Real-time forecasting of time series in financial markets using sequentially trained dual-LSTMs

Abstract

Financial markets are highly complex and volatile; thus, accurate forecasting of such markets is vital to make early alerts about crashes and subsequent recoveries. People have been using learning tools from diverse fields such as financial mathematics and machine learning to make trustworthy forecasting on such markets. However, the accuracy of such techniques had not been adequate until artificial neural network frameworks such as long short-term memory (LSTM) were utilized. Moreover, making accurate real-time forecasting, also known as nowcasting, of financial time series is highly subjective to the LSTM’s architecture in use and the procedure of training it. Herein, we forecast financial markets in real-time by training a dual version of LSTM which forecasts only one time step at each iteration so that the forecast for this iteration will be in the input for the next iteration. Semi-convergence is a prominent issue in a recurrent LSTM setup as the error could propagate through iterations; however, the duality of this LSTM aids in dwindling this issue. Especially, we employ one LSTM to find the best number of epochs associated with the least loss and train the second LSTM only through that many epochs to make forecasting. We treat the current forecast as a part of the training set for the next forecast and train the same LSTM. While classic ways of training cause more error when the forecast is made further away through the test period, our approach offers superior accuracy as the training increases when it proceeds through the testing period. The forecasting accuracy of our approach is validated using three time series from each of the three diverse financial markets: stock, cryptocurrency, and commodity. The results are compared with those of a single LSTM, an extended Kalman filter, and an autoregressive integrated moving average model.