光华讲坛——社会名流与企业家论坛第6487期
主 题：Data Driven Optimally Outperforming Without Dynamic Programming
主讲人：加拿大滑铁卢大学 李玉英教授
主持人：统计学院 兰伟教授
时间：5月26日 10:00-11:00
举办地点：柳林校区通博楼TB212会议室
主办单位：统计研究中心 统计学院 科研处
主讲人简介：
Yuying Li is a professor at the Cheriton School of Computer Science at the University of Waterloo in Canada. Prior to joining UW, she was a senior research associate at Cornell University 1988-2005. She is also the recipient of the 1993 Leslie Fox first Prize in numerical analysis competition held at Oxford England. Her research interests include financial data science, machine learning, computational finance, and computational optimization. Li is currently an associate editor for Journal of Computational Finance, as well as Journal of Finance and Data Science.
李玉英,加拿大滑铁卢大学Cheriton计算机科学学院的教授。在加入滑铁卢大学之前，她于1988年至2005年在康奈尔大学担任高级研究员。她也是1993年在英国牛津大学举行的Leslie Fox数值分析比赛一等奖的获得者。她的研究兴趣包括金融数据科学、机器学习、计算金融和计算优化。她目前是Journal of Computational Finance和Journal of Finance and Data Science的副主编。
内容简介：
We propose a data driven learning framework to learn stochastic optimal asset allocation strategies without dynamic programming (DP). Our proposed neural network (NN) Policy Function Approximation (PFA) approach learns the optimal dynamic policies directly from data samples, which can either come from simulating a parametric model or resampling market observations directly.
Resampling non-parametrically approximates the sampling distribution of the least prejudiced empirical distribution. We use block resample market data to generate training and testing data sets. We formally establish additionally that using block resampling, for typical data lengths and expected block sizes in finance, the probability of repeating a sample path, even with 1,000,000 random path draws, is negligible.
For outperforming a benchmark, we propose suitable objective functions, which are consistent with asset allocation performance evaluation metrics in financial industry. Specifically, we propose to use information ratio (IO) and tracking differences as objective functions. Using the proposed data driven approach, objective functions, and block resampled market data, we discover robust and higher performance strategies over a benchmark by allocating over equity and bond market indices, as well as factor investing assets. We contrast and assess testing outperformance based on terminal wealth distributions.
主讲人提出了一个数据驱动的学习框架来学习没有动态规划（DP）的随机最优资产配置策略。主讲人提出的神经网络（NN）策略函数近似（PFA）方法直接从数据样本中学习最佳动态策略，这些数据样本既可以来自模拟参数模型，也可以直接对市场观察进行重采样。
重采样非参数地近似于最小偏见经验分布的抽样分布。主讲人使用区块重采样市场数据来生成训练和测试数据集。我们还正式确定，使用区块重采样，对于金融中的典型数据长度和预期区块大。词褂 1000000 个随机路径绘制，重复样本路径的概率也可以忽略不计。
为了跑赢基准，主讲人提出了合适的目标函数，这些函数与金融行业的资产配置绩效评估指标一致。具体来说，主讲人建议使用信息比率（IO）和跟踪差异作为目标函数。使用提出的数据驱动方法、目标函数和块重采样市场数据，主讲人通过配置股票和债券市场指数以及因子投资资产，在基准上发现了稳健且性能更高的策略。主讲人根据终端财富分布对比和评估测试表现。
