This is an mini-course on "Advanced Methods in Computational Economics" to solve and estimate dynamic models, held from 14.11.2022 - 21.11.2022 at CREST/ENSAE.
- The lecture suite will take place in person on Monday, 14.11.2022 and on Thursday, 17.11.2022.
- The third lecture on the 21.11.2022 will be held via Zoom
- Zoom link for 21.11.2022, 13.00 - 16.00
- Meeting ID: 705 635 3422
Class enrollment on the Nuvolos Cloud
- All lecture materials (slides, codes, and further readings) will be distributed via the Nuvolos Cloud.
- To enroll in this class, please click on this enrollment key, and follow the steps.
-
This course is intended to confront M.sc. and Ph.D. students in economics, finance, and related fields with recent tools developed in applied mathematics, machine learning, computational science, and the computational economics literature to solve and estimate (dynamic stochastic) economic models.
-
The lectures will be centered around sparse grids, adaptive sparse grids, high-dimensional model representation, two types of machine learning methods (Gaussian Processes and Deep Neural Networks), and will be showcased in the context of application in macroeconomics, finance, and climate-change economics.
-
The lectures will be interactive, in a workshop-like style, that is, a mix of theory and actively playing with code examples (delivered in Python and deployed on a cloud computing infrastructure).
- Basic econometrics
- Basic programming in Python (see this link for a thorough Introduction)
- Basic calculus and probability
- Mathematics for Machine learning provides a good overview of skills participants are required to be fluent in.
Day 1, Monday, November 14th, 2022 (13.30 - 16.45; Room 2016)
- Introduction to Sparse Grids
- Introduction to Adaptive Sparse Grids
- Introduction to High-dimensional Model Representation
Day 2, Thursday, November 17th, 2022 (13.30 - 16.45; Room 2016)
- Deep learning basics
- Multi-layer perceptron
- Feed-forward networks
- Network training - SGD
- Error back-propagation
- Some notes on overfitting
- Throughout lectures - hands-on: Perceptron, gradient descent, Artificial neural networks: a simple MLP implementation and examples of applications
- Introduction to Tensorflow, applied to supervised machine learing
- Deep Surrogates/Deep Structural Estimation
- Deep Equilibrium Nets
Day 3, Monday, November 21st, 2022 (13.30 - 16.45; Online)
- Basics of Gaussian Process Regression
- Noise-free kernels
- Kernels with noise
- GP classification
- The curse of dimensionality and how to deal with it (e.g., active subspaces)
- Gaussian mixture models (unsupervised machine learning)
- Bayesian active learning
- Dynamic Programming/optimal control with GPs
- An outlook to frontier topics of GPs (Limitations of GPs and "big data"/scalable GPs
- A Gaussian Process Dynamic Programming Code Library
Lectures will be interactive, in a workshop-like style, using Python, scikit learn, Tensorflow, and TFP on Nuvolos, a browser-based cloud infrastructure in which files, datasets, code and applications work together, in order to directly implement and experiment with the introduced methods and algorithms.
- Simon Scheidegger (HEC, University of Lausanne)
- Simon Scheidegger: simon.scheidegger@unil.ch
- Nuvolos Support: support@nuvolos.cloud
Participants who take the course for credits are expected to propose a small project where they apply some of the methods learned to an application (e.g., apply deep learning to a data set, solve a dynamic model, etc.). The deliverable that will be graded is a short write-up (4-6 pages maximum), the data set (if any), and the code on which the presented results in the report were based. The participants will have four weeks time after the course is finished to complete the task. The participants can work alone, or in teams of two.