Unveiling Complex Seasonality in Stock Price Forecasting Using a Seasonal‐Adjusted Hybrid Machine Learning Approach

In emerging financial markets, stock price forecasting is challenged by nonstationarity, irregular trading calendars, and evolving structural dynamics that limit the effectiveness of conventional linear models. This study develops and evaluates a seasonal‐adjusted hybrid machine learning framework to forecast the daily closing stock prices of Square Pharmaceuticals PLC., one of the most actively traded firms on the Dhaka Stock Exchange. Using daily data from January 2000 to June 2025, STL decomposition is applied to separate trend and seasonal components prior to predictive modeling. A two‐stage hybrid architecture is implemented in which statistical and machine learning models are combined sequentially to capture structured temporal patterns and nonlinear residual behavior. Forecast performance is evaluated across a wide range of statistical, machine learning, and deep learning models using multiple accuracy metrics. Results demonstrate that hybrid configurations where statistical models are applied prior to machine learning consistently outperform single‐model and reverse‐sequencing approaches. In particular, the STL–TBATS–random forest regression model achieves the highest predictive accuracy (MAE = 4.33, RMSE = 5.90, MAPE = 2.04%, and MASE = 0.53), highlighting the importance of model sequencing in hybrid forecasting frameworks. Long‐horizon scenario projections suggest relatively stable conditional price trajectories under the assumption that recent historical patterns persist. The findings provide empirical evidence that seasonal‐adjusted hybrid architectures can enhance forecasting robustness in data‐constrained emerging markets. While the analysis focuses on a single benchmark stock, the proposed framework is methodologically generalizable and can be applied to other equities, sectoral indices, or broader financial markets characterized by complex temporal structures.