[R-sig-eco] PLSR: negative R2

[Thomas Möckel]

Dear list members,
I try to predict indicator values by means of spectral reflectance. I 
use the plsr function of the pls package with a leave-one-out 
validation. For one model I receive negative R2_CV values. I know that 
they mean that the model is bad in model my values. But when I test the 
model on a separate test dataset I get positive R2 values and the RMSEP 
is also not absolutely bad.
My question is:
How can it be that the R2_CV is bad but the R2_test is OK?
Is it a valid procedure to test a model with negative R2_CV values on a 
test dataset or should I rather stop after the LOO validation and say 
that it was not possible to build a reliable model for this variable?

Thank you very much for your answers!

-- 
*^*^*^*^*^*^*^*^*^*^*^*^*^*^*^*^*^*^*^*^*^*^*^*
Thomas Möckel, PhD student

Department of Earth and Ecosystem Science
Lund University
Sölvegatan 12
SE-223 62 Lund
Sweden

Tel:+46 (0)46 - 222 4887
E-mail: Thomas.Mockel at nateko.lu.se