Both are correct. It's just simple math: in data series, the time scale is reversed, i.e. higher indexes go backwards in time. This reverses the sign of first coefficient, of the third, fifth, and so on. LinearReg does not generate coefficients for a data series polynom, polyfit does.