Abstract:Most of the metrics used for detecting a causal relationship among multiple time series ignore the effects of practical measurement impairments, such as finite sample effects, undersampling and measurement noise. It has been shown that these effects significantly impair the performance of the underlying causality test. In this paper, we consider the problem of sequentially detecting the causal relationship between two time series while accounting for these measurement impairments. In this context, we first formulate the problem of Granger causality detection as a binary hypothesis test using the norm of the estimates of the vector auto-regressive~(VAR) coefficients of the two time series as the test statistic. Following this, we investigate sequential estimation of these coefficients and formulate a sequential test for detecting the causal relationship between two time series. Finally via detailed simulations, we validate our derived results, and evaluate the performance of the proposed causality detectors.
What problem does this paper attempt to address?
This paper aims to solve the problem that when detecting the causal relationships between multiple time series, existing methods usually ignore some important influencing factors in actual measurements. Specifically, these influencing factors include the finite - sample effect, undersampling, and measurement noise. These problems significantly affect the performance of causal relationship detection. For this reason, the paper proposes a new method to detect the causal relationship between two time series through sequential hypothesis testing and takes into account the above - mentioned influencing factors in the measurement.
### Main contributions:
1. **Proposal of test statistic**: Use the norm of the vector autoregressive (VAR) coefficient estimates of two time series as the test statistic for detecting causal relationships, and analyze the detection probability and false alarm probability of this test statistic under different sample sizes.
2. **Sequential detection algorithm**: Develop a sequential detection method for detecting the causal relationship between two time series by using the recursive least squares (RLS) algorithm.
3. **Performance evaluation**: Verify the performance of the proposed VAR - coefficient - based test statistic through detailed simulations and actual data tests, evaluate the performance of the sequential detection algorithm, and give its optimal parameter configuration.
### Problems solved:
- **Finite - sample effect**: Existing methods usually assume that the exact second - order statistics are known, but in practical applications, due to the limited number of samples, these statistics can only be estimated by a finite sample, resulting in a decline in detection performance.
- **Measurement noise**: There is noise in actual measurements, which will affect the accuracy of causal relationship detection.
- **Requirement for real - time detection**: Existing methods require all samples to calculate the test statistic, which is not suitable for real - time detection scenarios. The sequential detection method proposed in the paper can process newly arrived samples in real time, improving the feasibility of real - time detection.
### Formula examples:
- **VAR model**:
\[
\begin{aligned}
u[n] &= \sum_{k = 1}^{K}a_{uu,k}u[n - k]+\sum_{k = 1}^{K}a_{uv,k}v[n - k]+\eta_u[n] \\
v[n] &= \sum_{k = 1}^{K}a_{vu,k}u[n - k]+\sum_{k = 1}^{K}a_{vv,k}v[n - k]+\eta_v[n]
\end{aligned}
\]
- **Test statistic**:
\[
T_N=\frac{N - K}{\sigma^2_\phi}\left\|\Sigma^{- 1/2}\hat{w}_y\right\|^2
\]
- **ROC curve**:
\[
T_N\sim\begin{cases}
\chi^2_{2K}(0)&\text{H}_0 \\
\chi^2_{2K}(\kappa)&\text{H}_1
\end{cases}
\]
where \(\kappa=\frac{N - K}{\sigma^2_\phi}w_y^H\Sigma^{-1}w_y\).
Through these methods and formulas, the paper effectively solves the challenges of detecting the causal relationships of time series under actual measurement conditions.