Time-Aware Explainability for Cryptocurrency Forecasting: Evidence from a User Study

Abstract As Artificial Intelligence (AI) becomes increasingly integrated into financial decision-making, concerns about model opacity and regulatory compliance have come to the forefront. These concerns are particularly important in settings where forecasts influence risk assessment, portfolio decisions, and the credibility of AI-supported financial advice. This study examines how eXplainable AI (XAI) techniques can enhance transparency in the context of cryptocurrency forecasting–a domain characterized by extreme volatility and uncertainty. Using historical Bitcoin prices and macroeconomic indicators, we compare two deep learning approaches and apply TimeSHAP, a sequentially-aware explanation method, to clarify their predictions. Rather than focusing solely on model performance, we assess how different explanation formats are interpreted by finance-oriented users through a structured user study. Results show that local explanations are more easily understood and perceived as actionable, while more abstract forms–such as event-level and global plots–elicit mixed responses. This distinction matters because the practical value of explainability in finance depends not only on technical faithfulness, but also on whether explanations improve users’ confidence, support better-informed decisions, and help institutions justify model-based actions under scrutiny. To further support transparency and regulatory alignment, we document the entire machine learning process using a domain-specific ontology (XMLPO), offering a structured approach to traceability and reproducibility. Our findings suggest that effective explainability in financial AI requires both technically robust methods and communication strategies tailored to the needs of end-users. For financial economists and policy- and decision makers, the results indicate that the format in which AI explanations are presented can shape trust, perceived usefulness, and the governance value of predictive systems. This study contributes to the growing body of work on how AI systems can be made more interpretable, trustworthy, and aligned with economic governance in high-stakes domains.