An AI-driven market-surveillance system that fuses social-coordination signals with price data produces early warning scores—claiming to flag GME 22 days before the January 2021 squeeze—but the public ground-truth release covers only 33 ticker-days (3 positives) and relies on synthetic Reddit features.

Main Finding

AIMM (Automated Inference of Market Manipulation) is an AI-driven, multimodal framework that fuses social-media signals (Reddit-derived) with market microstructure features (OHLCV) to produce a daily AIMM Manipulation Risk Score (AMRS) per ticker. On a small curated ground-truth (AIMM–GT; 33 labeled ticker-days, 3 positives), AIMM shows preliminary discriminative capability and provides early warnings (notably flagging GME ~22 days before the January 2021 squeeze peak). The project releases code, dataset schema, and a Streamlit dashboard to support further research.

Key Points

• Core contribution: AMRS — a composite daily risk score combining five components (social volume, sentiment, bot activity, coordination, market anomalies).
• Multimodal fusion: aligns Reddit-derived social aggregates with daily OHLCV market data in parquet-native fused ticker-day windows.
• Feature highlights:
  • Social: social_volume, avg_sentiment (FinBERT primary + VADER fallback), unique_authors, avg_bot_score, bot_heavy_post_ratio.
  • Coordination: fraction of near-duplicate posts measured by TF–IDF embeddings and pairwise cosine similarity (threshold τc = 0.8; sample up to 200 posts/day).
  • Market: returns, rolling volume mean/std, volume_zscore, binary volume anomaly flag (z ≥ 2.0).
  • Bot-likeness heuristic: author-level score driven ~70% by posting frequency and ~30% by low subreddit diversity; ticker-day bot score = fraction of posts from high-bot-score authors.
• AMRS composition and thresholds:
  • Normalized components via max-scaling; AMRS = weighted sum with default weights (wvol, wsent, wbot, wcoord, wmkt) = (0.25, 0.15, 0.20, 0.20, 0.20).
  • Risk levels: Low < 0.2, Medium [0.2,0.5), High ≥ 0.5. A suspicious window requires High AMRS plus at least one supporting anomaly (e.g., volume spike, large return, high coordination).
• Evaluation approach:
  • Forward-walk evaluation and prospective logging to avoid temporal leakage (features normalized using only data up through day t).
  • Weight perturbation sensitivity: ±20% changes produce small ranking shifts; reported ROC–AUC range ~0.71–0.74 under perturbation (paper-level claim).
• Practical outputs: parquet pipeline, interactive Streamlit dashboard for analysts, and an annotation protocol for AIMM–GT.
• Data constraints: historical Reddit posts for 2021 unavailable due to Pushshift shutdown; social features in this release are calibrated synthetic signals (statistically matched to published meme-stock characterizations). Market OHLCV is real (Yahoo Finance).

Data & Methods

• Data sources:
  • Market: daily OHLCV from Yahoo Finance (real historical data).
  • Social: calibrated synthetic Reddit features (designed to match documented patterns from published meme-stock studies due to Pushshift/Reddit API gaps).
  • Ground truth (AIMM–GT v2.0): 33 labeled ticker-days across eight equities (Jan 2021–Dec 2024): 3 positive manipulation events (from SEC enforcement, community-verified cases) and 30 matched negative controls. Annotation protocol and provenance included.
• Storage & pipeline:
  • Columnar parquet format with fused ticker-day rows. Three stages: core fused windows → enhanced fused windows (coordination etc.) → scored windows (AMRS).
  • Deterministic forward-walk processing to ensure reproducibility and prevent look-ahead leakage.
• Feature engineering (selected formulas / heuristics):
  • returni,t = (closei,t − closei,t−1) / closei,t−1.
  • volume_zscorei,t = (volumei,t − rolling_mean(volume)) / rolling_std(volume); anomaly if z ≥ 2.0.
  • Bot-likeness per author: B(u) = wf·1[f(u) > τf] + wd·1[d(u) < τd], with wf=0.7, wd=0.3, τf=10 posts/day, τd=3 subreddits; ticker-day bot ratio = fraction of posts from authors with B>0.5.
  • Coordination: sample up to N=200 posts/day, TF–IDF vectors (unigrams + bigrams, d≈1000), compute pairwise cosine similarities and record fraction above τc=0.8.
  • Sentiment aggregation: Ssent = 0.4·VADERavg + 0.6·FinBERTmapped (FinBERT mapped to [−1,1] by P(pos)−P(neg)).
• AMRS calculation:
  • Components are max-scaled per series to [0,1] (plus small epsilon).
  • smkt = max(normalized volume_zscore, normalized |return|).
  • AMRS = weighted sum of normalized components, clipped to [0,1], then bucketed into risk levels.
• Evaluation:
  • Forward-walk per labeled day (compute features and score using only data available up to that day).
  • Prospective logging to match live predictions against future realized outcomes.
  • Robustness checks: ±20% weight perturbations, temporal normalization, ablations (described in appendices).
• System architecture:
  • Four layers: ingestion (Reddit, market, ancillary filings), feature engineering (sentiment, bot, coordination), risk scoring (AMRS), presentation (Streamlit dashboard + alerting rules).

Implications for AI Economics

• Operational surveillance: AIMM demonstrates a practical architecture for integrating social narratives with market data to produce explainable, auditable risk signals. This is directly relevant for exchanges, broker-dealers, and regulators seeking early-warning systems for social-media-driven manipulation.
• Measuring social influence: By formalizing coordination and bot-likeness alongside market anomalies, AIMM provides a template for quantifying the causal channel from coordinated online narratives to price/volume dynamics — a crucial measurement problem in AI economics and market microstructure.
• Research infrastructure: The parquet-native pipeline, fused ticker-day schema, and forward-walk evaluation pattern are useful methodological contributions for reproducible research on social-market interactions.
• Policy and ethics:
  • Transparency and explainability: AIMM’s componentized score and dashboard align with regulatory needs (SEC/ESMA emphasis) and support human-in-the-loop decision-making.
  • Data access gap: The reliance on calibrated synthetic social features highlights a major barrier — API/archival access to historical social data — that constrains empirical validation and policy oversight.
  • False positives / market impact: Deployment risks include spurious alerts that could affect trading decisions, and adversarial adaptation by manipulators (e.g., changing language or timing to evade heuristics). Any operational use must consider cost of action on false positives and require governance.
• Limitations and next steps for the field:
  • Ground-truth sparsity: AIMM–GT is very small (33 ticker-days, 3 positives). Statistical claims must be tempered; scaling labels (via regulator cooperation or larger investigative datasets) is essential.
  • Synthetic social features: While calibrated to literature, synthetic signals cannot substitute for full validation on real historical social traces. Future work should re-evaluate AIMM once reliable social archives are available.
  • Richer market data: Incorporating order-book/trade-level data (not just daily OHLCV) would likely improve detection of classical manipulative microstructure strategies combined with social campaigns.
  • Adversarial robustness & privacy: Research needed on robustness to evasion and privacy-preserving architectures for monitoring user-generated content.
• Practical value for AI economics: AIMM is a concrete step toward operationalizing the study of how algorithmically mediated social narratives alter financial markets. Even as a prototype, its fusion approach and evaluation discipline (forward-walk, prospective logging, explainable components) provide a useful blueprint for subsequent academic, regulatory, and industry work.

Notes and caveats - The reported positive result (early GME flagging) is promising but anecdotal given small labeled set and synthetic social inputs. - ROC–AUC and robustness claims are reported in the paper; they should be re-assessed on larger, real-data-labelled benchmarks before any deployment.