Variable selection in MIDAS models: A comparison of Ridge, Elastic Net, LASSO, sparse-group LASSO, OCMT, and BMT
Translated title
Variabeludvælgelse i MIDAS-modeller: En sammenligning af Ridge, Elastic Net, LASSO, sparse-group LASSO, OCMT og BMT
Author
Jensen, Jonas Vammen
Term
4. term
Education
Publication year
2026
Abstract
This thesis compares regularization and variable selection techniques in high-dimensional Mixed Data Sampling (MIDAS) models for nowcasting U.S. real GDP growth. The study combines macroeconomic, financial, and text-based news indicators observed at different frequencies and estimates Ridge, LASSO, Elastic Net, sparse-group LASSO, One Covariate Multiple Testing (OCMT), and Boosting Multiple Testing (BMT) versions of MIDAS. Forecasts are generated in a rolling pseudo-out-of-sample exercise from 2002Q1 to 2017Q2 at three nowcasting horizons (two months, one month, and end-of-quarter), and accuracy is evaluated using RMSE, MSFE, MAE, and the aSPA test. Results indicate that Ridge-U-MIDAS and sparse-group LASSO-MIDAS deliver the strongest out-of-sample performance, with BMT-U-MIDAS and Elastic Net-U-MIDAS also competitive, whereas OCMT-U-MIDAS performs weakest despite often selecting many predictors. Overall, the evidence supports shrinkage and structured regularization as effective strategies for GDP nowcasting with high-dimensional mixed-frequency data.
Dette speciale sammenligner regulariserings- og variabelselektionsmetoder i højdimensionelle MIDAS-modeller til nowcasting af amerikansk real BNP-vækst. Analysen kombinerer makroøkonomiske, finansielle og tekstbaserede nyhedsindikatorer observeret ved forskellige frekvenser og estimerer Ridge-, LASSO-, Elastic Net-, sparse-group LASSO-, One Covariate Multiple Testing (OCMT)- og Boosting Multiple Testing (BMT)-varianter af MIDAS. Prognoser fremstilles i et rullende pseudo-out-of-sample setup fra 2002K1 til 2017K2 ved tre horisonter (to måneder, én måned og kvartalsslut), og nøjagtigheden vurderes med RMSE, MSFE, MAE og aSPA-testen. Resultaterne viser, at Ridge-U-MIDAS og sparse-group LASSO-MIDAS giver den bedste out-of-sample præstation, mens BMT-U-MIDAS og Elastic Net-U-MIDAS også er konkurrencedygtige; omvendt klarer OCMT-U-MIDAS sig svagest trods ofte at vælge mange prædiktorer. Samlet peger fundene på, at shrinkage og struktureret regularisering er effektive til BNP-nowcasting med højdimensionelle mixed-frequency data.
[This abstract has been generated with the help of AI directly from the project full text]