Online Program

Probabilistic Methods in Time Series Analysis: A Case Study (308504)

Keywords: Bayesian, Time Series, Probabilistic

Modeling real-world time series data poses unique challenges. Properties such as infrequent reporting, data inconsistencies, and nonlinear behavior may prove difficult when employing traditional modeling techniques. A probabilistic approach helps circumvent these issues by natively handling missing, noisy data while providing a result with measurable uncertainty. Methods were applied to analyze the performance of a defined stage within a complex, multistage queuing process. Performance is measured by the reported number of days spent in each stage, a metric defined as a timeliness. To uncover seasonal patterns and provide a forecast of timeliness, Structural Time Series was used. Detecting structural changes in timeliness behavior was accomplished using a Hidden Markov Model. Lastly, Gaussian Process Regression in conjunction with Least Squares Anomaly Detection was used to characterize the overall process and identify systemic anomalies. Understanding the behavior of timeliness in an identified stage can help inform decision makers to streamline operations, anticipate backlog, and prevent bottlenecks in workflow.