Abstract:In this paper we consider two semimartingales driven by diffusions and jumps. We allow both for finite activity and for infinite activity jump components. Given discrete observations we disentangle the {\it integrated covariation} (the covariation between the two diffusion parts, indicated by IC) from the co-jumps. This has important applications to multiple assets price modeling for forecasting, option pricing, risk and credit risk management. An approach commonly used to estimate IC is to take the sum of the cross products of the two processes increments; however this estimator can be highly biased in the presence of jumps, since it approaches the quadratic ovariation, which contains also the co-jumps. Our estimator of IC is based on a threshold (or truncation) technique. We prove that the estimator is consistent in both cases as the number of observations increases to infinity. Further, in presence of only finite activity jumps 1) \hat{IC} is also asymptotically Gaussian; 2) a joint CLT for \hat{IC} and threshold estimators of the integrated variances of the single processes allows to reach consistent and asymptotically Gaussian estimators even of the $\beta$s and of the correlation coefficient among the diffusion parts of the two processes, allowing a better measurement of their dependence; 3) thresholding gives an estimate of IC which is robust to the asynchronicity of the observations. We conduct a simulation study to check that the application of our technique is in fact informative for values of the step between the observations large enough to avoid the typical problems arising in presence of microstructure noises in the data, and to asses the choice of the threshold parameters.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is how to distinguish between the integrated covariation (i.e., the covariation between the two diffusion parts, denoted as IC) and co - jumps between two semimartingales driven by diffusion and jumps given discrete observational data. Specifically, the paper focuses on how to effectively estimate IC and separate the influence of co - jumps from the global quadratic covariation in the presence of finite - activity jumps or infinite - activity jumps. This problem is of great significance for multi - asset price modeling, prediction, option pricing, risk management and credit risk management.
### Background and Problem of the Paper
In financial econometrics, the correct selection of asset price models is an important issue, which directly affects areas such as prediction, option pricing, portfolio risk management and credit risk management. Among them, the co - jump phenomenon refers to the situation where multiple assets jump at the same time point, and this phenomenon may imply certain unexpected events in the market or the mutual influence between companies. Therefore, accurately identifying and quantifying co - jumps is crucial for understanding asset price dynamics and market behavior.
### Common Methods and Their Limitations
The method usually used to estimate IC is to take the sum of the cross - products of the increments of the two processes, but this may be highly biased in the presence of jumps. This is because when the sample size tends to infinity, such a sum will approach the global quadratic covariation, and the global quadratic covariation includes not only IC but also the influence of co - jumps. In order to accurately estimate IC, a method is needed to isolate these jumps.
### Method Proposed in the Paper
The paper proposes an estimation method based on the threshold (or truncation) technique, which allows isolating all jumps in the case of finite - activity jumps and isolating jumps exceeding the threshold in the case of infinite - activity jumps. Through this method, an asymptotically unbiased IC estimator and a co - jump estimator can be constructed. Specifically, the paper constructs the following estimator:
\[ \hat{IC}_{T,n} := \sum_{j = 1}^n \Delta_j X^{(1)} 1\{(\Delta_j X^{(1)})^2 \leq r(h)\} \Delta_j X^{(2)} 1\{(\Delta_j X^{(2)})^2 \leq r(h)\} \]
where \(h = T/n\), \(r(h)\) is a deterministic function that satisfies certain conditions, such as \(\lim_{h \to 0} r(h) = 0\) and \(\lim_{h \to 0} h \log(1/h)/r(h) = 0\). This estimator only considers those increments whose changes are less than the given threshold, thus effectively excluding the influence of jumps.
### Main Results
1. **Consistency**: The paper proves that when the sample size tends to infinity, \(\hat{IC}_{T,n}\) converges uniformly to \(IC_T\).
2. **Central Limit Theorem**: In the case of only finite - activity jumps, \(\hat{IC}_{T,n}\) is also asymptotically normally distributed, and asymptotic normal estimators of the regression coefficient \(\beta\) and the correlation coefficient can be given.
3. **Robustness**: The threshold technique makes the estimator robust to the asynchrony of observations.
### Simulation Study
The paper also conducts a simulation study to verify the performance of the proposed estimation method in practical applications. The study finds that even when the observation interval is large enough (for example, five minutes), this method can effectively avoid the influence of microstructure noise and provide reliable estimation results.
### Conclusion
In general, this paper provides a method for effectively estimating the integrated covariation and co - jumps in the presence of jumps, which is of great significance for financial market modeling and risk management.