Joint return-risk DNN paper
An arXiv paper titled 'Joint Return and Risk Modeling with Deep Neural Networks' by Keonvin Park proposes combining return prediction and risk modeling inside a single deep learning framework—useful for ML-driven portfolio construction projects. The approach is pitched as a compact way to show end-to-end Python implementations for equities and risk controls. (x.com)
ArXiv submission arXiv:2603.19288 titled "Joint Return and Risk Modeling with Deep Neural Networks" lists Keonvin Park (Interdisciplinary Program in Artificial Intelligence, Seoul National University) and was submitted to q-fin.PM on March 9, 2026 (arXiv:2603.19288). (arxiv.org) The empirical sample uses daily prices for ten large-cap U.S. equities covering 2010–2024, with the paper reporting out-of-sample evaluation specifically over 2020–2024. (arxiv.org) The modelling backbone is a multivariate LSTM that the paper says captures nonlinear cross-asset dynamics and feeds a joint return-and-risk head for simultaneous forecast and covariance structure learning. (arxiv.org) Out-of-sample predictive results reported for 2020–2024 show a pointwise RMSE of 0.0264 on returns and a directional (sign) accuracy of 51.9%. (arxiv.org) At the portfolio level the author reports a "Neural Portfolio" implementation producing an annualized return of 36.4% with a Sharpe ratio of 0.91 versus equal-weight and historical mean–variance baselines in the same backtest window. (arxiv.org) The arXiv entry includes the PDF (1,927 KB) and an arXiv-issued DOI (arXiv:2603.19288) in its submission record, and the abstract page does not display an explicit external code repository link. (arxiv.org)
