Authors: Anna Scherbina, Bernd Schlusche

Publication Year: 2016

Abstract Summary

This paper explores the phenomenon where returns of certain stocks can predict the returns of others. By employing Granger causality regressions, the authors identify leader-follower pairs among individual stocks. These leader stocks can forecast the returns of follower stocks out-of-sample, and this predictability is shown to exist beyond industry-level correlations. A firm, regardless of its size, may emerge as a return leader if it is at the center of major news developments. Moreover, the number of followers a stock has is shown to increase when it experiences newsworthy events. The study provides evidence that return leadership can be both short-lived and tied to firm-specific news rather than broader market factors.

Key Ideas by Page

Page 1, Paragraph 2

This section raises the core research question: can one firm’s stock returns predict another’s, especially when firms are not linked by fundamentals? The authors highlight this as a sign of information flow inefficiencies within the market, setting the stage for exploring return leadership beyond traditional firm connections.

Page 3, Paragraph 4

Here, the methodology is outlined: the authors use Granger causality tests on stock returns to identify predictive relationships. Importantly, their approach does not rely on pre-defined connections like shared industries, allowing them to discover surprising and potentially lucrative leader-follower dynamics.

Page 15, Paragraph 3

This paragraph evaluates trading strategies based on the identified leader-follower relationships. The results show significant returns even after accounting for known return predictors and transaction costs, implying that these strategies effectively exploit slow information diffusion between firms.

Data and Data Fields Used

  • Data Sources: CRSP for stock returns, Thomson-Reuters News Analytics for news data.
  • Frequency: Monthly and weekly stock returns.
  • Fields:
    • Stock return data
    • Market capitalization
    • Industry classification
    • News story counts and relevance
    • Firm fundamentals like size, turnover, age
    • Institutional ownership and analyst coverage

The regression analysis uses lagged returns of stock pairs and market return to compute predictive relationships.

Potential Issues

1. False Discovery of Leaders

  • Noise from Estimation: Statistical noise in Granger regressions may falsely identify leaders.
  • Fat-tailed Distributions: The use of t-statistics (cutoff = 2.0) leads to roughly 150 false leaders per stock, as actual distribution of coefficients is non-normal.

2. Non-Stationarity in Returns

  • Time-varying Relationships: Leader-follower dynamics may shift over time, affecting predictive validity.
  • Decay of Signal: Effectiveness of leader signals diminishes beyond a one-month horizon, limiting long-term utility.

3. Data Mining Risk

  • Out-of-Sample Testing: Although attempted, repetitive data slicing may still yield spurious patterns.
  • Scrambled Data Test: While helpful, it doesn’t entirely eliminate concerns of overfitting due to retrospective analysis.

4. Limited Applicability in Real-World Trading

  • Transaction Costs: Weekly strategies require high turnover, incurring significant trading costs.
  • Scalability: Profitability may not scale due to market impact and capacity limits, especially for large portfolios.

Reflection

This paper makes a significant contribution to the understanding of cross-firm information diffusion and its practical implications for return predictability. It departs from standard approaches by not relying on structural firm links and instead applies a data-driven method to detect predictive stock pairs. The study’s demonstration of exploitable trading strategies based on leader-follower signals challenges the efficient market hypothesis and opens pathways for quantitative investors to develop new models.

Reference