Predictions poc jitter

[tiffanychu90]

• For metric 3, prediction inconsistency, let's just measure how inconsistent the prediction is. Take away any reference to actual_stop_arrival, except where it's used to set the 30 minute prediction window
  • For every minute, find the min / max of predictions. Prediction change = absolute value of max(prediction_pacific)-min(prediction_pacific)
  • For a rolling 2 min period, find the min / max of predictions again, and in that period, calculate the prediction_spread = max - min
  • For a given stop, prediction_inconsistency = mean(prediction_spread) across all the rolling windows
• Expand metric 1, update completeness, and do a version that's using vehicle positions
  • use actual_stop_arrivals for the window between the first stop's arrival and last stop's arrival
  • for every minute between that, count the number of vehicle position location_timestamp, and if it's more than 2, it's passes.
  • for prediction duration window, calculate the number of minutes with complete info (at least 2 updates)
  • this is calculated at trip-level, not stop-level
• Slack thread
• Research Request - RT Predictions / Real-Time Information POC metrics #709

[github-actions]

nbviewer URLs for impacted notebooks:

• rt_predictions/03_prediction_inconsistency.ipynb

[github-actions]

nbviewer URLs for impacted notebooks:

• rt_predictions/03_prediction_inconsistency.ipynb
• rt_predictions/06_update_completeness_vp.ipynb

[github-actions]

nbviewer URLs for impacted notebooks:

• rt_predictions/03_prediction_inconsistency.ipynb
• rt_predictions/06_update_completeness_vp.ipynb

[github-actions]

nbviewer URLs for impacted notebooks:

• rt_predictions/03_prediction_inconsistency.ipynb
• rt_predictions/06_update_completeness_vp.ipynb