How to choose the shift parameter to reproduce the result describes in the paper

[kunruiZhu]

Hi George,

Thanks for the great repository.

I am trying to reproduce the experiments results in the paper and run the experiments.py following the Run section in the REAME file.

However, shown in the figure below, when comparing the scores between the rule-based predictors(AVG/ AVG_WINDOW) whose output should be stable across different platforms, I notice that the average errors(AE) are closer to shift=1,3,6 instead of shift=2,6,13(equals to dt=3,7 and 14) reported in the paper.

For example, shown in the figure below, the score for AVG predictor in England with the setting of Up to next 3 Days shares the same AE 9.75 between the paper and the output from the code with shift=1 (highlighted with yellow), and for the AVG_WINDOW predictor in Italy, with the setting of Up to next 14 Days shares similar AE 19.45 and 19.24 in the paper the output from shift=6 respectively. On the contrary, in the paper, the MAE of AVG_WINDOW predictor in England with the Up to next 14 Days setting is 8.54(highlighted with orange) while the output from the code is 10.47 (highlighted with green box), which is quite different.

May i know the correct shift parameter to reproduce the average error reported in the paper?

Thanks, Konroy

[geopanag]

Hello Konroy thanks a lot for your detailed comment. Correct me If I m wrong, but I think your results describe the error at the shift (as computed by the code). The errors in the paper correspond to the average of these shifts e.g AVG_WINDOW up to 3 days in England means to take the average error for shift =0,1,2 i.e. ( 6.01+ 6.51+7.02)/3 = 6.513.

[kunruiZhu]

Got it, thanks for your explanation ;)