深層学習とウェーブレットを用いた多変量時系列予測器

抄録

The author integrated 1) multiple time-series analysis, 2) deep-learning, and 3) wavelet transform technics to forecast financial and economic time-series data. More specifically, the general purpose predictor was developed, which exploits large number of observable variables by summarizing them into latent factors through deep-learning. This can be regarded as a deep-learning version of Factor Augmented VAR model. As a preprocessing step, all observable variables are decomposed into cyclical components (waves) and a trend component by multiple resolution analysis based on "wavelet transforms". FAVAR model is fitted to the each decomposed series with extrapolating forecasts, and then integrated into the fitted values and forecasts of original series. The back-test for the period from Jan 2015 to Apr 2016 showed good performances of the 1-month- and 3-month-ahead predictions for TOPIX, USDJPY and other economic indicators, comparted with the simple VAR model.