Methodology of Learning Optimal Decisions from Market Data in Financial Technology

随着西方世界及其他地区养老金危机的加深，高效有效地解决长期动态资产配置问题变得比以往任何时候都更加关键。从固定收益到固定缴款的趋势对低收入人口来说尤其成问题。自动养老金投资可能会增强退休信心，特别是让女性和经济困难的人受益。总的来说，机器学习能否为动态资产配置提供一种新的解决方案？这是一个将金融数学和机器学习优化的专业知识结合起来最能解决的问题。加拿大皇家银行和其他类似组织最近创建的北极星人工智能研究所表明了金融技术中对机器学习研究的必要性。数学金融的标准方法总是从从市场数据中提取信息开始。直到最近，这种方法一直是一个两级解决方案。首先，从数据中估计出参数随机模型。其次，根据假设的参数模型计算相关答案，例如投资策略。在数学金融学中，这种两层式的方法遇到了严重的障碍。由于市场动态复杂多变，假设的模型不可避免地是错误的。当随后得出最优决策时，假设模型中的误差往往被放大。这限制了计算解的适用性。此外，许多动态投资组合优化和风险管理问题，遵循经典的参数方法，即使模型有缺陷，也非常难以解决。一种新的范式正在出现，它承诺提供一种全新的数据驱动解决方案。通常，相关的财务解决方案是最优决策。数据驱动优化的目的是直接从市场数据中确定最优解，代表了市场的随机性。这种直接优化框架绕过了进行错误模型假设的中间步骤。令人惊讶的是，这种方法通常避免了求解基于模型的最优控制问题的计算挑战。为了在这个新兴的跨学科领域取得成功，在优化、金融建模和机器学习方面的研究经验至关重要。我在这三个领域的研究专长使我能够独一无二地探索这一令人兴奋的领域的格局。这项研究的目标是开发一种数据驱动的方法，专门关注金融市场，直接从市场数据中学习更复杂但实际相关的动态决策。具体地说，我们将(A)开发解决(非参数)基于样本路径的优化问题的高效算法，并为财务策略设定合适的目标；(B)确保最优数据驱动模型是有效和稳健的，但通过彻底的经验验证是可解释的；(C)增加足够高质量的数据。

As pension crises deepen in the western world and beyond, efficiently and effectively solving long term dynamic asset allocation problems becomes more critical than ever. The trend from Defined Benefit to Defined Contribution is particularly problematic for lower income population. Automatic pension investing can potentially boost retirement confidence, benefiting particularly women and financially disadvantaged people. Can machine learning offer a new solution to dynamic asset allocation in general? This is a question which can best be addressed by combining expertise from the financial mathematics and machine learning optimization. The need for research on machine learning in financial technology is illustrated by recent creations of Borealis AI Institute by RBC and other similar organizations. The standard approach in mathematical finance has always started from  extracting information from market data. Until recently, the approach has been a 2-tiered solution. First, a parametric stochastic model is estimated from the data. Second, relevant answers, e.g., investment strategies, are computed from the assumed parametric model. Serious obstacles have been encountered from this 2-tirered approach in mathematical finance. Since the market dynamics are complex and changeable, the assumed model is inevitably erroneous. When optimal decisions are subsequently derived, errors in the assumed model are often magnified. This limits the applicability of the computed solution. In addition, many dynamic portfolio optimization and risk management problems, following the classical parametric approach, are incredibly hard to solve, even though the model is flawed. A new paradigm is emerging with the promise of a brand new data driven solution. Often, the relevant financial solutions are optimal decisions. Data driven optimization aims to determine the optimal solutions directly from market data, representing market stochasticity. This direct optimization framework bypasses the intermediate step of making erroneous model assumptions. Surprisingly, this approach often avoids computational challenges from solving model based optimal control problems. To succeed in this emerging interdisciplinary field, research experience in optimization, financial modelling, and machine learning is crucial. My research expertise in all three fields has uniquely positioned me to explore the landscape of this exciting area.  The goal of this research is to develop a data driven methodology, focusing specifically on financial markets , to learn more complex but practically relevant dynamic decisions directly from market data. Specifically we will  (a) develop efficient algorithms solving (non-parametric) sample path based optimization problems with suitable objectives for financial strategies; (b) ensure that optimal data driven model is effective and robust but interpretable through thorough empirical validation; (c) augment sufficiently high quality data.