Title: Forecasting during the recent financial crisis: Automatic versus adaptive exponential smoothing methods
Authors: Mohaimen Mansur - The University of Manchester (United Kingdom) [presenting]
Abstract: Central banks today need to produce reliable forecasts of a large volume of macroeconomic and financial time series in order to make well-informed forward-looking policy decisions. Accurate real time forecasting of these series, however, proved to be particularly challenging following the recent global financial crisis which induced moderate to large structural breaks to many of these series. We advocate use of two types of exponentially smoothing methods which are known to be robust to different types of structural changes. An automatic method chooses the best forecasting model based on in-sample performance and the adaptive method selects the down-weighting rate of past data based via cross-validation. A systematic forecasting exercise using the FED-MD dataset, an up-to-date data consisting of various macroeconomic and financial time series of the US demonstrates that forecasts of advised exponential methods are comparable to well performing no-change forecasts during the pre-crisis stable period, but the exponential methods outperform the benchmark for many of the series during the crisis and following recovery periods. We further investigate properties of those series for which the gains from exponential methods are particularly large.