A Time Series Forecasting Model Based on Machine Learning

Time series forecasting plays a critical role in domains such as finance, environmental science, and health management. Traditional methods like ARIMA and exponential smoothing often fall short in addressing the complexities of modern datasets, especially in modeling non-linear relationships and intricate dependencies. This study introduces the CLT model, an innovative ensemble approach that combines CatBoost, LightGBM, and Transformers to enhance forecasting accuracy and reliability. Employing a dataset of Bitcoin prices from 2014 to 2024, this research involves rigorous data preprocessing, sophisticated feature engineering, and systematic model selection. The CLT model significantly outperforms traditional and single-model approaches by effectively capturing complex dependencies and demonstrating remarkable stability across various validation folds. The results demonstrate that the CLT model outperforms traditional and single-model approaches by effectively capturing complex dependencies and showing remarkable stability across various validation folds. In terms of quantitative metrics, our model achieves an average Root Mean Square Error (RMSE) of 434.54 and an average Mean Absolute Error (MAE) of 374.9, underscoring its superior predictive capabilities. This research underscores the potential of integrating multiple advanced machine learning models to significantly improve the performance of time series forecasting, offering a scalable and efficient solution for high-stakes domains.