Currently, the Lo-Fi code in data reduction proxy records the accuracy of network quality prediction on every main frame request. For example, when a main frame HTTP request is seen (say at t=0), network quality prediction is stored in a variable.  When the next main frame HTTP request is seen (say at t=t1), the observed network quality during the time period from t=0 to t=t1 is computed, and compared with the stored prediction.

This has 2 problems: (i) User may not load the next web page (or may load a HTTPS web page), and thus the accuracy metric may not be recorded (ii) The interval between t=0 and t=t1 may be too large, and thus comparing the prediction with the observations may be meaningless.

Instead, we should record the accuracy of prediction at t=15, 30 and 60 seconds by posting tasks at t=0.