MiDaS R-squared

[KrilleJ]

Hi.

I am running MiDaS estimations and cannot get R-square to be the same as when I specify ordinary regression models to match up the MiDaS regression. The estimated parameters are the same, and the actual, fitted and residual values are the same (albeit that the MiDaS list some additional values for the dependent variable when there are missing high frequency values). But the R-square is different. Possibly, it traces back to the mean of the dependent variable (see output below).

It seems as I am missing out on something central, but I cannot pinpoint it. Any help would be useful.

/Krille
--------------------------------------------------------------------------------------------------------------
Sample (adjusted): 1996Q4 2016Q2
Included observations: 79 after adjustments
DY=C(1)+C(2)*DY(-1)+C(3)*KIFI_MONTH_3+C(4)*KIFI_MONTH_2+C(5)
*KIFI_MONTH_1+C(6)*KIFI_MONTH_3(-1)

Coefficient Std. Error t-Statistic Prob.

C(1) -0.049805 0.010499 -4.743994 0.0000
C(2) -0.046275 0.111160 -0.416292 0.6784
C(3) 0.000802 0.000338 2.370477 0.0204
C(4) 0.000805 0.000469 1.717588 0.0901
C(5) -0.001182 0.000451 -2.623780 0.0106
C(6) 0.000136 0.000308 0.442649 0.6593

R-squared 0.487987 Mean dependent var 0.006400
Adjusted R-squared 0.452918 S.D. dependent var 0.009386
S.E. of regression 0.006942 Akaike info criterion -7.029475
Sum squared resid 0.003518 Schwarz criterion -6.849517
Log likelihood 283.6643 Hannan-Quinn criter. -6.957378
F-statistic 13.91490 Durbin-Watson stat 2.028770
Prob(F-statistic) 0.000000
--------------------------------------------------------------------------------------------------------------


--------------------------------------------------------------------------------------------------------------
Sample (adjusted): 1996Q4 2016Q2
Included observations: 79 after adjustments
Method: U-MIDAS

Variable Coefficient Std. Error t-Statistic Prob.

C -0.049805 0.010499 -4.743994 0.0000
DY(-1) -0.046275 0.111160 -0.416292 0.6784

Page: M Series: KIFI Lags: 4

LAG1 0.000802 0.000338 2.370477 0.0204
LAG2 0.000805 0.000469 1.717588 0.0901
LAG3 -0.001182 0.000451 -2.623780 0.0106
LAG4 0.000136 0.000308 0.442649 0.6593

R-squared 0.538963 Mean dependent var 0.006561
Adjusted R-squared 0.455138 S.D. dependent var 0.009157
S.E. of regression 0.006760 Akaike info criterion -7.029475
Sum squared resid 0.003518 Schwarz criterion -6.849517
Log likelihood 283.6643 Hannan-Quinn criter. -6.957378
--------------------------------------------------------------------------------------------------------------

[EViews Gareth]

Hard to say without seeing the workfile.

[KrilleJ]

Sorry, I attach the workfile.

The midas_step object estimates the MiDaS model, while OLS_midas is the "ordinary" regression counterpart.

/K

[EViews Gareth]

Thanks.

Slight bug in the MIDAS R-squared calculation - it wasn't accounting for dropped observations from the high-frequency page. It will be fixed in next patch.