Research record

Affective Asymmetry in News Reactions across Reddit and Bluesky

M. Trerotola, A. Esposito, M. Parente · SHIELD-AI 2026 Workshop, ECML PKDD 2026

Six instruments — lexicons, classifiers, language models — disagree by an order of magnitude on how hostile a platform is (1.2%–24.4%), yet agree on the ordering, so only relative comparisons hold. On that footing Reddit turns negative sentiment into open hostility about three times as readily as Bluesky (19.5% vs 6.5%), while the underlying affect looks the same: a difference in expression, not in feeling.

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.

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.

Towards Explainable Detection of Money Laundering in Transaction Networks

M. Trerotola, D. Calvaresi, M. Parente · WIRN 2026

Recall reported per laundering typology rather than in aggregate. Fan-out is missed by roughly 30 percentage points across Random Forest, LightGBM, XGBoost and TabNet, and targeted feature engineering at a matched budget does not close it — a limit of the account-level representation, not of any one model. The gap survives a second, more realistic generator, where fan-out recall stays at 0.62 against 0.87 for cycle.

An Explainable Machine-Learning Framework for Detecting Fraudulent Bitcoin Addresses with Graph-Temporal Features

M. Trerotola, D. Calvaresi, M. Parente · Lecture Notes in Computer Science, Springer

Classifying Bitcoin wallets as fraudulent or legitimate from graph structure and temporal behaviour, with the reasoning exposed. The single-chain framework that the cross-ledger work above extends.

doi:10.1007/978-3-032-21811-7_16code

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

AI-Driven Multi-Agent Systems for Automated Regulatory Analysis of Crypto Projects

M. Trerotola, D. Calvaresi · IEEE MetroXRAINE 2025

A supervisory agent orchestrating peers for evidence retrieval, document normalisation, semantic extraction and rule-based verification, producing an auditable compliance report and a knowledge graph from crypto-asset whitepapers.

doi:10.1109/MetroXRAINE66377.2025.11340415

Enhancing Blockchain Transaction Tracking: A Systematic Review of DLT-Based Financial Systems

M. Trerotola, D. Calvaresi · IEEE MetroXRAINE 2025

A systematic review of transaction tracing on distributed ledgers: 120 publications screened across 2017–2025, 21 retained. The corpus converges on a persistent trilemma between scalability, attribution accuracy and privacy compliance.

doi:10.1109/MetroXRAINE66377.2025.11340456

A profitable trading algorithm for cryptocurrencies using a Neural Network model

M. Parente, L. Rizzuti, M. Trerotola · Expert Systems with Applications, 238, 121806

Neural-network model for cryptocurrency price movements, evaluated on out-of-sample trading performance.

doi:10.1016/j.eswa.2023.121806