Event Details

State Space Thinking in the Age of Foundation Models

In this talk, I reflect on a personal journey through decades of research in time series analysis, one that began at Lancaster University. The focus is on state space models for forecasting and on the often overlooked importance of careful model design. I will illustrate the power and versatility of state space models, emphasizing the many forms they can take and the flexibility they offer to practitioners. Both the literature and practice of forecasting with state space models have been largely dominated by the single source of error (SSOE) framework, most notably in the ETS family. The SSOE approach relies on a single disturbance term shared across all model equations. In contrast, multiple source of error (MSOE) models assign independent stochastic disturbances to each structural component—level, trend, and seasonality—providing a richer and more flexible representation of uncertainty. Despite their appeal, MSOE models have long lacked a systematic and unifying taxonomy. To address this gap, I introduce the PTS framework, a structured classification of MSOE state space models tailored to their specific properties. The taxonomy organizes models along three dimensions: P (power transformations, based on Box–Cox), T (trend: none, local, global, or damped), and S (seasonality: none, discrete, or trigonometric), yielding up to 24 well-defined model variants. All models are formulated within a general state space system and estimated via maximum likelihood using the Kalman filter, with model selection based on standard information criteria. Finally, I will not resist the temptation to step slightly beyond the technical core of this talk and invite the audience into a broader conversation on the current state of the forecasting profession.