The emergence and underdevelopment of machine learning (ML) methods to effectively process large-scale finance data and make accurate predictions encourage their application to traditional Portfolio Optimization (PO) models. This paper formulates a two-stage hybrid model for portfolio optimization. The first stage involves (i) Fuzzy Support Vector Machines (Fuzzy SVM) shortlisting of assets, which is an effective approach to deal with uncertainty and nonlinearity in finance data. Stage (ii) optimizes the chosen assets under Second-Order Stochastic Dominance (SSD) constraints in the second stage to address the pitfalls of the classical Markowitz model for maximizing investor satisfaction and utility. The new framework is tested on eight international equity markets, such as Bovespa 90 (Brazil), DAX 40 (Germany), Dow Jones Industrial Average (U.S.A.), EURO 50 (Europe), FTSE 100 (U.K.), Hang Seng (Hong Kong), Nifty 100 (India), and Nikkei 400 (Japan), and compared with models having no asset shortlisting and using traditional SVM-based selection. Empirical results indicate that Fuzzy SVM–SSD performs better consistently compared to the benchmarks on various measures of performance, such as average returns, Value at Risk (VaR), Conditional Value at Risk (CVaR), and minimum return. Also, better reward-to-risk performance is witnessed by virtue of higher STARR, Sharpe, and Sortino ratios. On the whole, the findings emphasize the significance of smart asset shortlisting and show the usefulness of combining fuzzy learning processes with risk-sensitive portfolio optimization for better financial decision-making.