EvoPort: An Evolutionary Framework for Portfolio Optimization via Randomized Alpha Discovery and Ensemble-Based Allocation

In this paper, we introduce EvoPort, a novel evolutionary portfolio optimization method that leverages stochastic exploration over a spectrum of investment pipeline depths. From raw equity data, we employ a randomized feature generation framework that hierarchically produces mathematical, logical, time-series, and cross-sectional operators for uncovering latent trading signals. Candidate alphas are then evaluated through a randomized hill-climbing optimization procedure, taking as guidance performance measures such as mean squared error (MSE) or Sharpe ratio. In order to increase robustness and generalizability further, we use a random ensemble model selection process whereby a heterogeneous set of machine learning models (e.g., linear regression, logistic regression, XG-Boost) are randomly drawn and combined to backtest the generated alphas. Finally, we use randomized portfolio weighting schemes based on the Markowitz modern portfolio theory with stochastic optimization techniques such as inverse volatility, risk parity, and variance-constrained approaches to optimally allocate assets. Our empirical results on real equity datasets demonstrate that EvoPort not only discovers rich sets of heterogeneous predictive signals but also constructs very robust and profitable portfolios. Compared to conventional alpha construction and allocation methods, our approach exhibits significant improvement in cumulative returns, Sharpe ratio, and drawdown control. We highlight the interpretability, scalability, and modularity of EvoPort, and speculate on its use as a general-purpose research pipeline for modern quantitative finance.