Portfolio optimization using deep learning with risk aversion utility function

This paper explores portfolio optimization with deep learning (DL), which can model non-linear returns that traditional methods cannot capture. While Sharpe loss addresses the risk-return trade-off in DL-based portfolio construction, it has limitations, including interpretability issues with negative PnL and biased gradients under stochastic gradient descent (SGD). We propose a new loss function based on a risk-averse utility function, which provides unbiased gradients and clear interpretation even with negative PnL. Additionally, we use DL outputs as adjustments to baseline weights, achieving improved portfolio performance. Experiments on S&P 500 data show that our method outperforms Sharpe loss-based models across several metrics, including the Sharpe ratio.