Component based stock price predicton with STL decomposition: A meta-learning based ensemble model

The stock market is highly arduous to predict due to its nature. As a result of market noise, volatility, and external influences such as economic indicators, geopolitical events, and investor sentiment, stock price prediction is inherently challenging. It's observed in the existing literature that when a single univariate model is used to model stock prices, the models have difficulty fully detecting price movements. For this reason, we developed a daily forecasting framework applied to the BIST30 index, focusing on the THYAO stock. We decomposed the data into components using STL and predicted each component individually by selecting an appropriate modeling approach based on its characteristics. The trend component is modeled by Ridge Regression thanks to its capability of applying regularization to prevent overfitting, and the seasonal component is modeled using TCN. The residual component which captures irregular and volatile fluctuations, is modeled using an ensemble approach This component is handled through a sophisticated ensemble method that combines XGBoost, GARCH, and spike detection techniques as input to XGBoost meta model. Finally, we combined their predictions into a meta-learning model based on Lineer Regression to generate the final forecast. Empirical evaluations demonstrate significant improvements on stock price prediction accuracy, with evaluation metrics of MAE: 1.5574 and MAPE: 0.53%. The study found that ensemble meta-learning techniques and decomposed component based predictions significantly increased the accuracy of stock price predictions.