XAIMSM: An Explainable AI Framework for Identifying Manipulated Sentiment in Social-Media-Driven Cryptocurrency Markets

Abstract

The rise of social media platforms has made it easier to coordinate sentiment manipulation in cryptocurrency markets, such as pump and dump schemes, which involve the excessive promotion of an asset followed by its rapid sell-off, leaving retail traders the victim of the action. Existing detection systems prioritise predictive accuracy but lack transparent, auditable explanations required for regulatory compliance. This study presents XAIMSM, an Explainable Artificial Intelligence framework combining four publicly available datasets (Telegram pump and dump events, Twitter seeding activity, sentiment of the influencers, Bitcoin tweets). FinBERT text embeddings are fused with 47 hand-crafted behavioral, temporal, network, and market features, resulting in an 815-dimensional feature vector. Global and instance level interpretation by using an XGBoost with SHAP TreeExplainer. On a held out 2023 test set consisting of 187,108 observations, XAIMSM outperforms all ablation baselines with F1 score 0.924, AUC ROC 0.961 and MCC 0.887. The most impactful features are behavioural indicators such as sentiment velocity, cross-platform co-occurrence, influencer weight, posting frequency deviation, and volume anomaly, while raw sentiment polarity plays a minor role, confirming manipulation as a behavioural phenomenon. This framework offers regulatory agencies auditable decision trails aligned with the transparency requirements of the EU MiCA regulation.