PhD student at WISE.

There is much evidence of structural changes in macroeconomic and financial data such that forecast accuracy is often sensitive to the selection of estimation windows. The local adaptive method (LAM) selects local homogeneous intervals through a backward testing procedure on a set of nested intervals, which is proved to be excellent in forecasting in many cases due to parameter stability. However, the LAM may not efficiently utilize pre-break information when a recently detected break is small. We design a dynamic adaptive method (DAM) with cross-validation on multiple sets of nested intervals. The DAM ensures the possibility of selecting longer intervals under small breaks or smooth changes to increase information efficiency, thus to improve the balance between bias and variance. Our simulation study shows that the DAM outperforms alternative methods, including the LAM, the rolling window selection method of Inoue et al. (2017), the pooled forecast of Pesaran and Timmermann (2007) and the ex-post best rolling window. A forecast application on the Chinese CPI inflation not only demonstrates its superiority with respect to the alternative methods, but also beats a well-known survey forecast.