Abstract
We propose a novel estimator of the autocorrelation function (ACF) in the presence of missing observations. We establish the consistency and asymptotic normality of this estimator, and derive deviation bounds for various classes of weakly dependent stationary time series, including both causal and non-causal models. Additionally, we introduce a modified version of the periodogram defined from these autocorrelation estimators and derive the asymptotic distribution of linear functionals of this estimator.
Introduction
In time series analysis, the Autocorrelation Function (ACF) is a fundamental tool used to measure how a signal correlates with its own past values. In an ideal setting, data points—such as financial indices or sensor readings—are recorded at perfectly regular intervals. However, real-world data is frequently plagued by missing observations due to equipment failure, human error, or irregular sampling protocols.
Standard methods for calculating the ACF typically assume a complete dataset. While previous research has attempted to extend these methods to handle missing values, such approaches often rely on the restrictive assumption that the underlying system's "noise" (innovations) is independent and identically distributed (i.i.d.). This paper addresses this gap by providing a robust mathematical framework for ACF estimation that remains valid even for complex, "weakly dependent" processes, such as those found in financial econometrics.
The Problem of Missing Data
When data is missing, the observed series can be viewed as a "censored" version of the true process. Let $X_i$ represent the underlying stationary process and $C_i$ represent a censoring process, where $C_i = 1$ if the data point is observed and $C_i = 0$ if it is missing. The resulting observed series is $Y_i = C_i X_i$.
Traditional ACF estimators struggle in this context because they do not account for the reduced number of available pairs at specific lags. Furthermore, many existing proofs for consistency in the presence of missing data fail when applied to models like ARCH or GARCH, where the variance changes over time (heteroskedasticity) and the noise is not strictly independent.
Methodology: The Parzen-Style Estimator
The authors employ a "Parzen-style" estimator to reconstruct the covariance. Instead of dividing the sum of observed products by the total sample size $n$, the proposed estimator calculates the covariance by taking the sum of observed pairs and dividing by the actual number of times those pairs were available at a given lag $\ell$.
The Weak Dependence Framework
The core innovation of this work lies in its departure from traditional "mixing conditions," which are often mathematically cumbersome and fail for simple models like AR(1). Instead, the authors utilize measures of weak dependence (specifically $\theta, \kappa, \text{ and } \lambda$). This framework allows the authors to prove that if both the underlying data ($X_i$) and the missingness process ($C_i$) are weakly dependent, the resulting observed series ($Y_i$) inherits this property, preserving the mathematical structure needed for reliable estimation.
Key Theoretical Results
The paper establishes three primary theoretical pillars for the proposed estimator:
- Asymptotic Normality (Theorem 1): The authors provide a Central Limit Theorem (CLT) for the estimator. They derive a concrete formula for the asymptotic variance $\sigma^2_\ell$, which allows researchers to determine the precision of their ACF estimates and construct confidence intervals even when data is missing.
- Almost Sure Convergence (Theorem 2): The estimator is shown to be consistent, meaning that as the sample size increases, the estimate converges to the true value. This holds under specific moment conditions, such as when the data possesses a finite $2+\delta$ moment.
- Spectral Density Estimation: Beyond the time domain, the authors introduce a modified periodogram. They prove that this periodogram is a consistent estimator for the spectral density of the process, enabling frequency-domain analysis for gapped or irregularly sampled data.
Applications and Insights
This theoretical framework is particularly relevant for non-linear time series models, including:
- Nonparametric ARCH models: Widely used in finance to describe volatility clustering.
- Bilinear models: Complex models where the current value depends on the product of past values and past noise.
- Markov Processes: Systems where the future state depends only on the current state.
The research demonstrates that independent noise is not a prerequisite for accurate ACF estimation. As long as the system's "memory" fades over time (the definition of weak dependence), the estimator remains reliable.
Limitations and Future Directions
While the paper provides a significant advancement in time series theory, it operates under two key assumptions:
- Independent Missingness: The censoring process $C_i$ must be independent of the data $X_i$. This means the data cannot be missing because of the value itself (e.g., a sensor failing only when a value exceeds a certain threshold).
- Stationarity: The underlying process is assumed to be stationary, meaning its statistical properties do not change over time.
Future research may explore extending these guarantees to non-stationary processes or scenarios where the missingness is correlated with the data values (Missing Not At Random).
Conclusion
By providing a rigorous proof of consistency and normality under weak dependence, this paper offers a versatile toolkit for researchers dealing with imperfect datasets. It bridges the gap between theoretical probability and practical application in fields like finance and signal processing, where missing observations are an unavoidable reality.
Summary updated