Forecast Combination for Discrete Choice Models: Predicting FOMC Monetary Policy Decisions
Social Science Research Network
This paper suggests combining multiple forecasts from discrete choice models for the FOMC decisions on the target rate instead of relying on one specific model. The main advantage of combining forecast is to improve forecast accuracy, as first shown by Bates and Granger . There are several other advantages to forecast combination; for example it allows to deal with a larger number of variables than a univariate model or it may be more robust to potential breaks in trends and intercepts.
... mmermann  provides a thorough overview of the sizeable forecast combination literature. First, since there is limited support for combining discrete outcomes, this paper considers combining probability forecasts estimated from the probability mass function associated with the one-period-ahead value of random variables. Furthermore, it is shown that probability forecasts from combining discrete choice models instead of combining probability forecasts from discrete choice models, do not yield a discrete choice model with tractable properties. Combination of probability forecasts has already been used in the context of aggregating probability distributions of expert opinions, as discussed in Genest and Zidek  and Clemen and Winkler . On the methodological side, Kamstra and Kennedy  and Clements and Harvey  among others, investigate different techniques for combining probability forecasts as well as their properties. Second, the paper proposes to use of log and quadratic scoring rules introduced by Brier  and Good  , both to evaluate forecasting accuracy of models and to construct adaptive weights for combination. The paper shows that forecast combination achieves greater accuracy in terms of scoring rules. Although probability scoring rules have been applied in the economics and finance literature, for example by Diebold and Rudebusch  , Ghysels  and Anderson and Vahid  , it has not been employed as a way to combine probability forecasts. Third, the proposed methodology is implemented empirically in forecasting the Fed decisions to change the target rate. The empirical exercise illustrates well that combining probability forecasts improve out-of-sample forecast performance especially when they are combined using weights based on either log or quadratic scoring rules. These results utilise the data and models used by Hu and Phillips [2004a] as a benchmark. The effect of sampling variation around probability forecasts are also assessed through simulations, which follows to some extend the work by McCabe et al.  . The paper is organised as follows. Section 2 sets out the discrete choice model. Forecast combination methodology is discussed in Section 3. Baseline results and robustness checks are provided in Section 4. Section 5 discusses the assessment of the effect of sampling variation associated with probability forecasts. Concluding remarks are presented in Section 6.