CRAN_Status_Badge Downloads


midasML - estimation and prediction for high-dimensional mixed frequency time series data.


The midasML package implements estimation and prediction methods for high dimensional time series regression models under mixed data sampling data structures using structured-sparsity penalties and orthogonal polynomials. For more information on the midasML approach see [1]. The package also allows to estimate and predict using single-variate MIDAS regressions. Note that such regressions are also implemented in midasr package. Functions implemented in this package allows to directly compare low-dimensional and high-dimensional MIDAS regression models.

The core of the midasML method is the sparse-group LASSO (sg-LASSO) estimator proposed by [2], and studied for high-dimensional time series data by [1, 3]. The sg-LASSO consists of group structures that are present in high-dimensional ARDL-MIDAS model, hence it is a natural estimator for such model.

The main algorithm for solving sg-LASSO estimator is taken from [2].

Functions that compute MIDAS data structures were inspired by MIDAS Matlab toolbox (v2.3) written by Eric Ghysels, see [4].

Main functions

Estimation and prediction functions


[1] Babii, A., Ghysels, E., & Striaukas, J. (2020). Machine learning time series regressions with an application to nowcasting.

[2] Simon, N., Friedman, J., Hastie, T., & Tibshirani, R. (2013). A sparse-group lasso. Journal of computational and graphical statistics, 22(2), 231-245. Related CRAN R package.

[3] Babii, A., Ghysels, E., & Striaukas, J. (2020). Inference for high-dimensional regressions with heteroskedasticity and autocorrelation.

[4] Ghysels, E. et. al. Mathworks Matlab toolbox.