毕业生知识库

金融科技作为新质生产力的一个重要组成部分，可以大幅提升金融服务的效率和安全性，其中资产管理是金融科技量化投资方向的一个重要应用，需要投资者利用算法和模型系统化地进行投资决策。金融科技领域下的资产管理被定义为一种基于数学模型、统计分析和人工智能的投资方式。资产管理强调的是通过算法和大数据分析来识别投资机会和风险，从而实现资金的有效管理和投资收益的最大化。论文强调了资产管理在全球金融市场中的重要性和应用范围，特别是在算法交易和资产价格趋势预测两方面。算法交易关注如何通过算法自动执行交易决策，以提高交易效率和获取更高收益；资产价格趋势预测则致力于利用历史数据和模型预测资产价格的未来走势，帮助投资者做出更准确的投资选择。论文的主要工作和创新点归纳如下：

（一）考虑金融数据分布对算法交易和资产价格趋势预测的增益

当前研究在处理金融数据分布变化方面有所不足。金融市场数据的信噪比低，随着金融数据的漂移，金融数据分布发生变化，导致了市场模式的不同。不同金融数据分布会导致多种市场情况，而忽略数据中的多个模式会破坏模型的性能。算法交易的成功往往依赖于交易策略的精确设计。近年来，基于强化学习的策略已经显示出解决算法交易问题的卓越能力。第三章提出了一个新的模型（a mixture of actors reinforcement learning method by optimal transport，MOT），在强化学习的基础上结合最优传输算法来解决以上问题。为了建模多种交易模式，模型设计了多个智能体。进一步地，模型引入最优传输算法，通过增加损失项将不同的样本分配给合适的智能体。资产价格趋势预测问题常用图模型来解决。在图模型中，传统的因果模型不能解释股票数据的分布变化，而数据分布变化的因果关系建模在金融领域非常重要。
本文提出了一个利用了对比学习方法的因果关系增强的多视角图模型CMG（a causal-enhanced multi-view temporal graph model，CMG），旨在改善资产价格趋势预测的准确性。CMG模型特别关注于挖掘随数据分布变化而变化的因果关系以及不同关系间的共性和差异。CMG模型通过恢复非平稳因果框架并识别因果方向来发现分布偏移数据的因果关系。实验表明考虑金融数据分布对算法交易和资产价格趋势预测问题有较好的增益。

（二）专家知识结合强化学习

金融市场的噪声大，潜在关系复杂，信息繁冗，基于强化学习的智能体在探索和利用之间难以平衡，引入专家知识和预训练方法可以改善模型初始化效果。具体的，MOT模型在模仿学习模块引入专家知识，智能体从专家策略轨迹中采样学习知识，一定程度上提高了探索的效率；在预训练模块，智能体模仿专家策略以对齐动作，再通过模仿学习使之初始化在较优的位置。在真实的期货市场上的实验结果表明，MOT在不同市场模式下具有卓越收益能力的同时还可以平衡风险。此外，消融实验验证了所提出方法中各模块的有效性。

（三）基于图神经网络挖掘金融资产间信息传播

探索更贴切的图神经网络方法去更合适地建模金融资产间信息交互会给模型带来增益。对于资产趋势预测问题来说，因果关系图可以揭示重要的隐藏信息，如股票之间的隐藏股权关系和供应链信息。这些发现对于理解和预测股票趋势至关重要。此外，通过在相关关系图模块和因果图模块中传播公司的嵌入信息，并构建多视图对比学习任务，CMG能够捕捉不同图关系之间的粗粒度共性和细粒度差异。实验结果和实际股市中的投资模拟表明，CMG模型在股价预测问题上表现出色，超越了基线模型。文章还讨论了模型的不足之处，如对高质量大数据的依赖和模型泛化能力的局限，并对未来的研究方向提出了展望。

综上所述，这篇论文在考虑金融数据分布对算法交易和资产价格趋势预测的增益，专家知识结合强化学习，和基于图神经网络挖掘金融资产间信息传播三个核心问题上做出了有意义的探索和贡献。通过引入最新的技术和方法，不仅提升了模型的预测准确度和交易效率，还为处理金融数据的复杂性和动态变化提供了新的解决思路。

Financial technology(Fintech), as an important component of new productive forces, can significantly enhance the efficiency and security of financial services. Among these, wealth management is a crucial application in the quantitative investment direction of fintech, requiring investors to systematically make investment decisions using algorithms and models. Wealth management in the fintech field is defined as an investment method based on mathematical models, statistical analysis, and artificial intelligence. It emphasizes identifying investment opportunities and risks through algorithms and big data analysis to achieve effective fund management and maximize investment returns. The paper highlights the importance and application scope of wealth management in the global financial market, particularly in algorithmic trading and asset price trend prediction. Algorithmic trading focuses on how to automatically execute trading decisions through algorithms to improve trading efficiency and achieve higher returns; asset price trend prediction aims to use historical data and models to forecast future asset price trends, helping investors make more accurate investment choices. The main work and innovations of the paper are summarized as follows:

(1) Considering the Gain of Financial Data Distribution on Algorithmic Trading and Asset Price Trend Prediction

Current research has some deficiencies in handling changes in financial data distribution. The signal-to-noise ratio of financial market data is low, and as financial data drifts, the distribution of financial data changes, leading to different market patterns. Different financial data distributions can result in various market situations, and ignoring multiple patterns in the data can impair model performance. The success of algorithmic trading often relies on the precise design of trading strategies. In recent years, strategies based on reinforcement learning have shown excellent capabilities in solving algorithmic trading problems. In Chapter 3, we propose a new model (a mixture of actors reinforcement learning method by optimal transport, MOT), which combines optimal transport algorithms with reinforcement learning to address the above issues. To model multiple trading patterns, we designed multiple agents. We further introduced the optimal transport algorithm, assigning different samples to appropriate agents by adding a loss term. The asset price trend prediction problem is often solved using graph models. In graph models, traditional causal models cannot explain the distribution changes in stock data, and modeling the causal relationships of data distribution changes is very important in the financial field. We propose a causal-enhanced multi-view temporal graph model (CMG) using contrastive learning methods to improve the accuracy of asset price trend prediction. The CMG model particularly focuses on uncovering the causal relationships that change with data distribution and the commonalities and differences between different relationships. The CMG model discovers the causal relationships of distribution-shifted data by restoring the non-stationary causal framework and identifying causal directions. Experiments show that considering financial data distribution has good gains for algorithmic trading and asset price trend prediction problems.

(2) Combining Expert Knowledge with Reinforcement Learning

The financial market is noisy, with complex potential relationships and abundant information. Agents based on reinforcement learning find it difficult to balance exploration and exploitation. Introducing expert knowledge and pre-training methods can improve model initialization effects. Specifically, the MOT model introduces expert knowledge in the imitation learning module, where agents sample and learn knowledge from expert strategy trajectories, improving exploration efficiency to some extent; in the pre-training module, agents imitate expert strategies to align actions and then initialize in a better position through imitation learning. Experimental results in the real futures market show that MOT has excellent return capabilities under different market patterns while balancing risks. Additionally, ablation experiments verify the effectiveness of each module in the proposed method.

(3) Mining Information Propagation Between Financial Assets Based on Graph Neural Networks

Exploring more appropriate graph neural network methods to better model the information interaction between financial assets will bring gains to the model. For asset trend prediction problems, causal relationship graphs can reveal important hidden information, such as hidden equity relationships and supply chain information between stocks. These findings are crucial for understanding and predicting stock trends. Moreover, by propagating company embedding information in the related relationship graph module and causal graph module and constructing multi-view contrastive learning tasks, CMG can capture the coarse-grained commonalities and fine-grained differences between different graph relationships. Experimental results and investment simulations in the actual stock market show that the CMG model performs excellently in stock price prediction, surpassing baseline models. The paper also discusses the model's shortcomings, such as reliance on high-quality big data and limitations in model generalization ability, and proposes future research directions.

In summary, this paper makes meaningful explorations and contributions in three core issues: considering the gain of financial data distribution on algorithmic trading and asset price trend prediction, combining expert knowledge with reinforcement learning, and mining information propagation between financial assets based on graph neural networks. By introducing the latest technologies and methods, it not only improves the prediction accuracy and trading efficiency of the model but also provides new solutions for handling the complexity and dynamic changes of financial data.