A Study of LSTM Optimisation for Forecasting Volatile Time Series Data

Forecasting time series data is an important subject in economics and business where the Autoregressive Integrated Moving Average (ARIMA) has been extensively used despite its weaknesses, from requiring a minimum number of data points to the assumption of linearity of data. With recent advancement, the Long Short-Term Memory (LSTM) shows potential to address such weaknesses. This research is aimed to identify the effect of hyperparameters such as the number of epochs, units, and optimiser on accuracy of prediction on volatile time series data. Performance metric used is the model accuracy measured with RMSE. This research concluded that the accuracy of LSTM can be improved by increasing the number of epochs, where the increase of the number of epochs from 25 to 125 can increase the prediction accuracy by 64%, and 77% when increased from 25 to 300 epochs. SGDM as the optimiser or solver was able to produce the most accurate result (RMSE at 0.51) compared to Adam and RMSProp (RMSE at 0.89 and 0.78, respectively). Meanwhile the number of units showed inconclusive impact to the prediction accuracy.