Working Papers

Predictive Factor Model for Jump Intensities

• with Yi Ding, Yingying Li, and Xinghua Zheng

(Abstract) In this paper, we propose a factor model for stock jump intensities, which we call the Jump Intensity Factor Model (JIFM). The model contains a latent jump process to capture the complete grids of potential jumps, and thinning processes for individual jumps. The latent jump process can have time series dependence, which leads to predictability. We develop estimators of the parameters of the latent jump process (factor) and the thinning processes (factor loadings), and establish their asymptotic normality. Empirically, our model leads to more accurate jump intensity predictions than state-of-the-art methods. Furthermore, we show that the jump intensities of stocks have implications on asset pricing, leading to abnormal returns when single sorted or double sorted with volatilities. JIFM is, to our knowledge, the first factor model to capture the large cross-sectional dependency of stock jump intensities. It is both parsimonious and enhances predictive power, offering new insights into the financial market.

Efficient High-Dimensional Covariance Matrix Estimation Incorporating Trading Information  

• with Dachuan Chen, Yingying Li, and Xinghua Zheng

 (Abstract) This paper proposes the first trading information incorporated estimation methodology for high-dimensional covariance matrix using high-frequency data. Our method extends the univariate trading information incorporated variance estimator of Li et al. (2016) to the high-dimensional setting, allowing the cross-sectional dimension d to grow exponentially in n^{1-\varepsilon} for any \varepsilon∈ (0,1), where n is the intraday observation frequency. Theoretically, under mild assumptions, we establish the tail property of the trading information incorporated covariance estimator. We then impose a factor structure on cross-sectional intraday returns and apply POET to estimate high-dimensional covariance and precision matrices. We demonstrate the effectiveness of the proposed estimators through simulations and an empirical study using second-by-second Trade and Quote data from S&P 500 index constituents, showing that the minimum variance portfolio constructed with our estimator achieves lower long-term risk than that relying on the pre-averaging-based estimator.

Improving High-Frequency Volatility Forecasting: Integrating Machine Learning Models with Intraday Periodic Patterns and Firm-Specific News

• with Bo Zhou and Yingying Li

 (Abstract) This paper introduces an innovative framework for high-frequency volatility forecasting by synergistically integrating intraday periodic patterns, advanced machine learning (ML) models, and firm-specific news. Utilizing one-minute price data and real-time news feeds from DJ30 index constituents, we benchmark the forecasting performance of Neural Network (NN) and Gated Recurrent Unit (GRU) models against the traditional Heterogeneous Autoregressive (HAR) model. Our approach flexibly incorporates both intraday patterns and overnight news to enhance predictive accuracy. Empirical results reveal that the NN reduces prediction error by 25% and the GRU by 10% relative to HAR. Moreover, integrating intraday patterns delivers an additional average error reduction of 8%, while overnight news contributes a further 7% reduction. Notably, the predictive power of news is strongest near market open and declines towards market close. Overall, our findings demonstrate that combining ML techniques with structural intraday market patterns and firm-specific news significantly improves high-frequency volatility forecasting.