Portfolio Optimization
A model that forecasts future returns can be wrong, so portfolio selection has to carry the uncertainty of the forecast with it. What I study here is how to join prediction to allocation, by constructing prediction intervals and by combining several models. The main result is Conformal Predictive Portfolio Selection (Kato, 2024, arXiv:2410.16333).
Problems addressed
Traditional portfolio selection, the mean-variance approach above all, estimates distributional quantities such as the mean and variance of returns from past data and then works with those estimates. The prediction model behind them carries error, and an allocation that rests on a point forecast alone has no defense against the way the model misses. I designed a scheme that builds a prediction interval for the portfolio return by conformal prediction and picks the allocation from that interval. It runs on any prediction model, from an autoregressive model to a random forest or a neural network.
This thread also includes the work on predictive synthesis, where several forecasting models are combined in a Bayesian way before the allocation is decided, and the work on reinforcement learning that keeps the variance of returns under control alongside the mean while it learns.
Core research
The Conformal Predictive Portfolio Selection page explains how the prediction interval is constructed and how selection from the interval is designed.
Research thread
- 2024
- Bayesian Portfolio Optimization by Predictive Synthesis. Deciding the allocation after combining several forecasting models in a Bayesian way.
- 2020
- Mean-Variance Efficient Reinforcement Learning with Applications to Dynamic Financial Investment. A formulation of reinforcement learning under the mean-variance criterion.
Relation to earlier work
Portfolio selection starts from the mean-variance theory of Markowitz (1952, Portfolio Selection). Conformal prediction comes from the other direction. It gives finite-sample guarantees for prediction intervals without assuming a distribution, and work such as the adaptive method of Gibbs and Candès (2021, Adaptive Conformal Inference Under Distribution Shift) has made it usable for time series and under distribution shift. My work joins the two lines and hands the uncertainty of the prediction model to the allocation decision as an interval that carries a guarantee.
Where it applies
When a prediction model is introduced or updated, the operational questions are how far its forecast can be trusted and how the allocation behaves once the forecast misses. That is the setting these methods are built for. For inquiries about designing a scheme for asset management or risk management, write to mkato-csecon@g.ecc.u-tokyo.ac.jp.