Much recent research in finance has found empirical evidence of jumps in equity returns. Their presence has been successfully used to better explain various market phenomena. Nevertheless, the role of real-time information for predicting jumps in stock markets has not been thoroughly investigated in the literature. In this article, I analyze the predictability of jumps in individual stock returns, using both macroeconomic and firm-specific news releases and I present how the information is reflected in stock prices as jumps. This analysis naturally allows a novel decomposition of individual stock jumps into systematic and idiosyncratic jumps.
To accomplish this goal, I identify important jump predictors and assess their relative importance and precision for the purpose of developing stochastic jump intensity models. Assuming that an individual equity price follows a jump diffusion process with stochastic jump intensity, I must resolve the econometric problem of identifying jump predictors using discrete data from continuous-time models. I refer to this as the mixed unobservability problem. It arises from the simultaneous presence of two unobservability problems. The first is caused by the difficulty we usually face when making an inference for a continuous-time jump counting process (without diffusion) using discrete observations. The second problem results from the presence of the diffusion process. The mixture of these two makes jumps in jump diffusion models unobservable; thus, the identification of jump predictors becomes difficult.
As a resolution, I propose an inference technique called the Jump Predictor Test (JPT). It allows us to estimate a regression-type jump intensity model and apply standard hypothesis tests in order to identify significant jump predictors. In this way, we can predict ex ante whether jumps are likely to occur, what kind of jumps are more likely to occur, and when they are more likely to occur, given the available information. The idea underlying this JPT is simple. I first detect the location of jumps from the return series by multiple nonparametric jump detection tests.
This is a necessary step before applying the JPT. Then, I suggest a likelihood inference for the JPT using time-series data for both jumps and information covariates. I prove that this technique asymptotically makes the effect of the mixed unobservability problem negligible, allowing good jump predictors to be identified. I discuss a theory of likelihood inference justifying this approach and provide a guide for tests and general applications.
