Paper: arXiv 2310.01319

Authors: Zhengyong Jiang, Jeyan Thiayagalingam, Jionglong Su, Jinjun Liang

Abstract

In this paper, we present a novel trading strategy that integrates reinforcement learning methods with clustering techniques for portfolio management in multi-period trading. Specifically, we leverage the clustering method to categorize stocks into various clusters based on their financial indices. Subsequently, we utilize the algorithm Asynchronous Advantage Actor-Critic to determine the trading actions for stocks within each cluster. Finally, we employ the algorithm DDPG to generate the portfolio weight vector, which decides the amount of stocks to buy, sell, or hold according to the trading actions of different clusters. To the best of our knowledge, our approach is the first to combine clustering methods and reinforcement learning methods for portfolio management in the context of multi-period trading. Our proposed strategy is evaluated using a series of back-tests on four datasets, comprising a of 800 stocks, obtained from the Shanghai Stock Exchange and National Association of Securities Deal Automated Quotations sources. Our results demonstrate that our approach outperforms conventional portfolio management techniques, such as the Robust Median Reversion strategy, Passive Aggressive Median Reversion Strategy, and several machine learning methods, across various metrics. In our back-test experiments, our proposed strategy yields an average return of 151% over 360 trading periods with 800 stocks, compared to the highest return of 124% achieved by other techniques over identical trading periods and stocks.

Complexity vs Empirical Score

  • Math Complexity: 7.5/10
  • Empirical Rigor: 6.0/10 (capped from 8.5: transaction costs only partially modelled; see penalty flags)
  • Quadrant: Holy Grail — high math complexity, high empirical rigor (effective empirical rigor capped from 8.5)

Why this score: The paper integrates advanced reinforcement learning algorithms (A3C, DDPG) with clustering, involving complex state/action spaces and optimization. It demonstrates high empirical rigor through backtesting on real datasets of 800 stocks across multiple periods with detailed metrics and comparisons against established baselines.

Research Flowchart

  flowchart TD
  A["Research Goal<br>Develop a Novel Multi-Period<br>Portfolio Management Strategy"] --> B["Data Preparation<br>800 Stocks (SSE/NASDAQ)"]
  B --> C["Clustering Stage<br>Categorize stocks by<br>financial indices using K-Means"]
  C --> D["Trading Decision Stage<br>Use A3C Algorithm to<br>determine Buy/Sell/Hold actions"]
  D --> E["Weight Allocation Stage<br>Use DDPG Algorithm to<br>generate Portfolio Weight Vector"]
  E --> F["Back-Testing Evaluation<br>Run strategy over 360 trading periods"]
  F --> G["Key Findings<br>Achieved 151% Average Return<br>Outperformed Baseline Strategies<br>(e.g., RMR, PAMR)"]

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 a novel multi‑period portfolio management strategy that combines clustering with reinforcement learning to generate trading actions and portfolio weights.

Source sentence
In this paper, we present a novel trading strategy that integrates reinforcement learning methods with clustering techniques for portfolio management in multi-period trading.

Assumes 2

  • Perfect liquidity: the market can absorb any trade size without affecting price. market
    Source sentence
    1. Perfect liquidity [13]. This assumption assumes that the market has perfect liquidity, allowing us to invest our capital in each asset with any possible proportion without considering whether there are enough order tickets in the market.
  • Zero impact cost: trades are executed immediately at the last price with no market impact. execution
    Source sentence
    2. Zero impact cost [2]. This assumption ensures that we can complete each trading immediately at the last price when the orders are put.

Data

Backtest sample: Shanghai Stock Exchange and National Association of Securiti · equities

  • 800 stocks from Shanghai Stock Exchange and NASDAQ · Yahoo Finance market-prices public
    2008/8/6 - 2016/3/9 (training, cross‑val · 800 stocks
    Source sentence
    In our experiments, a total of 800 stocks from the Shanghai Stock Exchange (SSE) or National Association of Securities Dealers Automated Quotations (NASDAQ) were utilized. These trading records can be freely downloaded from Yahoo Finance [6].

Recognised sources: Yahoo Finance

Costs

Unclear - sources disagree The platform classifier and the paper-text extraction give different answers, so no value is asserted.

Platform classifier: Partial

Paper text extraction: Modeled - 0.05% commission fee applied to each trade

Source sentence (the authors' words)
Table 6: Performance of CAD strategy in back-test of 800 stocks with 0.05% commission fee.

Finds 2

Reported performance (CAD: Clustering And Deep Reinforcement Learning): total return 151.0%

  • The CAD strategy achieves an average return of 151.29% over 360 trading periods, outperforming the next best method (A3C‑CAPM) which returns 124.06%. performance
    Source sentence
    The CAD strategy demonstrates the highest return (151.29%), while the A3C-CAPM strategy
  • Across six back‑test metrics (Final Value, Max Drawdown, Positive Days, Sharpe Ratio, Sortino Ratio, Calmar Ratio) the CAD strategy ranks first, with the second‑best results also achieved by CAD on all metrics. performance
    Source sentence
    As a result, the reinforcement learning algorithm can more effectively learn a suitable trading policy for stocks within the same cluster.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 28 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.