Speaker: Dr. Mike West
The Arts & Sciences Professor of Statistics & Decision Sciences
Bayesian Modelling and Forecasting of High Dimensional Count-Valued Time Series
Abstract: Problems of modelling for forecasting many, related time series of non-negative counts arise in a range of areas. They are particularly prevalent in commercial contexts such as consumer demand/sales prediction, transactions monitoring, network flow assessment, and modern IT applications. New classes of dynamic models address efficacy and scalability of Bayesian analysis for sequential learning and forecasting in such contexts, building on the decouple/recouple concept for scaling Bayesian analysis.
A general class of dynamic generalized linear models for binary and conditionally Poisson time series, with dynamic random effects for over-dispersion, allows use of dynamic covariates in both binary and non-zero count components. Sequential Bayesian learning is fast and parallel on the set of decoupled time series. New multiscale models enable information sharing when data at an aggregate level (or from any other relevant source or model) can be expected to provide useful information on shared patterns such as trends and seasonality. This multi-scale approach– one example of the concept of decouple/recouple– avoids the complexity and computational challenges of traditional hierarchical modelling. Analysis incorporates cross-series linkages while insulating parallel estimation of univariate models, hence enables scaling linearly in the number of series. Customized extensions address problems of consumer sales forecasting, focused on individual customer transactions, coupled with a novel probabilistic model for predicting counts of items per transaction. The latter involves a new dynamic binary cascade concept that contributes two main features: first, it aids in resolving some of the otherwise unpredictable variation in sales series; second, it allows probabilistic inference on rare events (otherwise unpredictable and very infrequent “high” outcomes on individual series).
A motivating case study context of many-item, multi-period, multi-step ahead supermarket sales forecasting provides examples that demonstrate improved forecast accuracy in a range of traditional and statistical metrics, while also illustrating the benefits of full probabilistic models. The talk will discuss a range of examples and highlight questions of forecast accuracy metrics and broader questions of probabilistic forecast accuracy assessment, decision analytic choice of point forecasts (when desired) and forecast accuracy comparison.
These models are, more recently, being deployed and extended in applications in other areas, including dynamic network monitoring contexts in cybers ecuritry studies as well as increasingly large-scale commercial applications.
- L. R. Berry and M. West (2019), Bayesian forecasting of many count-valued time series, Journal of Business and Economic Statistics
- L. R. Berry, P. Helman, and M. West (2019), Probabilistic forecasting of heterogeneous consumer transaction sales time series, International Journal of Forecasting
Mike West holds a Duke University distinguished chair as the Arts & Sciences Professor of Statistics & Decision Sciences in the Department of Statistical Science, where he led the development of statistics from 1990-2002. A past president of the International Society for Bayesian Analysis (ISBA), Mike has served the international statistics profession in founding roles for ISBA and other in other professional organizations and institutions.
Mike’s research and teaching activities are in Bayesian analysis in ranges of interlinked areas: theory and methods of dynamic models in time series analysis, multivariate analysis, latent structure, high-dimensional inference and computation, quantitative and computational decision analysis, stochastic computational methods, and statistical computing, among other topics. Interdisciplinary R&D has ranged across applications in signal processing, finance, econometrics, climatology, systems biology, genomics and neuroscience, among other areas. Main current interests are in macroeconomic forecasting and policy decisions, financial econometric forecasting and decisions, dynamic network studies in IT/commerce, and large-scale forecasting and decision problems in business and industry.
Mike has received a number of international awards for research and professional service, and multiple distinguished speaking awards. He has been a statistical consultant for various companies, banks, government agencies and academic centers, co-founder of a biotech company, and past or current advisor or board member for several financial and IT companies. Mike teaches in academia and through short-courses, works with and advises many undergraduates and Master’s students, and has mentored over 60 primary PhD students and postdoctoral associates, most of whom are now in academic, industrial or governmental positions involving advanced statistical research.