Explainable AI for tail-risk forecasting in cryptocurrencies: evidence from green and energy-intensive assets for risk-aware investment decisions

Purpose This study aims to develop an explainable deep learning framework for forecasting tail risk and volatility in cryptocurrency markets, with a particular focus on differences between energy-intensive and ESG-oriented digital assets. It aims to assess how volatility dynamics and downside risk exposure vary across environmental profiles and market regimes and how these differences can inform risk-aware investment decisions. Design/methodology/approach The proposed framework uses a quantile long short-term memory model to estimate predictive quantiles and Conditional Value at Risk (CVaR), enabling the modeling of asymmetric, fat-tailed and regime-dependent risks. The analysis focuses on four cryptocurrencies – Bitcoin and Litecoin (energy-intensive) and Algorand and Chia (lower-energy alternatives). The model integrates SHapley Additive exPlanations-based explainability and a regime classification (calm vs stress). Robustness is assessed using 20-, 40- and 60-day rolling evaluation windows, while predictive performance is evaluated through coverage rates, hit ratios and Kupiec and Christoffersen back testing procedures. Findings The results reveal substantial heterogeneity across the four cryptocurrencies rather than a uniform green-versus-energy-intensive ranking. Chia records the lowest Q50 root mean squared error (0.00149) and mean absolute error (MAE) (0.00119), whereas Algorand has the lowest full-sample CVaR (0.0293) and Litecoin the highest (0.0411). Regime results are also asymmetric: Bitcoin’s Q10 violation rate rises from 8.74% in calm periods to 11.74% in stress periods, while Litecoin and Algorand record more Q10 violations in calm than in stress regimes. Chia combines the lowest global Q10 violation frequency (2.98%) with the highest mean CVaR violation loss (0.127419), showing that infrequent breaches can still be severe. These findings support asset-specific tail-risk monitoring and cautious use of energy-profile information in risk-aware investment decisions. Originality/value This study contributes to the literature by integrating quantile-based deep learning, explainable artificial intelligence, regime-specific diagnostics and sustainability considerations into a unified risk-forecasting framework. Its main value lies in complementing connectedness and spillover studies with interpretable, asset-level distributional forecasts and formal tail-risk backtesting. The results provide decision-support information for risk monitoring and portfolio analysis, but do not imply that lower-energy cryptocurrencies are universally safer or that the observed differences are causal.