Project CEMAPRE internal
|Title||For an Efficient Quantification of Model Discrepancy in Non-linear Highly Parametrised State-Space Models using Deep Learning Techniques|
|Participants||Rui Paulo (Principal Investigator), Ivo Tavares|
|Summary||Uncertainty quantification (UQ) aims at characterizing and reducing the sources of uncertainty|
involved when modelling real world processes. It has recently contributed to the area of
Macroeconomics. Even the most complex model may be lacking in some modelling aspect, leading to the
presence of a particular source of uncertainty, namely, model discrepancy or structural uncertainty.
In our previous research, we used a bias correcting term with a Gaussian Process (GP) prior, and we
tried to quantify and correct the misspecification for a linear State-Space model (SSM), the type of
model in which many DSGE models are formulated after being solved. However, despite the interesting
capabilities of this technique, some shortcomings were encountered, namely the heavy computational
burden. Deep Learning methods have recently been used to learn such models. The purpose of this
research project is to investigate the usefulness of such class of techniques in reducing the
computational barriers previously encountered.