Mario Trerotola

Using Artificial Intelligence to enhance CONSOB’s enforcement against unauthorised financial activities

M. Trerotola, A. Boghi, P. Deriu, G. Frega · Consob, Quaderni FinTech, No. 19

A hybrid prototype — language models with deterministic rules — for investor complaints and the websites of reported entities, developed by Consob with the Politecnico di Torino and the Università Politecnica delle Marche. Written with the heads of the authority’s Studies and Regulation Division and of its Unauthorised Phenomena Supervision Office.

The system takes no decision on its own: it prepares, and the assessment, the decision and the responsibility for it stay with the operator.

consob.itJEL O32, O33, C55, G21

Blockchain Fraud Detection for Explainable Forensic Investigations

M. Trerotola, D. Calvaresi, M. Parente · DSA ISC 2026

Deep models are accurate but opaque; interpretable pipelines are usually tied to one chain. This method keeps both: architecture-neutral graph-temporal features, Balanced Random Forests for severe class imbalance, and a three-layer XAI stack (SHAP, CIU, DEXiRE). Macro-F1 up to 0.92 on 31,000 Bitcoin wallets and 0.985 on 157,000 Ethereum tokens, with the reasoning left inspectable.

A Hybrid Multi-Agent System for Early Scam Detection in Crypto-Assets

M. Trerotola, M. Parente, D. Calvaresi · Applied Sciences, 16(7), 3122

Seven specialised agents — coordination, retrieval, heuristic risk, MiCAR compliance, on-chain fraud detection, reconciliation — combining language models with rule-based compliance checks. Component-level testing on 150 projects gives 95% output reproducibility across five reruns and 210 s mean latency: proof of concept, not deployment evidence. A pilot with six researchers and two regulatory-authority experts supplies preliminary usability findings.

doi:10.3390/app16073122code