the differences between the previous year and the current year consistently are far apart, but that could probably make them more accurate. I know these models can be complicated, but I wouldn't be surprised if such a thing can be done, but I do know that tweaking weights and parameters in models is quite tricky, especially trying to do so dynamically.