Machine Learning and the Cross-Section of Cryptocurrency Returns

Aug 3, 2026

The study applies a diverse set of machine learning models to test cross-sectional return predictability in cryptocurrency markets. Despite consistently strong economic gains relative to other asset classes, the incremental value of model complexity is limited. Predictability concentrates in a small set of straightforward characteristics, including market price, past alpha, illiquidity, and momentum. Abnormal returns are traced to the long leg of trades and show persistence over time. Even with high turnover, many strategies remain profitable after transaction costs, with alphas concentrated in hard-to-trade assets driven by extreme returns on small, illiquid, and volatile coins.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages68

Reading Time2 h 51 min

AccessFree to download

Introduction

Data and scope

Machine learning approach

Prediction results and interpretation

Machine Learning and the CrossSection of Cryptocurrency Returns* Nusret Cakici †, Syed Jawad Hussain Shahzad‡, Barbara Będowska-Sójka §, Adam Zaremba 􀟜 ⁑ Abstract We employ a repertoire of machine learning models to investigate the cross-sectional return predictability in cryptocurrency markets. While all methods generate substantial economic gains—unlike in other asset classes—the beneﬁts from model complexity are limited. Return predictability derives mainly from a handful of simple characteristics, such as market price, past alpha, illiquidity, and momentum. Contrary to the stock market, abnormal returns in cryptocurrencies originate from the long leg of the trade and persist over time. Furthermore, despite high portfolio turnover, most machine learning strategies remain proﬁtable after trading costs. However, alphas are concentrated in hardto-trade assets and critically depend on harvesting extreme returns on small, illiquid, and volatile coins. Keywords: cryptocurrency markets, machine learning, return predictability, limits to arbitrage, asset pricing, the cross-section of returns JEL codes: G11, G12, G17 This version: 18 February 2024 * We thank the participants of the 39th Conference of the French Finance Association, the 2nd Spring Workshop on Fintech, the 43rd EBES Conference, as well as the seminars at John von Neumann University and Prince Sultan University for helpful comments and suggestions. All errors are our own. This research was funded in whole or in part by the National Science Center of Poland [Grant no. 2021/41/B/HS4/02443] . For the purpose of Open Access, the author has applied a CC-BY public copyright license to any Author Accepted Manuscript (AAM) version arising from this submission. ⁑ Corresponding author. † Nusret Cakici, Gabelli School of Business, Fordham University, 45 Columbus Avenue, Room 510, New York, NY 10023, USA, tel. +1-212-636-6776, email: [cakici@fordham.edu](cakici@fordham.edu). ‡ Syed Jawad Hussain Shahzad, Chair of Social & Sustainable Finance, Montpellier Business School, 2300 Avenue des Moulins, 34080 Montpellier, France, [j.syed@montpellier-bs.com](j.syed@montpellier-bs.com). § Barbara Będowska-Sójka, Department of Econometrics, Institute of Informatics and Quantitative Economics, Poznan University of Economics and Business, al. Niepodległości 10, 61-875 Poznań, Poland, Barbara.bedowska[sojka@ue.poznan.pl](sojka@ue.poznan.pl) 􀟜 Adam Zaremba, 1) Department of Investment and Capital Markets, Institute of Finance, Poznan. University of Economics and Business, al. Niepodległości 10, 61-875 Poznań, Poland, [adam.zaremba@ue.poznan.pl](adam.zaremba@ue.poznan.pl) ; 2) Montpellier Business School, 2300 avenue des Moulins, 34185 Montpellier, France, a .zaremba@montpellier[bs.com](bs.com) ; 3) Department of Finance and Tax, Faculty of Commerce, University of Cape Town, South Africa. 1. Introduction Cryptocurrency literature has documented a growing list of characteristics that predict cross-sectional returns. Some of them—such as momentum, size, or reversal (Liu et al. , 2022; Bianchi et al. , 2022)—are parallels of similar phenomena in equity markets; others, such as network activity measures (Liu & Tsyvinski, 2021; Cong et al. , 2022), are inherently speciﬁc to cryptocurrencies. The emerging cryptocurrency “factor zoo”—comprising potentially noisy and correlated predictors—may require methods beyond simple portfolio sorts or cross-sectional regressions. Recent advances in machine learning methods appear to be a natural response to this task. Due to their capacity to handle vast multidimensional data, select best predictors, and account for nonlinearities and interactions (Gu et al. , 2020; Giglio et al. , 2022), machine learning models are well poised to face the cryptocurrency landscape. In this paper, we combine machine learning with asset pricing research in order to gain new insights into cross-sectional return predictability in cryptocurrency markets.

FAQs

How does the paper assess return predictability in cryptocurrency markets?

It combines a large-scale dataset covering many coins and exchanges with eight machine learning models to test how well characteristics predict cross-sectional returns.

Do complex machine learning models outperform simpler ones?

No. The benefits from model complexity are limited, and simpler methods such as ordinary least squares often outperform regression trees and neural networks.

Where do the strongest predictive signals (alphas) come from?

Alphas concentrate in hard-to-trade assets and depend critically on harvesting extreme returns on small, illiquid, and volatile coins.

Aug 3, 2026

Upload by River Wang