Methods for fast rolling and expanding linear regression models. That is, series of linear regression models estimated on either an expanding window of data or a moving window of data. The methods use rank-one updates and downdates of the upper triangular matrix from a QR decomposition (see Dongarra, Moler, Bunch, and Stewart (1979)
grp value will have
at least width - 1L distance from min(grp) in roll_regres and
roll_regres.fit