This paper proposes the BLAME algorithm, an analytical Bayesian extension to the renowned EM algorithm. A simulation study is conducted to understand the strengths and weaknesses of the proposed modification and how these may affect forecasting performance. The simulation study finds that the classical EM method is powerful for latent coefficient estimation, whereas the BLAME method excels in observed coefficient estimation. When considering one-monthahead US macroeconomic forecasts, the EM algorithm performs very well, being comparable to VARIMA and VEC models, but does not beat a random walk. The Bayesian algorithm performs worse than expected, producing respectable forecasts only for inflation, the output gap, and cyclical unemployment. Overall, this research remains a strong advocate for the theoretical soundness of the applied shrinkage techniques exemplified by the general BLAME algorithm.

Lange, R.
hdl.handle.net/2105/50409
Econometrie
Erasmus School of Economics

Rallis, T. (2019, September 24). Bayesian Estimation of a Gaussian Macrofinance State Space Model. Econometrie. Retrieved from http://hdl.handle.net/2105/50409