Portfolio Optimization and Financial Engineering
Bringing prediction uncertainty, rather than point forecasts, into investment decisions — and applying statistical machinery to problems in financial practice.
- portfolio selection
- conformal prediction
- reinforcement learning
- financial engineering
The question
Portfolio construction consumes forecasts: expected returns, variances, quantiles. Those forecasts come from models, and the models are uncertain, but the allocation step usually treats them as if they were known. The work here asks how the uncertainty in a prediction should propagate into the decision it informs.
Predictive uncertainty in allocation
- Conformal Predictive Portfolio Selection forms conformal prediction intervals for future portfolio returns and selects the portfolio from the intervals. Conformal prediction gives coverage without distributional assumptions, so the framework accepts any predictor — AR models, random forests, neural networks — without changing the selection step.
- Bayesian portfolio optimization by predictive synthesis (IIAI-AAI 2024) combines several predictive models rather than selecting one, propagating the disagreement between them into the allocation.
Control of risk in sequential decisions
Direct expected quadratic utility maximization for mean-variance controlled reinforcement learning (arXiv:2010.01404) formulates mean-variance control as direct utility maximization, avoiding the difficulties that variance penalties create for standard reinforcement learning objectives.
Applied work
Other work in financial practice covers time series analysis of economic conditions — including analyses of the Cabinet Office Economy Watchers Survey using independent component analysis and linear discriminant analysis (IIAI-AAI 2024, CDEF 2025) — potential-customer identification from selectively labeled records, and audit sampling with statistical guarantees. Much of this is joint work carried out at Mizuho-DL Financial Technology.