基于机器学习算法的外汇汇率分析与预测

发布时间：2022-01-15 12:32

　　本文提出了一种预测外汇汇率序列的新方法。使用机器学习算法,尤其是深度Q学习,通过从技术指标中提取信息,发现外汇市场汇率变化的趋势。通过使用诸如技术指标之类的财务分析方法,与以前的工作相比,我们设法大大减少了神经网络中输入特征的数量。这可以让我们拥有更加轻松的网络,并具有相似的性能。为了优化输入信息,我们对技术指标提供的信息进行了分析,并使用了聚类算法来消除不必要的信息.通过本文提出的策略,经过验证,预测出的结果准确率优于传统的投资策略（例如“买入和持有”）,并可以获得更好的回报。我们能够预测市场的趋势:我们的算法在一个下降的市场上占据了一个空头,这意味着它成功预测了未来价格的下降趋势。这里介绍的研究工作对如何有效预测市场价格的策略研究具有启发性的意义,并且有助于解决构建智能投资组合管理系统的问题。 

【文章来源】：清华大学北京市 211工程院校 985工程院校 教育部直属院校

【文章页数】：105 页

【学位级别】：硕士

【文章目录】：
摘要
ABSTRACT
NOMENCLATURE
1.INTRODUCTION
    1.1 THE FOREIGN EXCHANGE MARKET
        1.1.1 A brief introduction to the Forex market
        1.1.2 Market mechanics
        1.1.3 Principal actors of the Forex market
        1.1.4 The Forex rates in global supply chain
    1.2 ALGORITHMIC TRADING
        1.2.1 The rise of electronic markets
        1.2.2 Trading system
        1.2.3 What is algorithmic trading?
        1.2.4 Different types of algorithmic trading
    1.3 MOTIVATIONS
    1.4 THESIS CONTENT
2.LITERATURE REVIEW
    2.1 MACHINE LEARNING
    2.2 MARKOV DECISION PROCESS
        2.2.1 Definition
        2.2.2 Bellman equation
        2.2.3 Solution
    2.3 REINFORCEMENT LEARNING
    2.4 Q-LEARNING
        2.4.1 Q-Learning implementation
        2.4.2 Deep Q-learning
    2.5 TECHNICAL ANALYSIS
        2.5.1 The“Moving Average Convergence Divergence”
        2.5.2 The Pivot Point
        2.5.3 Bollinger Bands
        2.5.4 Percent B
        2.5.5 Relative Strength Index
        2.5.6 On-Balance Volume
        2.5.7 Accumulation/ Distribution indicator
        2.5.8 Chaikin's Oscillator
        2.5.9 Sharpe Ratio
    2.6 NEW FEATURES
3.BUILDING THE FRAMEWORK
    3.1 CHOICE OF CURRENCIES
    3.2 DATA ACQUISITION
    3.3 STATES AND REWARDS
    3.4 TECHNICAL INDICATORS
        3.4.1 On-Balance Volume
        3.4.2 Accumulation/Distribution Index
        3.4.3 Chaikin's Oscillator
        3.4.4 Relative Strength Index
        3.4.5 Divergence
        3.4.6 Percentage Change
        3.4.7 Consecutive variation
    3.5 CLUSTERING
        3.5.1 Objective:discretizing the state
        3.5.2 Presentation
        3.5.3 Overview of clustering methods
        3.5.4 Performance metric
        3.5.5 Further clustering choice
        3.5.6 Relative Strength Index clustering
        3.5.7 Clustering example
        3.5.8 Results
        3.5.9 Turning Point Matrix
        3.5.10 Number of states
    3.6 VALUE FUNCTION SELECTION
    3.7 ACTIONS
        3.7.1 Short position
        3.7.2 Asset allocation
    3.8 LEARNING PROCEDURE
        3.8.1 Exploration
        3.8.2 Convergence
        3.8.3 Linear decrementation of ε
        3.8.4 Discontinuous decrementation of ε
        3.8.5 A faster way to reach convergence:dyna
        3.8.6 The role of α
        3.8.7 The role of γ
        3.8.8 Further hyperparameter fine-tuning
        3.8.9 Catastrophic forgetting
        3.8.10 Experience replay
    3.9 STEP BY STEP ALGORITHMIC EXPLANATION
        3.9.1 Initialization
        3.9.2 Selection of current state and action
        3.9.3 Taking action and getting the reward
        3.9.4 Experience replay
        3.9.5 Evaluation of the policy and ε update
4.EXPERIMENT
    4.1 INTRODUCTION
    4.2 BENCHMARK
        4.2.1 Profit and Loss
        4.2.2 Random pick
        4.2.3“Buy and Hold”
    4.3 HYPERPARAMETERS
        4.3.1 Technical analysis
        4.3.2 Reinforcement learning
        4.3.3 Neural network
    4.4 SANITY CHECK
    4.5 DEEP Q-LEARNING–FIRST EXPERIMENT
    4.6 DEEP Q-LEARNING–SECOND EXPERIMENT
        4.6.1 Training
        4.6.2 Testing
    4.7 BENCHMARK COMPARISON
    4.8 RESULTS DISCUSSION
5.CONCLUSIONS
    5.1 RESEARCH RESULTS
    5.2 FURTHER INVESTIGATION
        5.2.1 Fine-tuning difficulties
        5.2.2 Asset allocation
        5.2.3 Computational power
        5.2.4 Number of data points
        5.2.5 State representation
        5.2.6 Volume information
REFERENCES
ACKNOWLEDGEMENTS
APPENDIX A
APPENDIX B
APPENDIX C
APPENDIX D
RESUME



本文编号：3590623