Analyzing Effects of Time Series Data Characteristics on LSTM Performance

The use of deep learning techniques for time series forecasting has increased in recent years. Long Short-Term Memory (LSTM) is a deep learning technique that has been shown to be effective for time series forecasting. This research proposes a methodology to study the relationship between LSTM performance and time series data characteristics, including trend, seasonality, skewness, kurtosis, chaotic behavior, and coefficient of variation. The methodology was applied to 48 stock price datasets. Three techniques, namely multiple linear regression, feature importance, and SHAP value, were used to compare LSTM performance across different data characteristics. The results show that coefficient of variation and chaotic behavior are the two most crucial characteristics affecting LSTM performance.