K-Means++ framework flags 2.02% suspicious trades in 2012–2024

A Clustering-Based Framework for Identifying Suspicious Trading Patterns in Capital Market implements an unsupervised fraud-detection toolkit that begins with K-Means++ clustering and applies it to roughly one million financial transactions from 2012 to 2024, flagging 2.02% of trades as suspicious and reporting a Silhouette Score of 0.561.

What did the paper do?

The authors built a clustering-based pipeline that processes a dataset of roughly one million financial transactions spanning 2012 to 2024, uses K-Means++ as the initial clustering step, and reports that 2.02% of trades are identified as suspicious. The analysis assigns those flagged trades to categories using market-practice heuristic thresholds and presents a cluster-quality metric, the Silhouette Score, of 0.561.

The paper is authored by Asif Zaman, Romona Magdalene Sarkar, Sabiha Khair Ohi and Iftekharul Mobin, submitted on 5 Jul 2026 to arXiv, and accepted for publication in the IEEE 2nd International Conference on Quantum Photonics, Artificial Intelligence & Networking (QPAIN), 2026; the authors note the final version is available via IEEE Xplore.

How does the clustering pipeline work?

The pipeline begins with K-Means++ clustering, then applies market-practice heuristic thresholds to clusters to isolate and label suspicious activity; from the set of trades flagged as suspicious, the paper reports a categorical breakdown: 51.10% spoofing, 0.10% pump and dump, 0.55% insider trading, 1.43% fake breakout and 46.83% unclassified.

The authors describe the approach as an unsupervised fraud-detection toolkit that first groups transactions with K-Means++ and then uses rule-based heuristics to map cluster members to market-manipulation types. The paper emphasizes the lack of ground truth for fraud labels in their dataset and therefore evaluates clustering quality numerically, citing a Silhouette Score of 0.561 as evidence that the clustering produced meaningful separation.

Why it matters

Unsupervised clustering offers a way to surface anomalous trading patterns when labeled fraud data is unavailable. By flagging 2.02% of trades and breaking that subset into specific market-practice categories, the method gives investigators a focused set of leads to examine rather than attempting to classify every trade. The paper’s reliance on heuristic thresholds and the explicit acknowledgement of missing ground truth underline both the practical utility and the current limits of the approach: validation hinges on follow-up analysis or the emergence of labeled datasets.

What the results actually say

The numeric details matter: the toolkit marked a small fraction of the dataset, 2.02%, as suspicious, and within that fraction over half were labeled spoofing (51.10%). Nearly half of the flagged trades remained unclassified (46.83%), which signals either novel patterns that the heuristics did not map to known manipulation types or the need for richer labeling. The Silhouette Score of 0.561 is the paper’s stated evidence that clusters are coherent enough to merit downstream heuristic labeling despite the dataset’s unlabeled nature.

What to watch

Watch for the IEEE-published final paper on IEEE Xplore and any public releases of the authors’ code or data links that would let others reproduce the clustering and the heuristic classification. The most decisive next evidence will be tests of this pipeline on datasets with ground-truth fraud labels or regulatory case data that confirm whether the 2.02% flagged trades correlate with confirmed manipulative behavior.