Bayesian Compressed Vector Autoregression for Financial Time-Series Analysis and Forecasting

Abstract

Advanced time series models have been intensively developed and used to predict in financial data such as foreign exchange data (forex). In this paper, we implement the random compression method to reduce a large dimensional forex data into much smaller matrix form. Then, Bayesian inferences on vector autoregression are used to obtain all interesting parameters. Subsequently, the models are able to perform out-of-sample prediction up to 14 days ahead of forecast. For empirical works, 30 forex pairs are used in this paper. The results show that Bayesian compressed vector autoregression (BCVAR) and time-varying BCVAR (TVP-BCVAR) deliver excellent forecasting on AUD-JPY, CAD-CHF, CAD-JPY, EUR-DKK, EUR-MXN, and EUR-TRY forex datasets according to mean square forecasting error, outperforming the traditional benchmark Bayesian Autoregression.