Improving forecasting performance by window and model averaging

We present results from an extensive study on the benefits of rolling window and model averaging. Building on the recent work on rolling window averaging by Pesaran et al (2010, 2009) and on exchange rate forecasting by Molodtsova and Papell (2009), we explore whether rolling window averaging can be considered beneficial on a priori grounds, that is whether researchers can use it to improve forecasting performance and to avoid ‘window mining’ in short horizons. In addition, we investigate whether rolling window averaging can improve the performance of model averaging, especially when ‘simpler’ models are being used. Our results provide strong support for rolling window averaging, outperforming the best window forecasts more than 50% of the time across all rolling windows considered – with the outperformance being statistically significant. Furthermore, rolling window averaging smoothes out the forecast path and improves robustness of the forecasting model, thus minimizing the pitfalls associated with potential structural breaks. An illustrative simulation supports the proposed improvement with the double averaging approach. Afterwards the technique is applied in three datasets: exchange rates for 12 OECD countries, US inflation rate and US output growth rate. For exchange rates, we use the dataset of Molodtsova and Papell (2009) and replicate their analysis by considering rolling window and model averaging. The results reveal rolling window averaging can further improve the performance of the models and, in addition, when combined with model averaging brings forth the forecasting ability of ‘simpler’ economic models of exchange rates. With respect to US inflation and output growth forecasting, we again find that rolling window averaging outperforms the best individual window forecasts by more than 50% of the time, with significant differences from the benchmarks, and helps in model averaging by bringing forth the predictive power of economic variables in parsimonious models.