Effects of Macroeconomic Data on Bitcoin Price Prediction using Time Series Forecasting

Abstract

As Bitcoin evolves from a peer-to-peer transaction system into a widely traded financial asset, its price movements are increasingly influenced by broader market conditions and macroeconomic signals. This paper investigates whether incorporating continuous macroeconomic variables changes Bitcoin price prediction accuracy in multivariate time series forecasting techniques. A comparative experiment is conducted using LSTM, XGBoost, GRU, TCN, and GRU-based DeepVAR model, trained on daily data from 2015 to 2025. Macroeconomic indicators including SPX, XAU, VIX, and DXY are added and evaluated through a grid search approach to avoid hyperparameter bias. Based on repeated evaluations, macroeconomic data produces model-dependent changes in forecast accuracy, including marginal improvements, degradation, or statistically null effects, with no consistent benefit observed across models. HAC testing indicates that some performance differences are statistically detectable, while McNemar tests fail to reject the null of unchanged directional prediction behavior. Against a zero-change benchmark, no model improves on the naive forecast by more than 0.4%. The absence of a macroeconomic effect is therefore measured in a setting where predictive skill is already minimal.