This research paper introduces a novel deep learning framework for portfolio optimization that directly maximizes the Sharpe ratio without requiring traditional expected return forecasts. The approach updates model parameters to optimize portfolio weights directly, representing a significant departure from conventional portfolio optimization methods that typically rely on separate return prediction and optimization steps. By trading Exchange-Traded Funds (ETFs) of market indices rather than individual assets, the framework reduces the asset selection complexity while leveraging the robust correlations between different asset class indices.

The methodology was tested from 2011 through April 2020, including the financial instability period of early 2020, and demonstrated superior performance compared to a wide range of alternative algorithms. The paper includes comprehensive sensitivity analysis to understand input feature relevance and evaluates performance under varying transaction cost rates and different risk levels achieved through volatility scaling. This work contributes to the growing literature on machine learning applications in finance by providing a practical, end-to-end deep learning approach to portfolio optimization that shows robustness across different market conditions.