Deep learning for dynamic equilibrium models
Fernández-Villaverde et al. discuss in this paper how deep learning techniques, particularly deep neural networks (NNs), can help address challenges in quantitative economics, especially in solving high-dimensional dynamic equilibrium models (DEMs). Here are the key points:
- Neural networks and traditional methods:
– The authors argue that deep learning tools are generalizations of methods already familiar to economists.
– For example, linear regression is a special case of an NN, and NN-based functional approximations are similar to traditional projection methods. - Equilibrium conditions in economics:
– Unlike typical applications in other fields, economists must impose equilibrium conditions when using deep learning for DEMs.
– This creates unique challenges due to the feedback between agents’ optimizing behavior and equilibrium consistency requirements. - Solving dynamic equilibrium models (DEMs):
– DEMs are described as equilibrium Markov processes (EMPs) that satisfy both the agents’ Markov decision problems and consistency conditions.
– Finding this EMP numerically is what economists call “solving the model.” - Neural networks as function approximators:
– The paper explains how NNs can approximate unknown functions, such as value functions or policy functions in economic models.
– It describes the structure of basic and deep NNs, comparing them to traditional projection methods. - Application to stochastic neoclassical growth model:
– The authors demonstrate the application of deep NNs by solving the stochastic neoclassical growth model.
In conclusion, the paper argues that deep learning, particularly deep neural networks, offers a promising approach to addressing the curse of dimensionality in quantitative economics.
Taming the curse of dimensionality: quantitative economics with deep learning
Authors: Jesús Fernández-Villaverde, Galo Nuño, Jesse Perla
From: University of Pennsylvania, Banco de España, University of British Columbia
Deep learning methods for extracting structured data in economics
This working paper by Melissa Dell provides a comprehensive review of deep learning methods to extract structured information from unstructured text and image data at large scale. She covers several deep learning approaches such as classifiers and regression models, generative AI, and embedding models and discusses key architectures such as Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs),and Transformers. Dell emphasizes transfer learning – using pre-trained models and fine-tuning them on specific tasks with less data and provides guidance on when to use different approaches (e.g. classifiers vs embeddings) based on factors like number of classes and whether classes are pre-defined.
Main applications:
- Classification tasks:
– Text classification (e.g. news article topics)
– Named entity recognition
– Textual entailment - Embedding models for:
– Record linkage
– Tracking content dissemination
– Data exploration
– Optical character recognition - Regression and object detection for tasks like document layout analysis
Key findings:
- The author finds that custom-trained classifiers often outperform generative AI (like GPT) for nuanced classification tasks, but both perform well on straightforward tasks.
- Deep learning methods can scale affordably to problems with millions or billions of data points when suitable approaches are used.
- These techniques allow economists to extract novel, fine-grained data from unstructured sources at unprecedented scales.
The paper is accompanied by a website (EconDL) with demos, resources, and technical details to help economists implement these methods.
Deep Learning for Economists
Author: Melissa Dell
From: Harvard University
