ENET Elastic net regularization: EViews 13 vs 14

For technical questions regarding estimation of single equations, systems, VARs, Factor analysis and State Space Models in EViews. General econometric questions and advice should go in the Econometric Discussions forum.

Moderators: EViews Gareth, EViews Moderator

MCD_nz
Posts: 8
Joined: Fri May 17, 2019 2:45 pm

ENET Elastic net regularization: EViews 13 vs 14

Postby MCD_nz » Thu Jan 30, 2025 8:19 pm

I get different results when comparing the results of Ridge regressions and Lasso regressions in EViews 13 vs EViews 14 when lambda is left blank. The cross-validation options are K-fold (with 5 folds), the random number generator is Knuth (with seed set to 0), and the objective is set to Mean square Error. (Path definition and stopping rules are left at their defaults), and dependent scaling is set to "none". For the ridge regression, the optimal lambda is precisely 1000 times higher.

Would you happen to have any insights on why I cannot reproduce results across different versions of EVIEWS?

Output from EVIEWS 14
Dependent Variable: DLOG(RGDP)
Method: Elastic Net Regularization
Date: 01/31/25 Time: 16:11
Sample (adjusted): 2000Q2 2023Q1
Included observations: 92 after adjustments
Regularization penalty: Ridge regression (squared L2-norm)
Lambda specification: Automatic path (optimal = 17.822)
Cross-validation: K-Fold with MSE Objective (nfolds=5, rng=kn,
seed=0)
Variable scaling: Regressors (none), Dependent (none)

EViews Glenn
EViews Developer
Posts: 2679
Joined: Wed Oct 15, 2008 9:17 am

Re: ENET Elastic net regularization: EViews 13 vs 14

Postby EViews Glenn » Fri Jan 31, 2025 10:09 am

Hi. Thanks for your question.

The engine for elastic net estimation was completely revamped between the two versions. In particular, we moved to a coordinate descent algorithm with warm start path estimation which offers considerable computational efficiencies and more importantly, numeric stability. This change in method is the reason that we require older elastic net specifications to be reestimated in EViews 14, as we want people to use the updated estimates.

The differences in results should be small for most individual lambda specifications, but small differences are magnified where model selection is involved, and can lead to very different selected models. Indeed, I suspect that improvements in the model selection algorithm may prove to be more determinative of the differences in results that you are seeing, than are changes in the estimation algorithm.

I will note that we are generally very reluctant to make changes in computation when they may lead to differences in results between versions. In this case, the benefits to changing algorithms were so compelling, both for estimation of existing models, and most importantly, for planned future development, that we made the change. I hope that you will find that the added functionality we provide in the EViews 14 elastic net makes this worth it, but rest assured that we did not make these kinds of changes lightly, and we apologize for any inconvenience.

If you would like to provide your workfile, along with the estimates of interest, we may be able to offer additional commentary on the differences for your particular data and specification. It is certainly possible that there are settings which better align results. I do note that you have very few observations which will magnify differences between the results obtained by the two versions.


Return to “Estimation”

Who is online

Users browsing this forum: No registered users and 2 guests