The leverage effect puzzle: Disentangling sources of bias at high frequency

Yacine Aı̈t-Sahalia, Jianqing Fan, Yingying LiJournal of Financial Economics 2013 · DOI

TL;DR

This paper tackles the "leverage effect puzzle," where the expected negative correlation between asset returns and volatility disappears at high frequencies. The authors identify discretization, smoothing, estimation, and microstructure noise as key sources of bias, developing a novel bias correction method that successfully reveals the leverage effect in high-frequency data.

What Problem Does This Paper Address?

The "leverage effect" is a well-documented phenomenon in finance: when asset prices fall, their volatility tends to increase, and vice versa. This is intuitively linked to increased financial leverage when a company's stock price drops. However, when researchers try to measure this effect using high-frequency trading data (data sampled very rapidly), the expected negative correlation often vanishes, yielding results close to zero. This is the "leverage effect puzzle." The paper aims to understand why this puzzle exists by dissecting the various sources of error that arise when estimating volatility from high-frequency data and proposes a method to overcome these challenges.

Methods

The authors employ a multi-pronged approach:

  • Stochastic Volatility Model: They use the Heston (1993) stochastic volatility model as a theoretical framework. This model allows them to derive explicit expressions for biases under different assumptions.
  • Discretization and Smoothing: They analyze how the discrete nature of observed data (discretization) and the process of estimating spot volatility by smoothing over time (smoothing) introduce biases.
  • Noise-Robust Estimators: They investigate the impact of market microstructure noise (errors introduced by the trading process itself) and use noise-robust estimators like TSRV (Two Scales Realized Volatility) and PAV (Pre-Averaging) to mitigate these issues.
  • Bias Decomposition: The core of their method is to systematically decompose the total bias into its constituent parts: discretization, smoothing, estimation error, and noise correction.
  • Bias Correction Strategy: They propose a novel, model-independent bias correction strategy. This involves:
  • Computing preliminary bias-corrected estimates.
  • Running a simple linear regression on these preliminary estimates across different time horizons.
  • Taking the intercept of this regression as the final, bias-corrected estimate of the leverage effect parameter.

Datasets

The authors use high-frequency data for:

  • S&P 500 futures: From January 2004 to December 2007.
  • Microsoft (MSFT) returns: From January 2005 to June 2007.

They also utilize simulated data from the Heston model to verify their theoretical findings.

Key Insights and Ideas

  • The "Puzzle" is Real and Multi-faceted: The vanishing leverage effect at high frequencies isn't due to a single factor but a combination of biases that accumulate.
  • Discretization and Smoothing are Major Culprits: Even in idealized scenarios without microstructure noise, estimating volatility from discrete data and smoothing it introduces significant bias.
  • Noise-Robustness Can Worsen the Problem (Paradoxically): While noise-robust estimators are better for volatility estimation in general, they can sometimes exacerbate the bias in estimating the leverage effect itself.
  • A Linear Regression Approach is Key: Aggregating preliminary bias-corrected estimates over a range of time horizons via linear regression proves to be a robust way to extract the true leverage effect.
  • The Leverage Effect Does Exist at High Frequencies: The paper demonstrates that with the right correction, the expected negative correlation between returns and volatility can be clearly observed even in high-frequency data.

Key Results

  • The paper provides theoretical derivations for the biases introduced at each step of estimating the leverage effect.
  • Through Monte Carlo simulations, they show that their proposed bias correction method significantly reduces bias and leads to estimates that are close to the true leverage effect parameter.
  • Empirical analysis on S&P 500 and Microsoft data shows that naive estimation methods yield near-zero correlations, while their bias-corrected method reveals a strong, negative leverage effect, often around -0.7 to -0.9.
  • The effectiveness of the bias correction is demonstrated across different sampling frequencies (e.g., 1-minute, 5-second intervals).

Limitations and Future Work

  • Model Dependence (for some corrections): While the final proposed method is model-independent, some intermediate theoretical derivations rely on the Heston model.
  • Parameter Estimation: The accuracy of the bias correction can depend on estimating certain parameters (like quarticity) nonparametrically.
  • Future Work: The paper suggests that further investigation into the optimal choice of the tuning parameter 'm' (used in the regression-based correction) and exploring other data generating processes could be valuable. The authors also hint at the potential for reducing estimation errors by increasing sampling frequency.

Summary updated

BibTeX
@article{atsahalia2013leverage,
  author = {Yacine Aı̈t-Sahalia and Jianqing Fan and Yingying Li},
  title = {{The leverage effect puzzle: Disentangling sources of bias at high frequency}},
  journal = {Journal of Financial Economics},
  year = {2013},
  doi = {10.1016/j.jfineco.2013.02.018},
}
Figure