Dynamic model averaging in large model spaces using dynamic Occam׳s window

Abstract

Bayesian model averaging has become a widely used approach to accounting for uncertainty about the structural form of the model generating the data. When data arrive sequentially and the generating model can change over time, Dynamic Model Averaging (DMA) extends model averaging to deal with this situation. Often in macroeconomics, however, many candidate explanatory variables are available and the number of possible models becomes too large for DMA to be applied in its original form. We propose a new method for this situation which allows us to perform DMA without considering the whole model space, but using a subset of models and dynamically optimizing the choice of models at each point in time. This yields a dynamic form of Occam’s window. We evaluate the method in the context of the problem of nowcasting GDP in the Euro area. We find that its forecasting performance compares well with that of other methods.

Dynamic Model Averaging for Practitioners in Economics and Finance: The eDMA Package

“…However, handling larger number of combinations can quickly become very cumbersome and impose technical limits on the software at hand, especially with regards to memory consumption, see for example, Koop and Korobilis (2012). In order to deal with this issue, Onorante and Raftery (2016) suggest a strategy that considers not the whole model space, but rather a subset of models and dynamically optimizes the choice of models at each point in time. However, Onorante and Raftery (2016) have to assume that models do not change too fast over time, which is not an ideal assumption when dealing with financial and in some cases monthly economic data.…”

“…In order to deal with this issue, Onorante and Raftery (2016) suggest a strategy that considers not the whole model space, but rather a subset of models and dynamically optimizes the choice of models at each point in time. However, Onorante and Raftery (2016) have to assume that models do not change too fast over time, which is not an ideal assumption when dealing with financial and in some cases monthly economic data. Furthermore, it is not clear to us how one can incorporate the modifications suggested in Dangl and Halling (2012) within the framework of Onorante and Raftery (2016).…”

Modeling association in microbial communities with clique loglinear models

“…We can think of πpMq as a posterior distribution over M obtained by assuming a flat prior over M. Thus, the πpMq's can be considered to be the Bayes model weights, and, in the sequal, these weights will be used to perform model averaging using Occam's window. This methodology originates in [36], and has been developed in numerous other contexts, e.g., dynamic linear models [44] and graphical models [35]. For a more in depth discussion, see Supplementary Materials, Section B.2.…”

Forecasting Inflation in Russia Using Dynamic Model Averaging

“…Both papers forecast inflation in the US. Onorante and Raftery (2016) nowcast GDP for the Euro area, employing 30 predictors. The curse of model space dimensionality is circumvented by introducing stochastic search over the space of the models.…”