Forecasting cryptocurrencies in turbulent times: Evidence on parsimony versus model complexity

Abstract

This study examines short-term return forecasting for Bitcoin, Ethereum, and Litecoin over 2020–2024, comparing autoregressive benchmarks with Kitchen Sink and VARX-type models using point and density accuracy measures supported by Diebold–Mariano and Model Confidence Set inference. The results demonstrate that the AR(1) benchmark and parsimonious specifications incorporating cryptocurrency-specific variables consistently outperform the more elaborate linear frameworks considered, while the inclusion of macro-financial predictors offers limited benefits. Findings highlight the robustness of autoregressive dynamics for short-term cryptocurrency forecasting and underscore the importance of parsimony over model complexity. These results are consistent with a market environment characterised by high structural uncertainty, sentiment-driven trading and rapidly shifting regimes, in which additional macro-financial information contributes little to forecastability beyond short-run return momentum and crypto-specific volatility.