27 December, 2020 (Jung, bg, FLUXCOM carbon)
Gervasio Piñeiro, Susana Perelman, Juan P. Guerschman, José M. Paruelo, How to evaluate models: Observed vs. predicted or predicted vs. observed?, Ecological Modelling, Volume 216, Issues 3–4, 2008, Pages 316-322, ISSN 0304-3800, https://doi.org/10.1016/j.ecolmodel.2008.05.006.
Prediction (x-axis) vs. Observation (y-axis)
Martin added a comment to my proposal. He said that I should put prediction on the x-axis, referring to a paper:
I have just look over the conclusion of the paper. It says that:
Model evaluation based on linear regressions should be done placing the observed values in the y-axis and the predicted values in the x-axis (OP). Model evaluation based on the opposite regression leads to incorrect estimates of both the slope and the y-intercept. Underestimation of the slope and overestimation of the y-intercept increases as r2 values decrease.
We strongly recommend scientists to evaluate their models by regressing OP values and to test the significance of slope = 1 and intercept = 0.
RMSE should not be reported for the OP regression, but the RMSD adds important information to model evaluation.
Honestly, I don’t like the conclusion. For me, PO method is more intuitive as the larger (smaller) slope means overestimation (underestimation) from model’s point of view. But… I will try to get used to it.
Don’t report RMSE for OP regression!
On the other hand, the last sentence about RMSE is also interesting. I don’t know why for now, but I also don’t like to use RMSE as it is not intuitive and hard to interpret.