Financial Engineering
Financial practice needs forecasts that respect how time series actually behave, and it needs those forecasts to survive risk management and an audit, all at once. This topic gathers the financial applications that do not fit on the portfolio optimization page: the analysis of financial time series, reinforcement learning that controls the mean and the variance, and audit sampling with statistical guarantees. Much of it overlaps with my work at Mizuho-DL Financial Technology.
Problems addressed
Financial time series contain movements that cannot be treated as stationary, and regimes that change, so a textbook model applied as it stands will fail. My co-authored book Pythonによる金融時系列データ解析 (in Japanese) works through the whole path from preprocessing to model selection and validation, with the implementations alongside. The work on reinforcement learning for investment proposes a formulation that brings the variance of returns under control, not the mean alone. A separate strand takes the practical question of how many samples an audit of financial statements has to examine, and answers it as a sequential test with a statistical guarantee.
Research thread
- 2026
- Sequential Audit Sampling with Statistical Guarantees. Statistical guarantees for audit sampling.
- 2021
- Direct Expected Quadratic Utility Maximization for Mean-Variance Controlled Reinforcement Learning (IEEE CIFEr). Reinforcement learning that controls the mean and the variance.
- 2020
- Mean-Variance Efficient Reinforcement Learning with Applications to Dynamic Financial Investment. An earlier formulation on the same line.
Relation to earlier work
The mean-variance criterion itself goes back to the mean-variance analysis of Markowitz (1952, Portfolio Selection). The reinforcement learning work carries that criterion into sequential investment decisions. How the individual papers stand in relation to the literature is explained on the Portfolio Optimization and Anomaly Detection pages.
Where it applies
The work applies directly to building and validating models at a financial institution, to designing forecasting work that runs on time series, and to bringing statistical methods into audit and risk management. For inquiries about the analysis of financial data, write to mkato-csecon@g.ecc.u-tokyo.ac.jp.