Interpretable analysis of multiple ordinal time series data

In this thesis, we examine two questions posed by Lazar et al. in their paper in which they develop a recursive forecasting model for multiple ordinal time series accelerometer data. The first question is to what extent does having many of the same subjects present in the training and test data, but with different recorded accelerometer data, have on the reliability of any error rate estimates produced by the model. To address this, we discuss a method for simulating multiple ordinal time series data and simulate a training set and two different test sets to test our model, one which simulates a test set with identical subjects to the training set and one which simulates a test set with different subjects, and compare the estimates of the error rate produced from these two test sets and the real test set. We found a considerable improvement in test error for the former of the two simulated test sets, suggesting a severe limitation in the ability to rely on the results produced by such data. The second question is how the parameter ϵ used in the model weighting process should be tuned to produce the lowest error rates. To address this, we fit our model repeatedly using various values of ϵ and compare the results. On the real data, we found that increasing the value of the parameter ϵ corresponded to small improvements in the error rate, though the same conclusion cannot be drawn from the simulated data.