Overview
This project aims to develop two new machine learning-based strategies for dynamic asset allocation, to hopefully help pension funds avoid large drawdowns. Broadly, the first strategy is based on a funds-of-funds approach, and the second on a single-stock approach. Between them, they combine a variety of innovative aspects. Based on a deep reinforcement learning paradigm, they 1. Solve the investment problem, which is traditionally approached in two steps (estimation, then optimization), in a single step. 2. Introduce a new objective function replacing the commonly used Sharpe ratio to induce (approximate) time separability. 3. Exploit the resulting time separability of the investment problem to significantly reduce training and inference time. 4. Employ a novel differentiable approximation to the maximum drawdown metric, to facilitate penalizing downside risk. 5. Augment the state of the learning problem with factor portfolio data to incorporate the factor investing paradigm. 6. Allow for sparse portfolios (stock picking) by introducing L1 penalties.
