Automated misinformation detection flags events that coincide with sharp stock swings and retail demand shocks; a gradient-boosted classifier achieves near-perfect accuracy on the labeled corpus and the flagged events align with measurable financial and operational disruptions, suggesting firms could use such tools to spot information-driven risk earlier.

Main Finding

Automated detection of misleading news using standard ML (TF-IDF + gradient boosting or logistic regression) can reliably flag misinformation in a curated news corpus (45k articles), and aligning those detections with market and retail time series shows clear associations between misinformation events and extreme stock volatility, volume spikes, and large demand forecast errors. Integrating detection with time-series monitoring provides a practical early-warning signal for financial and operational risk.

Key Points

• Dataset and scope
  • Human-curated news dataset (≈45,000 articles: ~23.5k labeled “fake”, ~21.4k “true”) derived from public Kaggle sources; no synthetic text added.
  • Financial case studies use OHLCV series from Yahoo Finance for LLY, FPI, UAL and the S&P 500. Retail demand analysis uses Walmart weekly sales (Kaggle, ~2010–2012) plus industry reports for COVID-19 shocks.
• Misinformation detection models and performance
  • Text preprocessing: lowercasing, URL/HTML removal, punctuation removal, stop-word removal, tokenization; TF-IDF features performed best.
  • Logistic regression (L2 regularization): test accuracy 98.85%.
  • Gradient Boosting (XGBoost-style): best test accuracy 99.44%; produces a probability score usable as a “misinformation risk index.”
  • Bidirectional LSTM: underperformed (≈77.4% validation accuracy); likely data-size limits and simpler lexical cues dominated.
  • Robustness: 5-fold CV shows ≳98% mean accuracy for LR and GB (±0.1–0.2%), BiLSTM ≈76% across folds.
• Impact analysis (event study style)
  • Known misinformation incidents (e.g., Eli Lilly false tweet, UAL bankruptcy rumor, FPI short-and-distort) were flagged with high model confidence.
  • On misinformation event days: extreme returns and intraday ranges, and volume spikes (examples cited: UAL 9/8/2008 intraday plunge; FPI single-day collapse; LLY intraday move and doubled volume).
  • Retail demand: simulated sudden surges injected into historical Walmart data and external COVID-19 sales reports (e.g., March 2020 spikes) produced large Demand Forecast Errors (DFE) well beyond normal holiday/promo variation. NAB-inspired scoring used to evaluate anomaly detection timeliness and severity.
• Implementation notes
  • Models trained and evaluated in Python (scikit-learn, XGBoost, Keras/TensorFlow).
  • Detection outputs were used to label event days and correlate with volatility/DFE metrics rather than to produce causal inference.

Data & Methods

• News classification
  • Data: Fake.csv and True.csv from a public Kaggle compilation (Bozkus, 2022), cleaned to 45k entries.
  • Feature engineering: TF-IDF vectorization (outperformed bag-of-words CountVectorizer).
  • Models: Logistic regression (L2), Gradient Boosting (tuned tree ensemble), BiLSTM (embedding + LSTM layers, dropout).
  • Evaluation: 80/20 train/test split; 5-fold CV for robustness; confusion matrices inspected for false negatives/positives.
• Financial impact analysis
  • Data: Yahoo Finance OHLCV series, adjusted for splits/dividends; rolling 20-day moving averages and standard deviations for baseline volatility.
  • Event flags: dates of documented misinformation incidents; volume spikes defined as >3σ above 3-month moving average.
  • Metrics: daily returns, intraday high–low range normalized by prior close, volume ratios relative to baseline.
• Retail demand analysis
  • Data: Walmart weekly sales dataset (store & department aggregation) and external industry reports for pandemic period.
  • Forecasting: seasonal ARIMA and year-over-year naive forecasts to compute Demand Forecast Error (DFE = % deviation of actual from forecast).
  • Stress test: injected artificial surge weeks into historical data to gauge DFE magnitude under misinformation-driven panic buying.
  • Scoring: anomaly detection principles from Numenta Anomaly Benchmark to prioritize early detection and penalize misses.
• Limitations noted in methods
  • News dataset bias: “True” articles largely from reputable agencies; “Fake” from known hoax/propaganda sites — may make classification easier than in noisier real-world inputs (e.g., social media, short-form posts).
  • BiLSTM hampered by dataset size and dominant lexical signals; result may differ with far larger or more diverse corpora.
  • Retail surge tests partly simulated; direct retailer forecasts (proprietary) were not available for exact DFE comparison.

Implications for AI Economics

• Valuation of information quality
  • Firms and markets can treat misinformation risk as a quantifiable input into value-at-risk and scenario analysis. A reliable misinformation risk index can be priced into trading strategies, liquidity planning, and capital buffers.
• Market microstructure and automated trading
  • High-confidence, rapid detection systems can mitigate algorithmic over-reaction (flash crashes) by informing circuit-breaker logic or temporarily throttling automated order execution during detected misinformation spikes.
• Operational forecasting & inventory management
  • Retailers can integrate real-time misinformation signals into demand-forecast pipelines to trigger contingency actions (safety-stock adjustments, targeted communications, dynamic replenishment) and reduce forecast error costs during sudden information-driven runs.
• Policy and regulation
  • Regulators and exchanges could use detection outputs as part of market surveillance to identify potential market-manipulation campaigns (e.g., “short-and-distort”) and to design disclosure/flagging requirements for suspicious information that materially affects prices.
• Practical deployment considerations
  • False negatives are costly: models should be tuned to minimize missed misinformation even at some cost of false positives; interpretation and human-in-the-loop workflows remain important.
  • Generalization risk: models trained on curated news articles may underperform on social media, non-English sources, or multimedia misinformation; domain adaptation and multimodal models will be needed.
  • Adversarial dynamics: actors may adapt language to evade detection; continuous retraining, adversarial testing, and provenance/metadata signals (source reputation, publication timing) should be combined with text models.
• Research extensions valuable to AI economics
  • Move from association to causal inference: use high-frequency event-study methods, instrumental variables, or synthetic control approaches to better quantify causal financial losses attributable to misinformation.
  • Broaden inputs: include social media streams, messaging apps, images/video, and network propagation features to improve timeliness and coverage.
  • Estimate macroeconomic externalities: assess aggregate effects on market liquidity, cross-asset spillovers, and consumer welfare from misinfo-driven shocks.
  • Productization: build real-time pipelines that convert detection probabilities into actionable economic metrics (expected P&L impact, inventory shortfall probability) for decision-makers.

Limitations and caveats: the paper demonstrates strong predictive performance on a curated news corpus and shows clear temporal associations with market/retail anomalies, but its findings do not by themselves prove causality for all events, and real-world deployment will face additional data heterogeneity, adversarial adaptation, and integration challenges.