Forecasting Returns for High-Frequency Cryptocurrency WebSocket Data

Abstract

In this paper, I explore various machine-learning models for predicting high-frequency returns for four of the most popular cryptocurrency perpetual futures trading pairs: BTCUSDT, ETHUSDT, MATICUSDT, and SOLUSDT. Specifically, I train and evaluate models for classifying the direction of the smoothed mid-price change for high-frequency prediction horizons. I introduce a novel data set constructed from a live WebSocket stream from Binance, the world's largest centralized cryptocurrency exchange. I explore how different data representations affect model performance and how performance varies for different trading pairs and prediction horizons. I use a mixture of traditional machine learning and deep learning models and show that simple and explainable traditional models can rival the performance of far larger and more complex state-of-the-art deep learning models.