Clustering of Time Series Using Machine Learning

We investigate how human cooperation evolves over two different dimensions of time. Subjects in a laboratory experiment play a high-frequency two-player Cournot-Tullock market game over hundreds of timed periods with and without information about the payoff function. By varying the length of periods across treatments, we generate a data set that is stable in "physical time" (seconds) but varies in "period time". We apply machine learnings, specifically the dynamic time warping algorithm to cluster time series of market profits that exhibit similar behavior across the treatment dimensions and find that market behavior in such high-frequency games is determined mainly by physical time.