3.3 Autocorrelation and autocovariance

Autocorrelation and autocovariance are key concepts in analyzing time series data. They measure how a process relates to itself over time, helping identify patterns, trends, and seasonality in stochastic processes.

These tools are crucial for understanding the dependence structure of a process. By examining how values correlate with past versions of themselves, we can model and forecast future behavior, making them essential in fields like finance, economics, and signal processing.

Definition of autocorrelation

• Autocorrelation measures the correlation between a time series and a lagged version of itself
• Useful for identifying patterns, trends, and seasonality in time series data
• Autocorrelation is a key concept in stochastic processes as it helps characterize the dependence structure of a process over time

Autocorrelation vs cross-correlation

Top images from around the web for Autocorrelation vs cross-correlation

• 

  Autocorrelation - wikidoc View original

  Is this image relevant?

• 

  Autocorrelation functions of materially different time series View original

  Is this image relevant?

• 

  time series - How to interpret error autocorrelation? - Cross Validated View original

  Is this image relevant?

• 

  Autocorrelation - wikidoc View original

  Is this image relevant?

• 

  Autocorrelation functions of materially different time series View original

  Is this image relevant?

1 of 3

Top images from around the web for Autocorrelation vs cross-correlation

• 

  Autocorrelation - wikidoc View original

  Is this image relevant?

• 

  Autocorrelation functions of materially different time series View original

  Is this image relevant?

• 

  time series - How to interpret error autocorrelation? - Cross Validated View original

  Is this image relevant?

• 

  Autocorrelation - wikidoc View original

  Is this image relevant?

• 

  Autocorrelation functions of materially different time series View original

  Is this image relevant?

1 of 3

• Cross-correlation measures the correlation between two different time series
• Autocorrelation is a special case of cross-correlation where the two time series are the same, but with a time lag
• Cross-correlation can identify relationships between different stochastic processes, while autocorrelation focuses on the relationship within a single process

Mathematical formulation

• For a stationary process $X_t$, the autocorrelation at lag $k$ is defined as: $\rho(k) = \frac{\text{Cov}(X_t, X_{t+k})}{\sqrt{\text{Var}(X_t)}\sqrt{\text{Var}(X_{t+k})}} = \frac{\text{Cov}(X_t, X_{t+k})}{\text{Var}(X_t)}$
• The numerator is the autocovariance at lag $k$, and the denominator is the product of the standard deviations at times $t$ and $t+k$
• For a stationary process, the variance is constant over time, simplifying the denominator to $\text{Var}(X_t)$

Interpretation of autocorrelation values

• Autocorrelation values range from -1 to 1
  • A value of 1 indicates perfect positive correlation (linear relationship) between the time series and its lagged version
  • A value of -1 indicates perfect negative correlation
  • A value of 0 indicates no linear relationship between the time series and its lagged version
• The sign of the autocorrelation indicates the direction of the relationship (positive or negative)
• The magnitude of the autocorrelation indicates the strength of the relationship

Autocorrelation function (ACF)

• The ACF is a plot of the autocorrelation values for different lags
• Provides a visual representation of the dependence structure in a time series
• Helps identify the presence and strength of autocorrelation at various lags

ACF for stationary processes

• For a stationary process, the ACF depends only on the lag and not on the absolute time
• The ACF of a stationary process is symmetric about lag 0
• The ACF of a stationary process decays to zero as the lag increases (short-term memory property)

Sample ACF

• The sample ACF is an estimate of the population ACF based on a finite sample of data
• For a time series $\{X_1, X_2, \ldots, X_n\}$, the sample autocorrelation at lag $k$ is given by: $\hat{\rho}(k) = \frac{\sum_{t=1}^{n-k}(X_t - \bar{X})(X_{t+k} - \bar{X})}{\sum_{t=1}^{n}(X_t - \bar{X})^2}$
• The sample ACF is a useful tool for identifying the presence and strength of autocorrelation in a time series

Confidence intervals for ACF

• Confidence intervals can be constructed for the sample ACF to assess the significance of autocorrelation at different lags
• Under the null hypothesis of no autocorrelation, the sample autocorrelations are approximately normally distributed with mean 0 and variance $1/n$
• An approximate 95% confidence interval for the population autocorrelation at lag $k$ is given by: $\hat{\rho}(k) \pm 1.96\sqrt{1/n}$
• Autocorrelation values outside the confidence interval are considered statistically significant

ACF for non-stationary processes

• The ACF for non-stationary processes may not have the same properties as the ACF for stationary processes
• Non-stationary processes may exhibit trending behavior or changing variance over time
• Differencing or other transformations may be needed to achieve stationarity before analyzing the ACF

Properties of autocorrelation

• Autocorrelation has several important properties that are useful in analyzing and modeling time series data

Symmetry of autocorrelation

• The autocorrelation function is symmetric about lag 0: $\rho(k) = \rho(-k)$
• This property follows from the definition of autocorrelation and the properties of covariance

Bounds on autocorrelation

• Autocorrelation values are bounded between -1 and 1: $-1 \leq \rho(k) \leq 1$
• This property follows from the Cauchy-Schwarz inequality and the definition of autocorrelation

Relationship to spectral density

• The autocorrelation function and the spectral density function are Fourier transform pairs
• The spectral density function $f(\omega)$ is the Fourier transform of the autocorrelation function $\rho(k)$: $f(\omega) = \sum_{k=-\infty}^{\infty}\rho(k)e^{-i\omega k}$
• This relationship allows for the analysis of time series data in the frequency domain

Autocovariance

• Autocovariance measures the covariance between a time series and a lagged version of itself
• Autocovariance is a key component in the calculation of autocorrelation

Definition of autocovariance

• For a stationary process $X_t$, the autocovariance at lag $k$ is defined as: $\gamma(k) = \text{Cov}(X_t, X_{t+k}) = \mathbb{E}[(X_t - \mu)(X_{t+k} - \mu)]$
• $\mu$ is the mean of the process, which is constant for a stationary process

Autocovariance vs autocorrelation

• Autocorrelation is the normalized version of autocovariance
• Autocorrelation is obtained by dividing the autocovariance by the variance of the process: $\rho(k) = \frac{\gamma(k)}{\gamma(0)}$
• Autocorrelation is dimensionless and bounded between -1 and 1, while autocovariance has the same units as the variance of the process

Autocovariance function (ACVF)

• The ACVF is a plot of the autocovariance values for different lags
• Provides information about the magnitude and direction of the dependence structure in a time series
• The ACVF is not normalized, unlike the ACF

Properties of autocovariance

• Autocovariance is symmetric about lag 0: $\gamma(k) = \gamma(-k)$
• Autocovariance at lag 0 is equal to the variance of the process: $\gamma(0) = \text{Var}(X_t)$
• For a stationary process, the autocovariance depends only on the lag and not on the absolute time

Estimating autocorrelation and autocovariance

• In practice, the true autocorrelation and autocovariance functions are unknown and must be estimated from data

Sample autocorrelation function

• The sample autocorrelation function is an estimate of the population ACF based on a finite sample of data
• For a time series $\{X_1, X_2, \ldots, X_n\}$, the sample autocorrelation at lag $k$ is given by: $\hat{\rho}(k) = \frac{\sum_{t=1}^{n-k}(X_t - \bar{X})(X_{t+k} - \bar{X})}{\sum_{t=1}^{n}(X_t - \bar{X})^2}$
• The sample ACF is a consistent estimator of the population ACF

Sample autocovariance function

• The sample autocovariance function is an estimate of the population ACVF based on a finite sample of data
• For a time series $\{X_1, X_2, \ldots, X_n\}$, the sample autocovariance at lag $k$ is given by: $\hat{\gamma}(k) = \frac{1}{n}\sum_{t=1}^{n-k}(X_t - \bar{X})(X_{t+k} - \bar{X})$
• The sample ACVF is a consistent estimator of the population ACVF

Bias and variance of estimators

• The sample ACF and ACVF are biased estimators of their population counterparts
  • The bias is typically small for large sample sizes
• The variance of the sample ACF and ACVF decreases with increasing sample size
  • Larger sample sizes lead to more precise estimates

Bartlett's formula for variance

• Bartlett's formula provides an approximation for the variance of the sample ACF under the assumption of a white noise process
• For a white noise process, the variance of the sample autocorrelation at lag $k$ is approximately: $\text{Var}(\hat{\rho}(k)) \approx \frac{1}{n}\left(1 + 2\sum_{i=1}^{k-1}\rho(i)^2\right)$
• This formula can be used to construct confidence intervals for the sample ACF

Applications of autocorrelation and autocovariance

• Autocorrelation and autocovariance are powerful tools with a wide range of applications in various fields

Time series analysis

• Autocorrelation and autocovariance are fundamental concepts in time series analysis
• They help identify patterns, trends, and seasonality in time series data
• ACF and ACVF are used to select appropriate models for time series data (AR, MA, ARMA)

Signal processing

• Autocorrelation is used to analyze the similarity of a signal with a delayed copy of itself
• It helps detect repeating patterns or periodic components in signals
• Autocorrelation is used in applications such as pitch detection, noise reduction, and echo cancellation

Econometrics and finance

• Autocorrelation is used to study the efficiency of financial markets (efficient market hypothesis)
• It helps identify trends, cycles, and volatility clustering in financial time series (stock prices, exchange rates)
• Autocorrelation is used in risk management and portfolio optimization

Quality control and process monitoring

• Autocorrelation is used to monitor the stability and control of industrial processes
• It helps detect shifts, trends, or anomalies in process variables
• Autocorrelation-based control charts (CUSUM, EWMA) are used for process monitoring and fault detection

Models with autocorrelation

• Several time series models incorporate autocorrelation to capture the dependence structure in data

Autoregressive (AR) models

• AR models express the current value of a time series as a linear combination of its past values
• The order of an AR model (denoted as AR(p)) indicates the number of lagged values included
• AR models are useful for modeling processes with short-term memory

Moving average (MA) models

• MA models express the current value of a time series as a linear combination of past error terms
• The order of an MA model (denoted as MA(q)) indicates the number of lagged error terms included
• MA models are useful for modeling processes with short-term correlation in the error terms

Autoregressive moving average (ARMA) models

• ARMA models combine AR and MA components to capture both short-term memory and error correlation
• The order of an ARMA model is denoted as ARMA(p, q), where p is the AR order and q is the MA order
• ARMA models are flexible and can model a wide range of stationary processes

Autoregressive integrated moving average (ARIMA) models

• ARIMA models extend ARMA models to handle non-stationary processes
• The "integrated" component involves differencing the time series to achieve stationarity
• The order of an ARIMA model is denoted as ARIMA(p, d, q), where d is the degree of differencing
• ARIMA models are widely used for forecasting and modeling non-stationary time series

Testing for autocorrelation

• Several statistical tests are available to assess the presence and significance of autocorrelation in time series data

Ljung-Box test

• The Ljung-Box test is a portmanteau test that assesses the overall significance of autocorrelation in a time series
• It tests the null hypothesis that the first m autocorrelations are jointly zero
• The test statistic is given by: $Q = n(n+2)\sum_{k=1}^{m}\frac{\hat{\rho}(k)^2}{n-k}$
• Under the null hypothesis, Q follows a chi-squared distribution with m degrees of freedom

Durbin-Watson test

• The Durbin-Watson test is used to detect first-order autocorrelation in the residuals of a regression model
• The test statistic is given by: $d = \frac{\sum_{t=2}^{n}(e_t - e_{t-1})^2}{\sum_{t=1}^{n}e_t^2}$
• The test statistic d ranges from 0 to 4, with values close to 2 indicating no autocorrelation
• The Durbin-Watson test is sensitive to the order of the data and the presence of lagged dependent variables

Breusch-Godfrey test

• The Breusch-Godfrey test is a more general test for autocorrelation in the residuals of a regression model
• It tests for autocorrelation of any order and is not sensitive to the order of the data
• The test involves regressing the residuals on the original regressors and lagged residuals
• The test statistic follows a chi-squared distribution under the null hypothesis of no autocorrelation

Portmanteau tests

• Portmanteau tests are a class of tests that assess the overall significance of autocorrelation in a time series
• Examples include the Box-Pierce test and the Ljung-Box test
• These tests are based on the sum of squared sample autocorrelations up to a specified lag
• Portmanteau tests are useful for identifying the presence of autocorrelation but do not provide information about specific lags