Published December 2, 2025 | Version v2
Preprint Open

AIMM: An AI-Driven Multimodal Framework for Detecting Social-Media-Influenced Stock Market Manipulation

Authors/Creators

  • 1. Independent Researcher

Contributors

Researcher:

  • 1. Independent Researcher

Description

AIMM (Automated Inference of Market Manipulation) is an AI-driven, multimodal framework designed to detect social-media-influenced stock-market manipulation. The system integrates market structure, Reddit-based social features, transformer-based sentiment analysis, bot-likeness heuristics, coordination density, and optional SEC regulatory signals to assess manipulation risk in near real-time.

AIMM introduces the AIMM Manipulation Risk Score (AMRS), a weighted and normalized risk metric combining:

  • Reddit social volume and author activity

  • FinBERT/VADER sentiment signals

  • Bot-heavy posting ratios

  • TF–IDF text-similarity coordination metrics

  • Market anomalies such as returns and volume z-scores

  • Optional ownership and insider-transaction features from SEC EDGAR

The framework is implemented as a reproducible end-to-end pipeline with parquet-backed storage, configurable thresholds, a rule-based alerting system, and a publicly accessible Streamlit dashboard for interactive exploration.

AIMM extends the author's prior work, the Stock-Pattern-Assistant (SPA) framework (doi:10.5281/zenodo.17618798), by moving beyond deterministic price-pattern analysis into multimodal manipulation intelligence. AIMM models how online narratives, bot-amplified messaging, coordinated activity clusters, and market microstructure events jointly influence price behavior.

The paper provides:

  • A full feature-engineering pipeline for social, market, bot, sentiment, and coordination metrics

  • Formal AMRS definition, normalization, and weighting

  • Synthetic but realistic experiments across multiple tickers (AAPL, NVDA, SCHW, AMC, GME, PGR, ALAB, XYZ)

  • Threat model and responsible-use guidelines

  • A reproducible implementation with open-source code and live dashboard

This release corresponds to AIMM v1.0 – Initial Public Release.

Abstract

AIMM v2.0 is an AI-driven, multimodal system for detecting socially amplified
market manipulation by fusing Reddit activity, bot-likeness indicators,
coordination patterns, sentiment, and OHLCV market features into a daily
Manipulation Risk Score.

This v2.0 release introduces three major additions beyond the initial prototype:

1. AIMM–GT v2.0 ground-truth dataset:
   A curated set of 33 labeled ticker-days spanning eight equities, combining
   SEC enforcement actions, community-verified manipulation episodes, and
   matched negative controls.

2. Forward-walk evaluation pipeline:
   A no-lookahead historical evaluation framework that assesses how AIMM would
   have behaved if deployed during past events, enabling ROC–AUC, PR–AUC,
   confusion matrix analysis, and ranking-based diagnostics.

3. Prospective prediction logging and lead-time analysis:
   A live prediction log that records daily AIMM outputs over time and allows
   comparison with future labels to quantify real deployment performance.
   Lead-time analysis demonstrates that AIMM raises elevated risk for GME
   approximately 22 days before the January 2021 meme-stock event.

The paper included in this release describes the full architecture, feature
engineering pipeline, AMRS formulation, dataset schema, evaluation methodology,
and case studies. AIMM v2.0 transforms AIMM from an exploratory dashboard into a
quantitatively evaluated early-warning system for social-media-influenced market
surveillance.

This Zenodo release includes:
• AIMM v2.0 paper (PDF)
• AIMM–GT v2.0 labeled dataset (CSV)
• Evaluation scripts for forward-walk and prospective prediction logging
• Streamlit dashboard design for interactive exploration

AIMM is released to support reproducible research, regulatory studies, and the
development of trustworthy AI systems for market manipulation detection.

Files

AIMM_v2.0_Multimodal_Market_Manipulation_Detection_Preprint.pdf

Files (1.4 MB)

Name Size Download all
md5:dd98f317a0e45df9367d4216bf2c1ac3
3.3 kB Preview Download
md5:caf1a9dd2aa28861fa953146d78b4e4f
1.4 MB Preview Download

Additional details

Related works

Is supplement to
Preprint: 10.5281/zenodo.17618798 (DOI)

Dates

Submitted
2025-11-24

Software

Repository URL
https://github.com/sneela/aimm
Programming language
Python
Development Status
Active

References

  • [1] S. Neela, "Stock-Pattern-Assistant (SPA): Deterministic run detection and pattern-aware analysis of equity time series," Zenodo, Jan. 2025. doi:10.5281/zenodo.17618798. [2] F. Allen and D. Gale, "Stock price manipulation," The Review of Financial Studies, vol. 5, no. 3, pp. 503–529, 1992. [3] J. Bollen, H. Mao, and X. Zeng, "Twitter mood predicts the stock market," Journal of Computational Science, vol. 2, no. 1, pp. 1–8, 2011. [4] T. Rao and S. Srivastava, "Intra-day stock market prediction using online textual messages," in 2010 IEEE/WIC/ACM International Conference on Web Intelligence and Intelligent Agent Technology, vol. 1. IEEE, 2010, pp. 423–426. [5] J. J. Choi et al., "Reddit and the meme stock phenomenon," Yale School of Management Working Paper, 2022. [6] O. Varol et al., "Online human-bot interactions: Detection, estimation, and characterization," in Proceedings of the International AAAI Conference on Web and Social Media, vol. 11, no. 1, 2017, pp. 280–289. [7] D. Pacheco et al., "Uncovering coordinated networks on social media," in Proceedings of the 15th International AAAI Conference on Web and Social Media, 2021, pp. 128–139.