Open Access

ARTICLE

Cryptocurrency Market Trends: A Machine Learning-Driven Time Series Forecasting with Twitter Sentiment Integration

Mubariz Khan¹, Hafeez Ur Rehman Siddiqui², Adil Ali Saleem², Muhammad Amjad Raza^2,3, Lázaro Javier Hernández Rodríguez^4,5,6,7, Pablo Herrero García^4,8,9, Isabel de la Torre Díez^10,*

1 Department of Computer and Information Sciences, Northumbria University, Newcastle, UK
2 Institute of Computer Science, Khwaja Fareed University of Engineering and Information Technology, Rahim Yar Khan, Pakistan
3 Department of Computer Science and Information Technology, University of Lahore, 1-km Defense Road, Lahore, Pakistan
4 Escuela Politécnica Superior, Departamento de Soluciones Tecnológicas y Sistemas, Universidad Europea del Atlántico, Isabel Torres 21, Santander, Spain
5 Departamento de Ciencias de la Computación, Universidad Internacional Iberoamericana, Campeche, México
6 Department of Computer Science, Universidad Internacional Iberoamericana, Arecibo, PR, USA
7 Departamento de Ciencias de la Computación, Fundación Universitaria Internacional de Colombia, Bogotá, Colombia
8 Departamento de Ciências da Computação, Universidade Internacional do Cuanza, Cuito, Bié, Angola
9 Departamento de Ciencias de la Computación, Universidad de La Romana, La Romana, República Dominicana
10 Department of Signal Theory, Communications and Telematics Engineering, University of Valladolid, Valladolid, Spain

* Corresponding Author: Isabel de la Torre Díez. Email:

Computers, Materials & Continua 2026, 88(3), 46 https://doi.org/10.32604/cmc.2026.084269

Received 19 April 2026; Accepted 20 May 2026; Issue published 23 July 2026

Abstract

Accurate forecasting of cryptocurrency prices remains an open challenge because classical statistical models cannot capture the non-linear, sentiment-driven dynamics of these markets. This study compares three hybrid deep learning architectures—VAR-LSTM, XGBoost-LSTM, and CNN-LSTM—to determine which best forecasts Bitcoin (BTC), Ethereum (ETH), and Dogecoin (DOGE) closing prices, and to quantify the marginal predictive value of Twitter sentiment integration. Six years of hourly OHLCV data (2017–2023) are augmented with VADER-scored Twitter sentiment polarity. Each model is formulated mathematically, implemented with documented hyperparameters (epochs, dropout, units;), and trained for one-step-ahead next-hour price prediction. Performance is measured by RMSE, MAE, MAPE, R2, and Directional Accuracy (DA) across five random seeds, with paired Wilcoxon significance tests. XGBoost-LSTM achieves the best performance (RMSE = 81.547, R2 = 0.9254, DA = 80.0%), outperforming all nine literature baselines. Removing Twitter sentiment degrades DA by 14.3 percentage points (p<0.01), confirming that social media signals carry independent predictive information. Hybrid architectures consistently outperform single-model baselines; XGBoost-LSTM offers the best accuracy-to-compute ratio. VADER-enriched Twitter sentiment is a significant predictor beyond price history. Limitations include reliance on a single sentiment platform and a training window that predates several structural market events.

---

Keywords

---

Supplementary Material

Supplementary Material File

---