Time Travel Validation for Time Series Data

Author Information

Thomas Tiahrt, Univeristy of South DakotaFollow

Start Date

17-10-2025 4:30 PM

End Date

17-10-2025 5:00 PM

Location

MH 162

Presenter Information

Thomas Tiahrt is the POET Professor of Business Analytics and an Associate Professor of Decision Sciences at the University of South Dakota. He received his B.S. in Business Administration, his M.A. in Computer Science and his Ph.D. in Computational Science and Statistics from USD. Dr. Tiahrt spent over two decades in software development ranging from writing large-scale software systems to supervising staff. His work resulted in five software patents embodying innovation in indexing, categorization and dynamic database content capture. Those patents have been cited by Google, IBM, Trend Micro, and the Regents of the University of California as prior art for their patents. His research spans everything from physical therapy to predicting no-show patients.

Submission Type

Abstract

Track

Analytics and Statistics

Abstract

Model validation is necessary to evaluate the predictive effectiveness of time series forecasts. The standard validation method is to start at the earliest point in the data, train on two of the minimum meaningful number of periods or cycles of periods (i.e., two or more blocks) then test on the next block. Subsequent models are tested by extending the training period one block at a time and testing on the following block. Alternatively, the last block can be substituted for the following block. Other forward progressions can be used including forms of cross validation. In any case, the forecaster chooses the model with the best metric. The time travel validation method reverses the sequence. It begins by testing the model on the last or ultimate block using the antepenultimate and penultimate blocks. It extends backward one block at a time until it reaches the first block. The forecaster selects the best performing time travel model. Results show that time travel validation produces better than the standard method for many time series datasets.