Predicting Returns with Machine-Learned News Sentiment

Key Performance Metrics

📊 How Well Does This Strategy/Model Perform?

• Sharpe Ratio (Equal-Weighted L/S): 4.29
• Average Daily Return (L/S): 33 bps
• Risk-Adjusted Alpha (FF5+MOM): 4.3% per day (EW L/S)
• Outperformance vs RavenPack: SESTM earns 33 bps vs 18 bps/day

💡 Takeaway:
SESTM delivers strong predictive power and out-of-sample alpha. It outperforms both RavenPack and dictionary-based methods while using a transparent, white-box ML approach.

Key Idea: What Is This Paper About?

The paper introduces SESTM (Sentiment Extraction via Screening and Topic Modeling), a three-step supervised learning model to extract predictive sentiment from financial news. Instead of relying on pre-built dictionaries, SESTM learns which words matter for future returns and assigns them custom weights. It shows this model yields better predictions and more profitable trading strategies.

Economic Rationale: Why Should This Work?

📌 Relevant Economic Theories and Justifications:

• Limited Attention: Investors process news with delay, especially for small or volatile stocks.
• Textual Complexity & Novelty: Fresh news and complex sentiment signals are assimilated into prices more slowly.
• Information Frictions: Differences in speed of assimilation suggest inefficiencies that can be exploited.

📌 Why It Matters:
SESTM shows how tailored sentiment extraction reveals information traditional methods miss, enabling consistent alpha generation using public news.

How to Do It: Data, Model, and Strategy Implementation

Data Used

• News Source: Dow Jones Newswires (1989–2017)
• Price Data: CRSP, TAQ
• Training Sample: 1989–2012
• Out-of-Sample: 2004–2017
• Articles Processed: ~6.5 million single-stock articles

Model / Methodology

1. Screening: Select sentiment-charged words using correlation with return direction.
2. Topic Modeling: Assign word weights based on predictive ability using supervised topic modeling.
3. Scoring: Aggregate to article-level sentiment with penalized likelihood.

• Key Tuning: Optimized using cross-validation on 5-year rolling windows.
• Portfolio Construction:
  • Long top 50 sentiment stocks
  • Short bottom 50
  • Rebalanced daily, open-to-open returns

Trading Strategy (SESTM Long-Short)

• Weighting Schemes: Equal-weighted (preferred), and value-weighted
• Holding Period: 1 day (can be extended with decay-weighting)
• Transaction Costs: Handled via turnover-constrained variants (EWCT strategy)

Key Table or Figure from the Paper

📌 Explanation:

• Shows performance of equal- and value-weighted long, short, and long-short portfolios using SESTM sentiment.
• The equal-weighted long-short portfolio consistently outperforms S&P 500 (SPY), including during crises.
• The strategy generates pure alpha: over 90% of return not explained by FF5+MOM factors.

Final Thought

💡 A machine that reads news and beats the market? SESTM proves it's possible—with transparency and theory to back it up. 🚀

Paper Details (For Further Reading)

• Title: Predicting Returns With Text Data
• Authors: Zheng Tracy Ke, Bryan T. Kelly, Dacheng Xiu
• Publication Year: 2019 (NBER Working Paper 26186)
• Journal/Source: NBER Working Paper
• Link: https://www.nber.org/papers/w26186