Everthing I've said is about data from a single voxel.
The coefficient of 1(t) from 3dDeconvolve can be considered to be the 'baseline' -- it is the constant 'a' in the time series model
v(t) = a + b*r_1(t) + c*r_2(t) + .....
where the other model functions r_i(t) are prescribed by the -stim_file (etc.) input options. Some of the r_i(t) are in the 'signal model' and some in the 'baseline model' (the latter r_i(t) are from -polort or -stim_base).
However, the other coefficients in the baseline model in 3dDeconvolve do represent something -- they are the magnitudes of the temporal fluctuation of the baseline (e.g., the linear drift over time). Usually, people do not consider these when computing percent change.