Open mancag opened 7 years ago
In the graphical interface, the forecasts that appear in the stochastic series panel are the forecasts computed in Tramo (=forecasts of the series - regression effects) (see also the y_lin_f file). The values of decomposition.y_lin_f correspond to the forecasts computed in Seats (=forecasts of the linearized series). Those forecasts might differ for several reasons: actual definition of those forecasts (see above) or model changed by Seats... We decided to use in the graphical interface the forecasts of Tramo to avoid confusion for the users (even if the use of the forecasts of Seats would be more coherent). Comparing the different forecasts may be instructive, especially when the model is changed in Seats (non decomposable models).
Some additional precisions/corrections: Except when Seats changes the model, the linearized forecasts of Tramo and of Seats should be identical: (forecasts of the series) - regression effects = forecasts of (series-regression effects). However, for the moment, we use two different algorithms to compute the forecasts: for comparability issues, we use in the regarima part the same algorithm as in X13; that algorithm produces forecasts conditional to the first (D+BD+P+BP) observations. In Seats, we use the exact solution, conditional to the first (D+BD) observations. When the model doesn't contain stationary auto-regressive polynomials, the forecasts are identical. In the future (JD+ 3.0), we should use the exact solution; that will solve the differences between Seats and Tramo forecasts (except in the cases mentioned above), but it implies some (usually small) differences in comparison with the "official" X13 implementation.
Should the values for decomposition.y_lin_f in the output file be equal to italic values of »Decomposition → Stochastic series → Series (cmp)«? In JD+ 2.2.0 RC1, they are not equal.