romainlafarguette / gar

Conditional Density Projection via Quantile Regressions, Resampling and Multifit Models

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Conditional Density Projection via Quantile Regressions, Resampling and Multifit Models

The developer version of the Growth at Risk model used at the IMF -- still beta

The official, IMF-approved version is available in https://github.com/IMFGAR/GaR

This version contains the new functionalities I am developing, without being reviewed by colleagues. Use them at your own risk !!

https://romainlafarguette.github.io/software/

The project is split along different steps, that have to be ran sequentially:

  • step 001: Group variables into partitions, to reduce parametric noise and provides more degrees of freedoms. The partitions are estimated using either Principal Component Analysis (PCA) or Partial Least Squares (PLS - also called projections on latent structures)

  • step 002: estimate the quantile regressions, project GDP growth at different horizons and generate term structure and fan chart plots. Note that the fan charts rely on quantile rearrangement

  • step 003: fit the sampled density using kernel and parametric densities. Best parametric family is assessed using AIC, BIC or RSS criteria.

  • step 004: fit the density using Gaussian mixtures

  • step 005: measure the performance of the density forecasts using PIT, logscores and entropy tests

  • step 006: try different quantiles interpolation methods. This script is not very important for a "standard" GaR use, it was rather to test the robustness of the rearrangement approach in quantiles uncrossing and sampling.

About

Conditional Density Projection via Quantile Regressions, Resampling and Multifit Models

License:GNU General Public License v3.0


Languages

Language:Python 100.0%