Why is my R-squared value different?

I have used a feedforward backprop to train a network.

When I extract the output from the network and calculate the R-squared value with MS Excel, it is different from what MATLAB calculates from the network when the plotregression command is activated. Any reasons for this? Which one do I report as the actual value?

[John D'Errico]

Does it matter which one you report? There is more than one possible method to compute R^2. Worse, there are ways to compute an adjusted R^2, which is sometimes arguably more appropriate. So you might be just seeing an adjusted R^2.

Or it might be simply that you extracted a set of coefficients without taking ALL of the significant digits of those parameters, putting essentially the wrong results into Excel. This is a common mistake in fact.

Don't get hung up on a single number that tries to predict goodness of fit. R^2 simply is not that important.