GA-LSTM: Performance Optimization of LSTM driven Time Series Forecasting

These past few years, technology has simplified the process of gathering and arranging time series data. This paves the way for tremendous opportunities to gain helpful insights by analysing these data. Historically, statistical models have been used for time series analysis. These models work well for linear or univariate data but struggle to accurately capture complex nonlinear trends or when unpredictable external factors impact the data (such as stock or trading prices). Long Short-Term Memory (LSTM) has emerged as one of the most popular options for analysing time series data efficiently. However, there are still a few challenges associated with it. The main reason for this study is to discuss these shortcomings and propose an intelligent system to deal with these flaws. One of the challenges is the choice of hyperparameters, which are handled successfully using the Genetic Algorithm (GA) to optimise the LSTM method’s hyperparameters and its variants. In this case, the GA solves the combinatorial optimization problem of finding the optimal hyper-parameters for the LSTM model and selecting the appropriate features from the data set. The proposed model has experimented on the National Stock Exchange-Fifty (NIFTY-50) dataset and found that the accuracy of all LSTM variants was improved. On average, the GA improved the performance of LSTM models by 60.11%. The Bidirectional LSTM model performs best with a root mean square error of an average of 66.39% and mean absolute error of an average of 49.43% after optimization, followed by the classic LSTM and the stacked LSTM models. The combination of LSTM with GA holds promise for enhancing the predictive power of time series analysis in various fields.