<![CDATA[Background: Although researchers have increasingly explored the combination of machine learning based return forecasts with traditional portfolio construction, the discussion about how these predictive models reshape established methods is still developing. One prominent direction involves extending the classical mean variance approach so that it incorporates forward looking estimates, which is often referred to as the Mean Variance with Forecasting (MVF) framework. In parallel, approaches such as Risk Parity Portfolios (RPP) and Maximum Drawdown Portfolios (MDP) continue to gain attention because they represent different perspectives on risk management. However, despite this growing interest, there is still limited empirical evidence on how Support Vector Regression (SVR) and Random Forest (RF) forecasts affect performance within these three frameworks, and this gap is particularly evident in emerging markets. Objective: This study examines how SVR and RF one day ahead return forecasts influence the risk adjusted performance, drawdown control, and diversification outcomes of the MVF, RPP, and MDP frameworks when applied to stocks in the VN-100 index between 2017 and 2024. The choice of these frameworks is intentional, as each reflects a different level of investor tolerance for risk. MVF tends to appeal to investors who place greater weight on potential returns, RPP seeks a more even distribution of risk which suits investors with a moderate stance, and MDP focuses on limiting losses, making it more suitable for investors who are highly cautious about downside risk. Methods: Daily returns of VN 100 stocks were standardized and then used as inputs for the SVR and RF models. The models were tuned through a grid search on data from 2017 to 2021 and evaluated on the remaining period up to 2024. After generating the return forecasts, portfolios were constructed under the MVF, RPP, and MDP frameworks, and their performance was assessed using monthly excess returns, the information ratio, and total returns in comparison with the VN-100 index. Results: The forecasts generated by SVR showed greater reliability than those obtained from the RF model, and this contributed to stronger risk adjusted performance when applied within the MVF framework. The MDP strategy, which places emphasis on limiting drawdowns, delivered solid protection against large losses, whereas the RPP approach produced more moderate returns along with improved consistency. Conclusion: In the end, matching forecasting techniques and portfolio construction methods with an investor’s risk preferences and view of the market is crucial, since overall performance is shaped by the interaction between predictive inputs and allocation rules. Looking ahead, future studies could investigate a wider range of forecasting models, incorporate transaction costs more explicitly, and explore adaptive forms of optimization that are capable of responding to changing market conditions. Keywords: Machine Learning, Maximum Drawdown, Mean-Variance, Portfolio Optimization, Random Forest, Risk Parity, Support Vector Regression, VN-100 Index]]>
