped Diebold–Mariano tests. Evidence against this testing procedure dominated, however, and training-sample evaluations without accuracy tests performed best in many cases.
File in questo prodotto:
File Dimensione Formato  
Costantini_Kunst_IJF_2021.pdf

non disponibili

Tipologia: Documento in Versione Editoriale
Licenza: Creative commons
Dimensione 556.82 kB
Formato Adobe PDF
556.82 kB Adobe PDF   Visualizza/Apri   Richiedi una copia
Pubblicazioni consigliate

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11697/162310
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus 9
  • ???jsp.display-item.citation.isi??? 6
social impact