OPTIMIZATION OF LSTM NETWORKS FOR TIME SERIES FORECASTING

Abstract

Background. LSTM neural networks are a very promising means to develop time series analysis and forecasting. However, as well as neural networks for other fields and applications, LSTM networks have a lot of architecture versions, training parameters, and hyperparameters, whose inappropriate selection may lead to unacceptably poor performance (poor or badly unreliable forecasts). Thus, optimization of LSTM networks is still an open question. Objective. The goal is to ascertain whether the best forecasting accuracy is achieved at such a number of LSTM layer neurons, which can be determined by the time series lag. Methods. To achieve the said goal, a set of benchmark time series for testing the forecasting accuracy is presented. Then, a set-up of the computational study for various versions of the LSTM network is defined. Finally, the computational study results are clearly visualized and discussed. Results. Time series with a linear trend are forecasted worst, whereas defining the LSTM layer size by the lag in a time series does not help much. The best-forecasted are time series with only repeated random subsequences, or seasonality, or exponential rising. Compared to the single LSTM layer network, the forecasting accuracy is improved by 15 % to 19 % by applying the two LSTM layers network. Conclusions. The approximately best forecasting accuracy may be expectedly achieved by setting the number of LSTM layer neurons at the time series lag. However, the best forecasting accuracy cannot be guaranteed. LSTM networks for time series forecasting can be optimized by using only two LSTM layers whose size is set at the time series lag. Some discrepancy is still acceptable, though. The size of the second LSTM layer should not be less than the size of the first layer.

Keywords

• time series forecasting
• LSTM network
• LSTM layer size
• forecasting accuracy
• root-mean-square error
• maximum absolute error