算法交易是利用计算机算法进行证券交易的一种创新交易手段。机器学习专门研究从数据中学习出有效的预测模型，从而被用来解决算法交易中的问题。本项目针对当前组合算法交易策略在仿真测试中性能不足以支撑合理交易成本的问题，设计新策略来提升回溯测试的性能。主要研究内容包括：（1）在现有基于机器学习的交易框架上，设计综合考虑金融市场动量和均值回归特性的策略，以提升当前策略在零交易成本时的性能；（2）并利用机器学习正则化的理念，设计一套主动考虑交易成本的组合算法交易策略，以提升现有策略在非零交易成本时的性能；（3）针对现有策略没有解释收益的薄弱环节，拟运用常见的定价模型来解释策略的收益。另外，本项目拟公开研究中实现的仿真交易系统及策略，以供同行参考。本项目对于深入理解金融市场的特性，探索组合交易策略的机理，促进市场的有效性和研究计算方法论在金融中的应用具有一定的理论意义和实践指导价值。

Algorithmic trading, which utilizes computational algorithms for trading securities, is an innovative trading techniques. Machine learning specializes on learning efficient predictive models from data, and thus becomes a nature candidate for algorithmic trading. As the state of the art portfolio trading algorithms cannot sustain reasonable transaction costs in existing back tests, this project aims to design new strategies that improve existing strategies’ performance. Firstly, based on existing online machine learning framework for trading, we design a new strategy that utilizes both properties of momentum and mean reversion in financial markets and is expected to show better performance than the state of the art. Secondly, utilizing the regularization technique in machine learning, we design a new strategy that actively evaluates unknown transaction costs and is expected to improve the strategies’ performance in case of non-zero transaction costs. Thirdly, although existing strategies have shown amazing excess return in back tests, they did not explain it. As a result, we would like to explain the strategies’ excess return using common asset pricing models. In addition, we will publicize our simulation system and strategies, including their source codes. This project will help us understand the properties of financial markets, explore the mechanism of algorithmic trading strategies, improve market efficiency and study the application of computational methods in finance.