AIMM-X: An Explainable Market Integrity Monitoring System Using Multi-Source Attention Signals and Transparent Scoring
Kyunghyun Cho [email protected], Raia Hadsell [email protected], Hugo Larochelle [email protected] · 2026
Read AIMM-X: An Explainable Market Integrity Monitoring System Using Multi-Source Attention Signals and Transparent Scoring by Kyunghyun Cho [email protected], Raia Hadsell [email protected], Hugo Larochelle [email protected] (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-15304.txt.
Introduction
Kyunghyun Cho, Raia Hadsell, and Hugo Larochelle — as credited on the record — present a market watchdog built to be read, not just trusted. Their 2026 arXiv paper deals with market manipulation, which is hard to catch in practice. The strongest surveillance systems lean on private order-book data and opaque models.
That secrecy blocks reproduction and independent checks. AIMM-X uses only public signals instead: standard price and volume data, plus proxies for public attention from several sources. It flags suspicious windows where returns, volatility, and attention all stray from their past baselines. Its scoring stays transparent.
Catalog record
| CREATOR | Kyunghyun Cho [email protected], Raia Hadsell [email protected], Hugo Larochelle [email protected] |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 98 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2601-15304
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
AIMM-X: An Explainable Market Integrity Monitoring System Using Multi-Source Attention Signals and Transparent Scoring Kyunghyun Cho [email protected], Raia Hadsell [email protected], Hugo Larochelle [email protected] Abstract Abstract Market manipulation remains difficult to detect in practice because the strongest surveillance systems rely on proprietary order-book data and opaque models, limiting reproducibility and independent validation. We present AIMM-X , an explainable market-integrity monitoring framework that uses only publicly accessible signals —standard OHLCV price/volume data combined with multi-source attention proxies—to surface suspicious windows : contiguous time intervals where returns, volatility, and public attention jointly deviate from their historical baselines. AIMM-X applies transparent statistical scoring and hysteresis-based segmentation to generate candidate windows, then ranks them using an interpretable Integrity Score that decomposes into factor contributions ( ϕ 1 \phi_{1} – ϕ 6 \phi_{6} ), enabling auditability and analyst review. In a one-year demonstration on a 24-ticker universe (daily bars, 2024), AIMM-X detects 233 suspicious windows and produces fully reproducible artifacts (window lists, rankings, factor attributions, and case study reports). The system is intentionally framed as triage rather than accusation : outputs represent evidence for human investigation, not proof of wrongdoing. We conclude by outlining a validation roadmap that combines retrospective checks against publicly documented enforcement actions, expert annotation studies, and higher-frequency data integration to support operational deployment and responsible regulatory engagement. Keywords: Market manipulation detection, explainable AI, market microstructure, social attention signals, algorithmic surveillance, financial market integrity, reproducible research 1 Introduction Financial markets serve as the lifeblood of modern economies, channeling capital from savers to productive enterprises, enabling risk management, and facilitating price discovery. Yet these critical functions depend entirely on market integrity—the confidence that prices reflect genuine supply and demand rather than manipulation, fraud, or coordinated distortion. When manipulative practices go undetected, the consequences extend far beyond individual losses: investor confidence erodes, capital allocation becomes inefficient, and market participation shrinks ( 36 ; 4 ; 25 ) . The challenge of detecting market manipulation has intensified dramatically in recent years. Modern markets operate at speeds measured in microseconds, with algorithmic trading accounting for the majority of volume in major exchanges. Retail investors now coordinate through social media platforms, creating attention-driven volatility spikes that can overwhelm traditional surveillance systems ( 31 ; 24 ) . Meanwhile, sophisticated actors have adapted their techniques, making manipulation harder to distinguish from legitimate trading ( 49 ; 52 ) . 1.1 The Gap in Current Surveillance Capabilities Most manipulation detection systems today face a fundamental tension between effectiveness and accessibility. On one hand, exchange-operated surveillance platforms like NASDAQ’s Smarts and NYSE’s Mercury have access to comprehensive order book data, proprietary trade-level feeds, and sophisticated pattern recognition algorithms. These systems can detect complex manipulation schemes like spoofing, layering, and momentum ignition with reasonable accuracy ( 60 ; 59 ) . However, these powerful tools remain locked behind proprietary walls. Academic researchers, smaller broker-dealers, compliance teams at regional firms, and independent analysts cannot validate their methods, replicate their findings, or extend their approaches. This creates several critical problems: Reproducibility crisis. When surveillance methodologies remain opaque, the research community cannot build on prior work, validate detection algorithms, or establish best practices. Published studies often rely on restricted datasets or proprietary features, making independent replication nearly impossible ( 33 ; 44 ) . Black-box opacity. Modern detection systems increasingly employ machine learning models that, while potentially accurate, provide little insight into why a particular trading episode was flagged. For compliance reviews, regulatory proceedings, or enforcement actions, this opacity is problematic—suspicious activity needs to be explained, not just identified ( 43 ; 9 ) . Access inequality. Only the largest market participants can afford comprehensive surveillance infrastructure. This creates an asymmetry where sophisticated manipulators can exploit blind spots at smaller institutions, regional exchanges, or emerging markets where surveillance capabilities lag ( 2 ; 16 ) . Data dependency. The most effective detection approaches require order book depth, trade-level microstructure, or proprietary attention feeds that are either unavailable to most researchers or prohibitively expensive ( 23 ; 28 ; 46 ) . 1.2 The AIMM-X Approach: Transparency Meets Practicality This paper introduces AIMM-X (AI-driven Market Integrity Monitor with Explainability), a detection framework designed from the ground up to address these limitations while maintaining practical effectiveness. Unlike proprietary surveillance systems, AIMM-X operates on publicly accessible data: standard OHLCV (Open, High, Low, Close, Volume) price series supplemented with attention signals derived from social media, news, Wikipedia page views, and search trends—data sources available to any researcher or practitioner. The core innovation lies not in data exclusivity but in methodological transparency and interpretability. AIMM-X detects "suspicious windows"—short temporal intervals where price movements, volatility, and attention exhibit unusual co-movement patterns—and ranks them using an Integrity Score that can be completely decomposed into interpretable factors. Every alert comes with evidence: which signals triggered the flag, how much each factor contributed, and what patterns drove the scoring. This transparency serves multiple purposes. For researchers, it enables validation, replication, and systematic improvement. For compliance teams, it provides an auditable trail from raw data to final alert. For regulators, it offers a framework where detection logic can be examined, tested against known manipulation cases, and refined through open collaboration. Importantly, AIMM-X is designed as a triage system rather than a definitive accusation engine. Detected windows represent candidates for deeper investigation—not proof of wrongdoing. This framing aligns with responsible surveillance practices where automated detection surfaces anomalies for human review, expert analysis, and potential escalation ( 59 ; 60 ) . 1.3 Motivation: Learning from Historical Detection Failures Recent market events underscore the need for better surveillance tools. The January 2021 GameStop episode demonstrated that massive price and volume surges driven by coordinated retail attention can emerge with minimal warning, overwhelming traditional volatility-based alerts ( 31 ) . The 2010 Flash Crash revealed how algorithmic cascades can create systemic instability faster than human supervisors can respond ( 35 ) . Multiple pump-and-dump schemes propagated through spam and social media have gone undetected until after significant retail losses accumulated ( 26 ) . These cases share a common pattern: by the time traditional surveillance systems flagged the activity, much of the damage had already occurred. AIMM-X’s approach—combining price dynamics with multi-source attention signals—is designed to provide earlier, more contextual alerts by explicitly modeling the co-movement between market activity and public attention that characterizes many modern manipulation schemes. While we cannot claim that AIMM-X would have definitively detected these historical episodes in real-time (ground truth labels are inherently difficult in manipulation detection), the framework provides a foundation for systematic validation against known enforcement cases. Such validation represents a key direction for future work and a pathway toward operational deployment. Objective. Our objective is to build an auditable and explainable market-integrity monitoring framework that surfaces suspicious windows —short time intervals where price/volume dynamics co-move unusually with public attention—using only data that is realistically accessible to researchers and smaller institutions. Rather than attempting to prove manipulation, AIMM-X produces a ranked set of windows with transparent factor attribution so analysts can decide what warrants deeper review. Limitation addressed. Much of the market-manipulation and surveillance literature either (i) depends on proprietary, transaction-level microstructure data and trader identifiers that are unavailable to most researchers, or (ii) uses black-box anomaly models that are difficult