Stepwise VIF elimination

Analysis Master 21-07-2026

Removing weak terms one at a time from Model 6, re-derived at calibrated alpha. At each step, find the extra term (not one of the core terms) with the highest VIF, remove it, and fit the model again. Stop when every extra term has a VIF of 10 or less. The 13 core terms from Model 4a are never removed (Multicollinearity is structural).

Step	Worst droppable	VIF	Action
0	σ²	18.38	drop
1	f_distribution × share_type	9.82	stop

Removing σ² costs nothing: the fit does not get significantly worse (LR = 1.18, df = 1, p = 0.278). The condition number drops about 6 times (14,851 → 2,430), so the model becomes much more stable.

Result, the 14-parameter model:

\[V_t = w_{1,t} f_\text{distribution} + w_{2,t} f_\text{distance} + w_{3,t} f_\text{directional} + w_{4,t}\,share_t + w_5\,\sigma + w_6\,(f_\text{distribution}\times share_\text{type})\]

This is Model 4a plus f_distribution × share_type. Accuracy 0.724, accuracy on unseen data 0.709 ± 0.075, condition number 2,430. It has the same form as the model reported on 19-07-2026, now built again from the start.

Later it was made smaller, to 12 parameters: f_directional: shared vs split. Part of Calibration.

Linked from Calibration · Model 6, re-derived at calibrated alpha · Model progression · Multicollinearity is structural · RQ1 rebuilt with grouped CV · Sigma in tree importance