<p>Since Markowitz initially came up with the theory of portfolio selection way back in 1952, risk-return optimization has been the core subject of finance studies. Traditional risk measures, such as variance, while popular, may, at times, fall short when trying to mitigate downside risk. New approaches, in a bid to counteract these limitations, more and more utilize advanced risk measures such as Value-at-Risk (VaR) and Conditional Value-at-Risk (CVaR). This research introduces a hybrid portfolio optimization model that combines Conditional Value-at-Risk with Fuzzy Least Squares Support Vector Machines (FLSSVM), an enhanced version of the conventional Least Squares Support Vector Machines (LSSVM) model. The fuzzy logic feature provides better data uncertainty and market noise management, while CVaR supports good representation of tail risk in portfolio returns. In addition, a key to effective portfolio building is asset selection. The nature and behavior of individual assets largely determine the performance of the portfolio. Proper asset categorization and choosing, by machine learning algorithms like FLSSVM, allow for better-informed and data-based decisions that optimize return potential and minimize risks. Machine learning is superior at detecting hidden patterns, classifying assets, and learning from non-linearities, and hence is perfectly capable of identifying the right assets for a portfolio by predicting their behavior under different market conditions. After intelligent classification has identified high-quality assets, portfolio optimization can be conducted more efficiently. The chosen assets are more suited to the investor’s goals and limitations, leading to more effective and reliable portfolios. With the incorporation of machine learning-powered asset selection with risk-sensitive optimization methods like CVaR, the overall model has better performance. An empirical experiment is performed using equity data of the two biggest financial indices–Nifty 50 (India) and Dow Jones 30 (USA). Outcomes invariably show that the FLSSVM-CVaR model significantly performs better than the traditional LSSVM-CVaR model on the CVaR efficient frontier, also on different performance measures, such as returns, risk, risk-adjusted returns, distributional characteristics, statistical validation, tail-risk estimates, and across different market regimes, including bull, bear, and volatile markets. By integrating machine learning with risk management methods, this study illustrates an even more active and reliable option for portfolio optimization in the context of actual-world financial risks.</p>