Description
Hello sr.
I have the following problem. I used the fit.variogram function to fit several variogram models. The problem is that the one that fits best, is not the one that shows the best performance after analyzing the cross validation results. Is this possible? This is a problem since it makes no sense to me that the method that best fits my data set is the one that shows the worst results in predictive performance tests. Considering that the Gaussian model is the best fitting but worst performing model, should I give priority to the visual fit in choosing my model or to the performance shown in the cross-validation results? Here is my code and the graph with the results and models:
v <- variogram(Val_log~1, data_krig)
#Exp
expo <- fit.variogram(v, vgm("Exp"))
krig_expo<-krige.cv(Val_log~1,data_krig,expo)
#Sph
sph <- fit.variogram(v, vgm("Sph"))
krig_sph<-krige.cv(Val_log~1,data_krig,sph)
#Rest of the models..
rmse_expo <- sqrt(mean(krig_expo$residual^2))
mse_expo <- mean(krig_expo$residual^2)
var_expo <- var(krig_expo$residual)
rmse_sph <- sqrt(mean(krig_sph$residual^2))
mse_sph <- mean(krig_sph$residual^2)
var_sph <- var(krig_sph$residual)
#Rest of the statistics evaluation...
I would like to know if my code formulation is correct or why am I getting this? I should add that I am working with ERA5 - LAND surface pressure data.