# Geometric Mean Bias result for air quality models

i'm calculating the Geometric Mean Bias (MG) for the results of an air quality dispersion model against observations with the following equation from Chang and Hanna (2004);

$$MG = exp(\overline{LnC_o} - \overline{LnC_p})$$

A result of 1 indicates a perfect match of data means. My value for 32 observation & prediction pairs comes out at 1.2, fine. But my reading of various papers and this site suggests that values > 1 indicate the model is overestimating.

The thing is, my model is clearly underestimating, all but 4 predictions are below the matching observation. If to have MG > 1 you need $$(\overline{LnC_o} - \overline{LnC_p})$$ to be > 0 (for the exponential), then surely a model must be underestimating if MG > 1 ? i.e. $$\overline{LnC_o} > \overline{LnC_p}$$ ? I must be missing something basic.