The Research Department of Economics and Management of Tadbir Nikan Business, Marketing, and Finance Open 3092-6238 In Press 2027 08 01 Deep Learning versus Classical Machine Learning in Portfolio Optimization with Downside Risk: Evidence from the Abu Dhabi Securities Exchange (ADX) Deep Learning versus Classical Machine Learning in Portfolio Optimization with Downside Risk: Evidence from the Abu Dhabi Securities Exchange (ADX) 1 20 EN Journal Article 2026 05 05 <p>This study aimed to develop and evaluate a hybrid portfolio optimization framework integrating deep learning-based stock-return forecasting with Mean Semi-Absolute Deviation (MSAD) optimization to improve risk-adjusted asset allocation in the Abu Dhabi Securities Exchange. An applied quantitative design was employed using daily Open, High, Low, Close, and trading-volume data for 25 active ADX stocks. Three deep learning architectures—Deep Multilayer Perceptron (DMLP), one-dimensional Convolutional Neural Network (1D-CNN), and Long Short-Term Memory (LSTM)—were compared with Support Vector Regression (SVR). Predictors were constructed from rolling 20-day windows, and the dataset was partitioned into training, validation, and out-of-sample testing subsets. Forecasts were incorporated into an MSAD linear optimization framework, and stocks were ranked into five quintile portfolios. Predictive performance was assessed using MAE, MSE, and directional Hit Rates, while portfolio performance was evaluated through daily return and return-to-risk ratios. DMLP produced the lowest forecasting errors, with MAE=0.0163 and MSE=0.0013, and achieved the highest positive directional accuracy (HR+=0.5163). In portfolio optimization, DMLP+MSAD generated the strongest results for aggressive portfolios, producing daily returns of 0.1817 and 0.2164 and return-to-risk ratios of 0.49 and 0.78 for P1 and P3, respectively. SVR+MSAD showed greater stability in the defensive P2 portfolio, whereas CNN+MSAD demonstrated comparative resilience in negatively performing portfolios. Across the quintiles, MSAD-optimized AI strategies consistently exhibited superior risk-adjusted performance relative to equal-weighted allocation. Integrating predictive artificial intelligence models with downside-risk-based MSAD optimization improves portfolio construction under asymmetric and volatile market conditions. DMLP+MSAD is particularly effective for high-return strategies, whereas SVR and CNN may offer advantages under more defensive or adverse market regimes, indicating that optimal model selection should reflect investors’ risk-return preferences.</p> https://bmfopen.com/index.php/bmfopen/article/download/662/479