توظيف التعلم الآلي في تعزيز دقَة التنبؤ بعوائد المحفظة الاستثمارية مقارنة بنماذج تقليدية (دراسة تحليلية مقارنة لعَينة من الشركات المدرجة في سوق العراق للأوراق الماليًة للمُدة 2016-2024)

Rp-Employing Machine Learning to Enhance the Accuracy of Portfolio Return Predictions Compared to Traditional Model .pdf

This study aims to evaluate and analyze the explanatory and predictive power of individual stock returns within an investment portfolio context in the Iraq Stock Exchange (ISX) for the period (2016–2024). This is conducted through an empirical comparison between the Fama-French Three-Factor Model and the Long Short-Term Memory (LSTM) model. The originality of this research stems from being a pioneering local study addressing the problem of prediction accuracy as a fundamental cornerstone for constructing efficient portfolios in the future. The methodological framework was adapted by adopting the “Horizontal Average of Averages” approach for the Book-to-Market (B/M) ratio in firm classification—an innovative procedure designed to neutralize the impact of outliers and ensure the stability of input data amidst market volatility. The study sample comprised (22) joint-stock companies distributed across (6) active economic sectors, with a total of (2,354) monthly observations spanning (107) months. The Coefficient of Determination (R2) was employed as an indicator of explanatory power; the traditional model succeeded in capturing (67.9%) of return fluctuations, recording a significant negative effect for the size factor (SMB = -0.019), reflecting the dominance of large-cap companies. Conversely, the value factor (HML) yielded a positive result of (0.004) but lacked statistical significance (p = 0.056), revealing an explanatory gap and a non-linear pattern in return movements. To measure predictive efficiency, the traditional model recorded error metrics of (0.00056) for MSE, (0.0238) for RMSE, and (0.0178) for MAE, reaching a Mean Absolute Percentage Error (MAPE) of (2.49%).In contrast, the LSTM model, designed with a deep architecture using the R language, demonstrated superiority in enhancing excess return prediction by capturing the latent temporal patterns within the portion unexplained by linear models. The model successfully reduced all error metrics to their lowest levels, with MSE decreasing to (0.00035), RMSE to (0.0189), and MAE to (0.0092), while MAPE reached approximately (1.08%). The study concludes that the prediction accuracy achieved at the individual asset level serves as the cornerstone for enhancing investment portfolio performance. It recommends adopting deep learning techniques as a benchmark tool to transcend the limitations of classical analysis and support investment decisions within the Iraqi market environment.