Adaptive Hyperparameter Optimization for Financial Time Series Forecasting: A Chameleon Swarm-driven XGBoost Approach

Abstract

Time series forecasting is a central theme in the financial market, and the ability to estimate stock prices and trends accurately has a direct impact on investment strategies and risk management decisions. Statistical methods and neural network-based models often struggle to cope with the nonlinear and erratic nature of financial data. This work is aware of these shortcomings and proposes a new model, Weighted Chameleon Swarm-driven eXtreme Gradient Boosting (WCS-XGBoost), to improve prediction performance in challenging time series cases. Historical stock price data from credible public sources is collected, emphasizing daily closing prices and corresponding technical indicators. The data is normalized, then undergoes feature extraction via Principal Component Analysis (PCA) to reduce dimensionality while maintaining signal integrity. The predictive engine's central component, WCS-XGBoost, utilizes Chameleon Swarm Optimization to fine-tune XGBoost hyperparameters adaptively, maximizing accuracy and generalization. The model was implemented in python that exhibited superior performance with a Root Mean Square Error (RMSE) of 0.2312, accuracy of 98.69%, and Mean Absolute Percentage Error (MAPE) of 0.321, all of which were significantly better than baseline models. This framework highlights the potential of hybrid evolutionary learning in the advancement of stock market forecasting methodologies.