Supervised learning models, statistical models or hybrid models? A prediction of clean energy stock based on fear and fundamental factors

This paper explores several time series models for predicting the S&P Clean Energy Index. We begin by identifying factors previously found to influence the clean energy market and use eXtreme Gradient Boosting (XGBoost) to rank and filter feature importance of variables, followed by further validation and variable selection using SHAP values. Next, we simulate future feature data using methods like Random Forest and Long Short-Term Memory (LSTM). For the LSTM-based simulations, the data is generated through a classification-then-prediction approach using the K-Nearest Neighbors (KNN) algorithm. To predict the index’s volatility, we employ statistical models such as AutoRegressive Integrated Moving Average (ARIMA) and Vector Autoregression (VAR). Additionally, we use advanced methods like LSTM, Supervised Autoencoder (SAE), and hybrid models such as Prophet Features and LSTM_Autoregressive(LSTM_AR). Each model’s parameter-tuning process will be explained in detail. Finally, we compared the models’ performance and prediction results, discussing their strengths and suitability for different scenarios. We confirmed that the Prophet-based models performed well on Random Forest simulated data when predicting both the trend and actual values of the S&P clean energy index.