There are two common ways to compute the baseline for the purpose of being the denominator for the percent change calculation:
(1) Average of the entire data time series.
(2) Average of the 'non-stimulus' time points values.
To me, (2) make more sense. If you are using 3dDeconvolve, this is equivalent to the value that will be computed as the coefficient of the 1(t) baseline model function, since all the other baseline model functions have their mean removed before being inserted into the design matrix.
If you are using the AFNI 'FIM' functions or programs, then they use choice (1) by default.
ZtMM