Paper: arXiv 2309.00626

Authors: Shuyang Wang, Diego Klabjan

Abstract

We propose an ensemble method to improve the generalization performance of trading strategies trained by deep reinforcement learning algorithms in a highly stochastic environment of intraday cryptocurrency portfolio trading. We adopt a model selection method that evaluates on multiple validation periods, and propose a novel mixture distribution policy to effectively ensemble the selected models. We provide a distributional view of the out-of-sample performance on granular test periods to demonstrate the robustness of the strategies in evolving market conditions, and retrain the models periodically to address non-stationarity of financial data. Our proposed ensemble method improves the out-of-sample performance compared with the benchmarks of a deep reinforcement learning strategy and a passive investment strategy.

Complexity vs Empirical Score

  • Math Complexity: 8.5/10
  • Empirical Rigor: 6.0/10 (capped from 7.0: transaction costs not modelled; see penalty flags)
  • Quadrant: Holy Grail — high math complexity, high empirical rigor (effective empirical rigor capped from 7.0)

Why this score: The paper employs advanced mathematical concepts like MDPs, policy ensembling, and distributional views of performance, but also includes significant empirical elements like a 4-year backtest, portfolio trading, and performance metrics.

Research Flowchart

  flowchart TD
  A["Research Goal: Improve DRL Strategy Generalization for Crypto Trading"] --> B["Data: Intraday Cryptocurrency Prices & Volumes"]
  B --> C["Methodology: Ensemble DRL with Model Selection"]
  C --> D{"Key Component: Mixture Distribution Policy"}
  D --> E["Retraining: Address Data Non-Stationarity"]
  E --> F["Outcome: Out-of-Sample Evaluation"]
  F --> G["Result: Superior Performance vs. DRL & Passive Benchmarks"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes an ensemble method with a novel mixture distribution policy to combine multiple deep reinforcement learning models for cryptocurrency portfolio trading.

Source sentence
We propose a novel mixture distribution policy for DRL trading strategies to effectively ensemble the best performance models on multiple validation periods, and achieve high returns on out-of-sample data.

Assumes 1

  • The market is liquid enough with negligible bid‑ask spread. market
    Source sentence
    We assume that the market is liquid enough with negligible bid ask spread.

Data

  • 4‑year cryptocurrency price data market-prices unknown
    07/01/2018 to 06/30/2022 · five cryptocurrencies
    Source sentence
    We use the 4-year data from 07/01/2018 t0 06/30/2022 for back‑testing. ... The

Recognised sources: Kraken

Costs

Not modeled A trading or execution result is reported without costs.

Finds 3

  • The ensemble policy achieves an annualized return of 0.9319, 47% higher than the FinRL benchmark (0.6320) and 23% higher than buy‑and‑hold (0.7571). performance
    Source sentence
    The ensemble policy method outperforms the FinRL model in the annualized return by 47% and the buy and hold strategy by 23%.
  • The ensemble method improves the Sortino ratio by 18% over FinRL (1.6218 vs 1.3762) and by 29% over buy‑and‑hold (1.6218 vs 1.2560). performance
    Source sentence
    our ensemble method still outperforms FinRL in the Sortino ratio by 18% and the buy and hold strategy by 29%.
  • The ensemble method attains a cumulative return of 7.9361 over the 4‑year test period, compared with 4.8851 for FinRL and 2.6088 for buy‑and‑hold. performance
    Source sentence
    Cumulative return 7.9361 4.8851 2.6088

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 12 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.