Proposal Description

Forecasting under uncertainty is among the oldest and most consequential problems in statistics, yet it remains an active frontier as the environments in which prediction is required grow more complex, dynamic, and data-rich. This session brings together four speakers whose work reflects the breadth of modern forecasting methodology — spanning financial time series and AI-generated investment portfolios, environmental systems such as coastal wave dynamics, and the stochastic reliability of components operating in random environments. The talks draw on a diverse toolkit: classical statistical models, deep learning architectures including LSTMs and transformer-based large language models, Bayesian dynamic models, and information-theoretic measures. Together, they address a shared challenge: how to produce reliable, transparent, and actionable forecasts when the underlying environment is uncertain, evolving, or subject to structural change. The session is particularly timely given the rapid expansion of AI-driven forecasting tools into high-stakes domains — from financial markets and renewable energy to marine navigation and engineering reliability — where the statistical properties of forecasts, and the uncertainty surrounding them, have direct consequences for safety, policy, and investment. The talks approach forecasting from distinct but complementary angles, collectively demonstrating the range and vitality of modern statistical forecasting.

Ensor and co-authors examine the intersection of time series methodology and AI-driven financial decision-making. As auto-investing and generative AI portfolio tools proliferate, the talk asks what statistical rigor is needed to evaluate and bring transparency to these systems: embedded assumptions, obscured risks, and methodological standards. The talk gives a critical appraisal of current practice and a call for statistically sound standards in a rapidly evolving domain.

Ravishanker and co-authors discuss forecasting significant wave heights, with application to coastal resilience, marine navigation, and wave energy. Three frameworks are developed and benchmarked: classical statistical models, deep learning (LSTM), and modified LLMs, all incorporating historical wave height and coastal wind data. The talk explains implications for coastal infrastructure and real estate risk assessment.

Soyer and co-authors address forecasting from a reliability and information-theoretic perspective, studying how lifetime distributions of system components change when they transition from a controlled testing environment to a real-world operating environment. Using proportional hazards with frailty models as foundations, they introduce novel dynamic formulations for environments governed by Markov and geometric processes, and information measures for comparing lifetime distributions across environments, with application to maintenance planning, warranty modeling, and system reliability forecasting.

Holan and co-authors describe efficient Echo State Networks (ESNs) for capturing nonlinear temporal dynamics and generating forecasts from area-level spatio-temporal data. Spatio-temporal area-level datasets play a critical role in official statistics, providing valuable insights for policy-making and regional planning. Incorporating approximate graph spectral filters improves forecast accuracy while preserving the model's computational efficiency during training. They apply their methods to Eurostat's tourism occupancy dataset.