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].

ARDL-MIDAS model with Beta density and exponential Almon specifications use analytical gradients derived in [5]. Also, MIDAS Logit model is available for linear in parameters specifications (soon: exponential Almon with analytical gradients)

• midasml_forecast - midasML estimation and prediction function.
• midas_ardl - ARDL-MIDAS single-variate estimation and prediction function (accomodates different weight functions and loss functions, e.g. quantile regression loss).
• midas_dl - DL-MIDAS single-variate estimation and prediction function (accomodates different weight functions and loss functions, e.g. quantile regression loss).

• reg_sgl - sg-LASSO regression estimation (currently only for mse loss).
• panel_sgl - panel sg-LASSO regression estimation (currently only for mse loss).

• qtarget.sort_midasml - transforms data into format suitable for midasML technique, creating in-sample and out-of-sample observations for quarterly target variable. Output could be directly inputed into midasml_forecast (note: currently does not handle real-time data vintages. in case real-time experiment is considered for a specific application, this function can help to setup up the data for each quarter prediction separately. future updates will contain functions capable of handling real-time data vintages.)
• mixed_freq_data transforms data into MIDAS regression format creating in-sample and out-of-sample observations. Output is subsequenlty used in midas_ardl & midas_dl

# CRAN version
install.packages("midasml")

# Development version
# install.packages("devtools")
library(devtools)
install_github("jstriaukas/midasml")

[1] Babii, A., Ghysels, E., & Striaukas, J. (2020). Machine learning time series regressions with an application to nowcasting. https://arxiv.org/abs/2005.14057

[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. https://CRAN.R-project.org/package=SGL

[3] Babii, A., Ghysels, E., & Striaukas, J. (2020). Inference for high-dimensional regressions with heteroskedasticity and autocorrelation. https://arxiv.org/abs/1912.06307.

[4] Ghysels, E. et. al. Mathworks Matlab toolbox. https://www.mathworks.com/matlabcentral/fileexchange/45150-midas-matlab-toolbox

[5] Kostrov (2020) Estimating MIDAS regressions via MIDAS-NLS with revised optimization. Working paper.