Not able to reproduce LEAR forecasts stored in epftoolbox/forecasts/

[aleszani]

Dear @jeslago, as a first step before using your package for research purposes I tried to reproduce the BE forecasts here obtained with LEAR. To do so I've kept it as simple as possible by using the evaluate_lear_in_test_dataset function, with parameters:

• dataset='BE',
• years_test=2,
• calibration_window=84,

Unfortunately the price forecasts I obtain are similar to yours but not the same. I have also tried to change calibration window to 56 and 1092 but the problem still persists. Am I missing something on the methodology you use for obtaining LEAR forecasts stored inside this file Forecasts_BE_DNN_LEAR_ensembles.csv?

Many thanks in advance for your help! Alex

[jeslago]

Can you share the numbers you get? Both in terms of MAE as well as the forecasts themselves? We detected some issues with replication due to the underlying libraries leading to small numerical differences. We have noticed that when using newer scikit-learn versions the numerical output changes slightly.

@gmarcjasz Can you pitch in here?

[aleszani]

Follow MAE values and forecasts for all datasets that have been calculated in the following conditions: years_test = 2 and calibration_window=84. The version of scikit-learn I am using is 1.01 following issue #10 .

Here are MAE values I obtain for all datasets:

While here are the forecasts:

• PJM
• NP
• BE
• FR
• DE