[Seminar] MLDS Unit Seminar 2026 by Dr. Riku Green, University of Bristol

Speaker: Mr. Riku Green, the University of Bristol

Title: Machine Learning for Forecasting Multiple Steps Ahead 

           -Strategy choice, model architecture, recursive composition, and evaluation under uncertainty-

Abstract:
Multi-step forecasting asks machine learning models to predict not just a single next value, but an entire future trajectory. This makes the forecasting strategy—the way multiple future values are generated—a central modelling choice. In this talk, I will focus on three often-hidden aspects of this problem: how forecasts are formed across the horizon, how these strategies interact with model architecture, and how the forecast is evaluated under uncertainty.
Together, these ideas reframe multi-step forecasting as a joint problem of strategy choice, model choice, and objective choice. I will show how this perspective helps explain when forecasting strategies matter empirically, why standard model comparisons can miss important interactions, and how strategy choice can help navigate the trade-off between point accuracy and distributional realism under long-horizon uncertainty.
The talk is aimed at a broad machine learning audience and does not assume prior background in forecasting.

Short bio:
Riku Green is a final-year PhD student in the Interactive AI CDT at the University of Bristol. His research focuses on foundational machine learning methods for multi-step time-series forecasting, especially forecasting strategy design, uncertainty, statistical learning, and model selection.