Financial Time Series Prediction via Neural Ordinary Differential Equations Approach

This paper considers the prediction problem of financial time series, namely, the prediction of exchange rate for four currencies with the Chinese Yuan (CNY). A novel approach is developed for this challenging topic by neural ordinary differential equations (NODEs). The comparing simulations with long short-term memory (LSTM) and gated recurrent unit (GRU) show that NODEs have better precision and accuracy in exchange rate prediction. Especially, for the drastic change case, the NODEs approach gives more flexible technique.