Macroeconomic Forecasting Based on LSTM-Conditioned Normalizing Flows

Macroeconomic forecasting is a key task of developing the outlook for economy of a country and supporting decision making. In this paper we present a novel approach to macroeconomic forecasting based on LSTM-based encoder-decoder and conditional normalizing flows. First, we employ LSTM-based encoder-decoder to learn vector representations of the input data. The obtained representations are then transformed by using conditional normalizing flows. In such a way, the distribution of the data encoded in the representations is transformed into a more complex distribution. We evaluate the proposed approach on 215 macroeconomic variables of FRED-QD dataset. In experiments, we use two kinds of normalizing flows: Real-valued Non-Volume Preserving and Masked Autoregressive Flow. Experimental results show that proposed approach outperforms VAR in several orders of magnitude and also outperforms DeepVAR on the benchmark dataset. To the best of our knowledge, this is the first time LSTM-based encoder-decoder in combination with conditional normalizing flows has been successfully adopted to forecast hundreds of macroeconomic variables simultaneously.