Recording Prediction Error for Travel and Dwell Times Seperately

[simonberrebi]

Measuring prediction error separately would allow to benchmark how travel and dwell times contribute to prediction errors. This benchmarking could inform the prioritization of new development efforts. The benchmarking could also assess the performance of prediction methods and help identify bugs. Recording prediction errors for constituent parts separately is therefore necessary to guide the development of prediction methods addressing dwell time and travel time specifically.

[scrudden]

Code on branch tc_issue_225

[simonberrebi]

Based on the new data fields developed by @scrudden, we are now able to record prediction errors for travel and dwell times separately. The following graph shows 12.5 to 87.7 percentiles of dwell and travel prediction errors against the horizon on the MBTA's Route 66 in Boston. We can see that dwell time predictions tend to be overly optimistic whereas travel time predictions are pessimistic. They cancel each other out. This shows the value of considering travel and dwell time predictions separately.