Exploring the Characteristics of Time Series Data Affecting Forecasting Errors Using LSTM
Lugnapis Kaewwhangsakul, Nantachai Kantanantha · 2024
This paper investigates the predictive capabilities of Long Short-Term Memory (LSTM) models using diverse time series data relevant to lifestyle decisions and business planning, such as stock prices, exchange rates, gold prices, and weather conditions. The main contributions include analyzing 42 datasets from Kaggle.com to identify influential features impacting LSTM prediction accuracy. Findings reveal varying prediction accuracy across datasets due to different data characteristics, with some datasets exhibiting high Mean Absolute Percentage Error (MAPE) when they process Coefficient of Variation (CV) exceeding 50%. Additionally, datasets with non-normal distributions and high kurtosis show greater prediction errors. These findings underscore the significance of considering dataset features for accurate LSTM predictions in real-world applications.