Deep reinforcement learning for portfolios, replicated
Reinforcement learning appears on more than 700 recent papers, and 41 of them apply it to portfolios, trading or hedging, almost all in AI and operations research journals. The results are rarely reproduced and rarely compared with simple rules after costs. Which of the reported gains survive replication on new data?
What prompted this
Since October 2025, 738 papers carry a reinforcement-learning keyword (577 in AI journals, 120 in operations research): "reinforcement learning" on 255, "deep reinforcement learning" on 191, "multi-agent reinforcement learning" on 97. Forty-one of them also carry a portfolio, trading, stock, asset-allocation, hedging, market-making, finance or crypto keyword; "portfolio optimization" appears on 153 papers in the archive and "portfolio management" on 18.
Questions
- When the published agents are retrained on the same data with different random seeds and then run on the two years after each paper's sample, how often do they beat equal weights and mean-variance after transaction costs?
- How much of the reported performance comes from look-ahead in data preparation and from tuning on the test period?
- Does the hedging use case, with a clearer objective, replicate better than the return-maximizing one?
Data that could help
The code and data of recent papers, daily prices for the post-sample period, and realistic transaction cost models.
Next step
Replicate ten published deep reinforcement learning portfolio papers with multiple seeds and evaluate them on the two years after each sample, after costs.