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

Abstract

Raftery, Kárnỳ, and Ettler (2010) introduce an estimation technique, which they refer to as Dynamic Model Averaging (DMA). In their application, DMA is used to predict the output strip thickness for a cold rolling mill, where the output is measured with a time delay. Recently, DMA has also shown to be useful in macroeconomic and financial applications. In this paper, we present the eDMA package for DMA estimation implemented in R. The eDMA package is especially suited for practitioners in economics and finance, where typically a large number of predictors are available. Our implementation is up to 133 times faster then a standard implementation using a single-core CPU. Thus, with the help of this package, practitioners are able to perform DMA on a standard PC without resorting to large clusters, which are not easily available to all researchers. We demonstrate the usefulness of this package through simulation experiments and an empirical application using quarterly U.S. inflation data.

Forecasting Indonesian Inflation Within an Inflation-Targeting Framework: Do Large-Scale Models Pay Off?

“…Suppose that indicates the model that is used at each time period, and . Then, the DMA approach entails computing and averaging forecasts across models using these probabilities to forecast inflation at time t using inflation predictors through time t-1 (for details, see Koop and Korobilis, 2012;Catania and Nonejad, 2018).…”

An Overview of Dynamic Model Averaging Techniques in Time‐series Econometrics

“…However, the package does not allow for the extensions mentioned in Dangl and Halling (2012), see Section 4.1. Likewise, as shown in Catania and Nonejad (2018), this package requires a very large amount of RAM even for moderately large applications and does not allow for parallel computing. The eDMA package in R (Catania and Nonejad, 2017) suggested in Catania and Nonejad (2018) implements DMA and DMS following Koop and Korobilis (2012).…”

“…Evidently, not everyone has access to external serves with fast computing power. Parallelization can be considered as a solution, see Catania and Nonejad (2018) for an illustration. However, this requires a high degree of technical knowledge from the practitioner's part.…”

Forecasting Inflation in Russia Using Dynamic Model Averaging

“…Handling such a large number of models requires considerable computational power and a prohibitive amount of time for computation on a desktop intended for research or a laptop computer with a typical technical specification. For this project, I have employed the R package called eDMA (Catania and Nonejad, 2018), which employs efficient algorithms and parallel computation. If M = 20 and the number of time observations is 189, my ASUS laptop with Intel Core i9 12-core processor and 64 GB of RAM completes all computations for the main exercise within about ten minutes.…”

“…The curse of model space dimensionality is circumvented by introducing stochastic search over the space of the models. An important assumption needed for the approximation to be well-grounded is that the DGP smoothly transitions from one model to another (Catania and Nonejad, 2018). Inspection of the posterior inclusion probabilities in Onorante and Raftery (2016) and other papers, which feature occasional swings, suggests that this is unlikely to be the case.…”