Stock returns, quantile autocorrelation, and volatility forecasting

We examine stock return autocorrelation at various quantiles of the returns' distribution and use it to forecast stock return volatility. Our empirical results show that the strength of the autoregression varies across the quantiles of the returns' distribution in terms of both magnitude and persistence. Specifically, the autoregression order and magnitude of the coefficients is lower in the left tail in comparison with the right tail. Additionally, we show that the quantile autoregressive (QAR) framework proposed in this study improves out-of-sample volatility forecasting performance compared to the generalised autoregressive conditional heteroscedasticity (GARCH)-type models and other quantile-based models. We also observe greater outperformance in QAR estimates during periods of financial turmoil. Moreover, the QAR method also explains the stylized 'leverage effect' associated with asset returns in the presence of volatility asymmetry.