Deep reinforcement learning for automated stock trading: an ensemble strategy

Three actor-critic algorithms — PPO, A2C, and DDPG — are trained separately and combined into one strategy that picks among them by rolling Sharpe ratio, so as to “inherit and integrate the best features of the three algorithms, thereby robustly adjusting to different market situations”. It is tested on the 30 Dow Jones constituents against the DJIA index and a minimum-variance portfolio. The reported outperformance is a 2020 backtest on that data, not a live track record.

The method is still runnable today: it lives inside FinRL (MIT) as DRLEnsembleAgent in finrl/agents/stablebaselines3/models.py, driven by finrl/applications/stock_trading/ensemble_stock_trading.py, with a companion notebook at examples/FinRL_Ensemble_StockTrading_ICAIF_2020.ipynb. Two published pointers to the code are dead ends — the standalone repository named in the arXiv abstract returns 404, as does the notebook path in FinRL’s own README — so use the paths above.

The proceedings version (pp. 1–8) sits behind ACM copyright; an openly readable author version was posted in November 2025 as arXiv:2511.12120. The conference’s own 2020 program lists the work under a variant title, “Deep Ensemble Reinforcement Learning for Automated Stock Trading”; the proceedings title is used here.