Paper: arXiv 2404.08935
Authors: Zhenglong Li, Vincent Tam
Abstract
In recent years, deep or reinforcement learning approaches have been applied to optimise investment portfolios through learning the spatial and temporal information under the dynamic financial market. Yet in most cases, the existing approaches may produce biased trading signals based on the conventional price data due to a lot of market noises, which possibly fails to balance the investment returns and risks. Accordingly, a multi-agent and self-adaptive portfolio optimisation framework integrated with attention mechanisms and time series, namely the MASAAT, is proposed in this work in which multiple trading agents are created to observe and analyse the price series and directional change data that recognises the significant changes of asset prices at different levels of granularity for enhancing the signal-to-noise ratio of price series. Afterwards, by reconstructing the tokens of financial data in a sequence, the attention-based cross-sectional analysis module and temporal analysis module of each agent can effectively capture the correlations between assets and the dependencies between time points. Besides, a portfolio generator is integrated into the proposed framework to fuse the spatial-temporal information and then summarise the portfolios suggested by all trading agents to produce a newly ensemble portfolio for reducing biased trading actions and balancing the overall returns and risks. The experimental results clearly demonstrate that the MASAAT framework achieves impressive enhancement when compared with many well-known portfolio optimsation approaches on three challenging data sets of DJIA, S&P 500 and CSI 300. More importantly, our proposal has potential strengths in many possible applications for future study.
Complexity vs Empirical Score
- Math Complexity: 8.0/10
- 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 mathematics including transformer architectures, multi-agent systems, attention mechanisms, and POMDP formulation (high math complexity). It also demonstrates strong empirical rigor with backtests on three major datasets (DJIA, S&P 500, CSI 300) and comparisons against established methods.
Research Flowchart
flowchart TD A["Research Goal:<br>Develop attention-based<br>ensemble framework for<br>portfolio optimisation"] --> B["Data Preparation<br>Directional Change &<br>Price Series Data<br>(DJIA, S&P 500, CSI 300)"] B --> C["Multi-Agent Architecture<br>Multiple Trading Agents<br>Self-Adaptive Mechanism"] C --> D["Spatial-Temporal Analysis<br>Attention-based Cross-Sectional<br>Analysis Module<br>Attention-based Temporal<br>Analysis Module"] D --> E["Portfolio Generation<br>Ensemble Fusion of<br>Spatial-Temporal Information<br>Balance Returns & Risks"] E --> F["Key Outcomes<br>Improved Signal-to-Noise Ratio<br>Reduced Biased Trading Actions<br>Superior Performance vs.<br>Baseline Methods"]
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 MASAAT, a multi‑agent self‑adaptive portfolio optimisation framework that integrates attention mechanisms, directional‑change features and temporal analysis to generate ensemble portfolios.
Source sentence
a multi-agent and self-adaptive portfolio optimisation framework integrated with attention mechanisms and time series, namely the MASAAT
Assumes 4
- The portfolio optimisation problem is modelled as a partially observable Markov decision process (POMDP). model
Source sentence
the PM problem can be modelled as a POMDP
- Reward is defined as the sum of the logarithm of returns over the trading horizon. model
Source sentence
the sum of the logarithm of returns as the reward function of this work is defined as J(θ) = 1/T log C0
- Training data consists of ten years of historical prices (Jan 2008–Dec 2017) and validation/test periods are later years. data
Source sentence
all the models are trained on ten-year data from January 2008 to December 2017 and then validated by the subsequent data set of three years. Afterwards, the test set from January 2021 to December 2023 is applied
- Only the top 10 stocks of each index are used as the investment universe. data
Source sentence
the top 10 stocks of each market index are selected as the available stocks in a portfolio
Data
- Dow Jones Industrial Average (DJIA) index data market-prices unknown
Source sentence
three challenging data sets of Dow Jones Industrial Average (DJIA), S&P 500 and CSI 300 indexes from U.S. and China markets
- S&P 500 index data market-prices unknown
Source sentence
three challenging data sets of Dow Jones Industrial Average (DJIA), S&P 500 and CSI 300 indexes from U.S. and China markets
- CSI 300 index data market-prices unknown
Source sentence
three challenging data sets of Dow Jones Industrial Average (DJIA), S&P 500 and CSI 300 indexes from U.S. and China markets
Costs
Not modeled
Finds 2
- MASAAT achieves the highest annualised return of 14.28% on the DJIA test period, at least 1.13% higher than the best baseline. performance
Source sentence
the proposed MASAAT framework achieves the highest AR of 14.28% during the backtesting period of three years, which is at least 1.13% higher than that of the best baseline frameworks achieved by the DPM
- The framework demonstrates strong capability to balance overall returns and portfolio risks across DJIA, S&P 500 and CSI 300 compared with state‑of‑the‑art approaches. performance
Source sentence
the empirical results on three challenging data sets of DJIA, S&P 500, and CSI 300 market indexes reveal the strong capability of the proposed MASAAT framework to balance the overall returns and portfolios risks against the state-of-the-art approaches
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 15 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.
Cited References (37)
- Universal portfolios (1991)
- On-line portfolio selection using multiplicative updates (1998)
- PAMR: Passive aggressive mean reversion strategy for portfolio selection (2012)
- Nonparametric kernel-based sequential investment strategies (2006)
- A deep reinforcement learning framework for the financial portfolio management problem (2017) — link
- Cost-sensitive portfolio selection via deep reinforcement learning (2020)
- Relation-aware transformer for portfolio policy learning (2021)
- DeepTrader: a deep reinforcement learning approach for risk-return balanced portfolio management with market conditions Embedding (2021)
and 29 more
- Addressing function approximation error in actor-critic methods (2018)
- Machine learning and portfolio optimization (2018)
- A brief review of portfolio optimization techniques (2023)
- How are reinforcement learning and deep learning algorithms used for big data based decision making in financial industries–A review and research agenda (2022)
- An innovative neural network approach for stock market prediction (2020)
- Applications of deep learning in stock market prediction: recent progress (2021)
- Forecasting stock prices using stock correlation graph: A graph convolutional network approach (2021)
- A novel graph convolutional feature based convolutional neural network for stock trend prediction (2021)
- Trading algorithms built with directional changes (2019)
- Generating directional change based trading strategies with genetic programming (2015)
- Evolving trading strategies using directional changes (2017)
- Attention in natural language processing (2020)
- Swin transformer: Hierarchical vision transformer using shifted windows (2021)
- Speech-transformer: a no-recurrence sequence-to-sequence model for speech recognition (2018)
- Attention is all you need (2017)
- Llama 2: Open foundation and fine-tuned chat models (2023) — link
- GPT-4 Technical Report (2023) — link
- Online portfolio selection: A survey (2014)
- Stockformer: Learning hybrid trading machines with predictive coding (2023)
- Reinforcement-learning based portfolio management with augmented asset movement prediction states (2020)
- An Adaptive News-Driven Method for CVaR-sensitive Online Portfolio Selection in Non-Stationary Financial Markets. (2021)
- Graph-based stock correlation and prediction for high-frequency trading systems (2022)
- The kelly growth optimal portfolio with ensemble learning (2019)
- Portfolio rebalancing based on a combined method of ensemble machine learning and genetic algorithm (2023)
- Developing A Multi-Agent and Self-Adaptive Framework with Deep Reinforcement Learning for Dynamic Portfolio Risk Management — → 2402.00515
- An attentive survey of attention models (2021)
- Efficient transformers: A survey (2022)
- A general survey on attention mechanisms in deep learning (2021)
- Recent advances in reinforcement learning in finance (2023)
1 of 37 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.