Predicting Overnight Stock Changes Post-Earnings: Tomorrow’s Price Hidden in Today’s Language

2026-05-24 · Source: Machine Learning on Medium · Field: Finance & Economics — Capital Markets & Investment Management, FinTech & Digital Financial Services · Depth: Advanced, extended

Summary

A machine learning project successfully predicted overnight stock price movements following earnings announcements by analyzing language patterns in earnings call transcripts. Utilizing data from Wharton Research Data Services (WRDS) spanning 2007-2024, the project combined guidance surprise metrics, domain-specific NLP via the Loughran-McDonald financial dictionary, and custom word importance analysis. Stacked ensemble models achieved 58.2% directional accuracy in forecasting whether stocks would open higher or lower, representing an 8 percentage point improvement over random chance after rigorous, leak-proof validation. Key findings revealed that while forward-looking confidence correlated with returns, it degraded predictive performance due to market overhype. Conversely, clear communication negatively correlated with complexity. The ensemble's meta-learning approach notably exploited anti-correlated errors between XGBoost and LightGBM models, contributing to a 1.67 Sharpe ratio for the strategy.

Key takeaway

For quantitative trading firms or risk managers building systematic strategies around earnings events, this research demonstrates that linguistic analysis of earnings call transcripts provides a genuine statistical edge. You should integrate domain-specific NLP and stacked ensemble models to predict overnight stock movements with 58.2% directional accuracy. Prioritize rigorous temporal validation to avoid data leakage, ensuring your models deliver a reliable 1.67 Sharpe ratio for alpha generation and enhanced risk management.

Key insights

Earnings call language, processed with domain-specific NLP, predicts overnight stock movements.

Principles

• Domain-specific NLP tools outperform generic ones in specialized contexts.
• Rigorous, leak-proof validation is crucial for honest model performance in finance.
• Ensemble models can exploit anti-correlated errors for incremental gains.

Method

A two-layer ML system: classification for direction, regression for magnitude. Combines financial features with NLP features (sentiment, tone, forward-looking language, readability). Uses stacked ensemble with meta-learning.

In practice

• Integrate NLP models into systematic earnings trading strategies.
• Prioritize earnings calls for analysis based on linguistic signals.
• Optimize corporate earnings communication for clarity and impact.

Topics

• Financial NLP
• Stock Market Prediction
• Earnings Call Analysis
• Ensemble Machine Learning
• Data Leakage Prevention
• Quantitative Trading

Best for: NLP Engineer, Research Scientist, Machine Learning Engineer, Data Scientist, AI Scientist

Related on AIssential

• Efficient Financial Language Understanding via Distillation with Synthetic Data — Distillation with synthetic data, guided by clustering real examples, enables efficient financial NLP in low-resource settings.
• McKenzie Marshall: NLP in Asset Management (Barings) — Effective NLP in asset management augments human analysis by precisely identifying and contextualizing company-specific information from diverse text sources.
• I spent 4 days training an AI model to understand my country. It barely works — Low-resource NLP projects face significant challenges in data collection, translation, and accurate labeling, often yielding modest baselines.
• Ensemble RL through Classifier Models: Enhancing Risk-Return Trade-offs in Trading Strategies — Combining ensemble RL with classifiers and variance-based filtering significantly enhances financial trading risk-return profiles.
• The Math Behind the Bots Dominating Prediction Markets — Strict mathematical models, not sentiment, drive consistent profitability in prediction markets.
• FinBalance: A Multi-Document Accounting Reconciliation Benchmark — LLMs struggle with multi-document accounting reconciliation, showing low accuracy and issues with document binding and consistent aggregation.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning on Medium.