Rolling Regression with Time Blocks

One idea is to model the alpha and beta as some gaussian process with a Periodic covariance matrix, or put some smoothing team on the random walk (which would be roughly equal to a gaussian process).