A Neural Network Based Factor Model for Cryptocurrencies

An empirical application of the Conditional Autoencoder to the cryptocurrency market. The Conditional Autoencoder is a deep learning based nonlinear extension of Instrumental Principal Component Analysis.

0 Rating
0 Reviews
5 Students Enrolled

Wishlist

Free

Enroll Now

Tags:

Courselet Content

2 components

	A Neural Network Based Factor Model for Cryptocurrencies (video)	35 min
	A Neural Network Based Factor Model for Cryptocurrencies (pdf)	17.29 M

Requirements

Basic knowledge of statistics, cryptocurrencies and neural networks.

General Overview

Description

We apply a neural network based factor model to the cryptocurrency market to describe individual asset returns in terms of latent risk factors and time-varying risk exposures. Our study makes five contributions to the literature. First, we show that pricing performance is improved by adding nonlinearities to the risk exposures. Second, we establish that the risk dynamics of the cryptocurrency market evolve more quickly than for equity. Third, we identify that cryptocurrency prices were more predictable before the COVID-19 pandemic than thereafter. Fourth, we find latent risk factors to be related to observables, but to additionally include idiosyncratic variance. Last, we observe that asset characteristics which are important for the estimation of risk exposures are commonly found in the literature on observable factor models.