Transfer Learning for Macroeconomic Forecasting

In this paper we present a novel approach to macroeconomic forecasting based on transfer learning using normalizing flows conditioned on LSTM-based encoder-decoder as building blocks. The approach consists of two steps: (1) pretraining and (2) fine-tuning. At the pre-training step, we train a model based on macroeconomic data of many different countries. The obtained pre-trained model can capture hidden patterns in temporal changes of macroeconomic indicators. The pre-trained model is then fine-tuned on macroeconomic data of the target country. In the approach, LSTM-based encoder-decoder aims at learning vector representations of the input data. The obtained representations are then transformed by using conditional normalizing flows so that the distribution of the data encoded in the representations is transformed into a more complex distribution. We evaluate the proposed approach on seventeen macroeconomic variables of a public dataset. The experimental results show that transfer learning using normalizing flows conditioned on LSTM-based encoder-decoder as building blocks significantly improves the performance of macroeconomic forecasting with one-step ahead forecasts. To the best of our knowledge, this is the first time neural transfer learning has been successfully applied to forecast many macroeconomic variables simultaneously.