Autoregressive Models

AR Processes, Difference Equations & Spectral Analysis

2-3 hours

Advanced

Time Series

Essential Definitions

Core concepts and mathematical foundations

Autoregressive Model AR(p)

Formal Definition

A stochastic process $\{X_t\}$ is an autoregressive process of order $p$ if it satisfies:

X_t = \sum_{j=1}^{p} a_j X_{t-j} + \epsilon_t, \quad t \in \mathbb{Z}

where $\{\epsilon_t\} \sim WN(0,\sigma^2)$ is white noise and $a_1, \ldots, a_p$ are real coefficients with $a_p \neq 0$ .

Interpretation

The current value is a linear combination of past values plus an innovation term. This models persistence and momentum in time series data.

Key Property

AR models exhibit short memory: influence of past values decays exponentially over time, controlled by the characteristic roots.

Lag Operator (B)

Operator Definition

The lag operator (backshift operator) $B$ is defined by:

B X_t = X_{t-1}

B^n X_t = X_{t-n}

Properties

Linearity: $B(aX_t + bY_t) = aB X_t + bB Y_t$
Constants: $B c = c$ for any constant $c$
Composition: $B^n B^m = B^{n+m}$
Polynomial Action: For $\psi(z) = \sum c_j z^j$ , $\psi(B)X_t = \sum c_j X_{t-j}$

AR(p) in Operator Form

Using the lag operator, AR(p) becomes:

A(B) X_t = \epsilon_t

where $A(z) = 1 - \sum_{j=1}^p a_j z^j$ is the characteristic polynomial.

Stationarity Condition (Minimum Phase)

Critical Requirement

An AR(p) process is weakly stationary if and only if:

\text{All roots of } A(z) = 0 \text{ satisfy } |z| > 1

Equivalently, all roots lie outside the unit circle in the complex plane.

|z| > 1

Stationary: Solutions decay exponentially to zero, ensuring bounded variance.

|z| = 1

Unit Root: Persistent oscillations, non-stationary (requires differencing).

|z| < 1

Explosive: Solutions grow exponentially, variance → ∞.

Key Theorems & Proofs

Rigorous mathematical foundations with step-by-step derivations

Theorem 1: AR(p) Stationarity Criterion

Necessary and sufficient condition for weak stationarity

Theorem Statement:

The AR(p) process $X_t = \sum_{j=1}^p a_j X_{t-j} + \epsilon_t$ is weakly stationary if and only if all roots of the characteristic equation $A(z) = 0$ satisfy $|z| > 1$ .

Proof:

Setup: Consider the homogeneous difference equation $A(B)X_t = 0$ . The characteristic equation is $1 - \sum_{j=1}^p a_j z^j = 0$ .
Let $z_1, \ldots, z_k$ denote the distinct roots with multiplicities $r(1), \ldots, r(k)$ where $\sum r(j) = p$ .
General solution form: The solution to the homogeneous equation is:
$X_t^{(h)} = \sum_{j=1}^k \sum_{m=0}^{r(j)-1} C_{j,m} t^m z_j^{-t}$
where $C_{j,m}$ are constants determined by initial conditions.
Stationarity requires convergence: For the particular solution to exist and be stationary, we need:
$\lim_{t \to \infty} X_t^{(h)} = 0$
This ensures the process doesn't explode or maintain unit root oscillations.
Convergence condition: The term $t^m z_j^{-t} \to 0$ as $t \to \infty$ if and only if $|z_j^{-1}| < 1$ , i.e., $|z_j| > 1$ .
Polynomial growth $t^m$ is dominated by exponential decay $|z_j|^{-t}$ when $|z_j| > 1$ .
Wold representation existence: When all $|z_j| > 1$ , we can invert the operator:
$X_t = A(B)^{-1} \epsilon_t = \sum_{j=0}^{\infty} \psi_j \epsilon_{t-j}$
with $\sum |\psi_j| < \infty$ (absolutely summable), guaranteeing finite variance.
Conclusion: The process is weakly stationary with:
$E[X_t] = 0, \quad \text{Var}(X_t) = \sigma^2 \sum_{j=0}^{\infty} \psi_j^2 < \infty$

Practical Implication:

Before fitting an AR model, verify that estimated characteristic roots lie outside the unit circle. If not, the data likely requires differencing (ARIMA modeling) or has structural breaks.

Theorem 2: Yule-Walker Equations

Autocovariance structure and parameter estimation

Theorem Statement:

For a stationary AR(p) process, the autocovariance function satisfies:

\gamma_k = \sum_{j=1}^p a_j \gamma_{k-j}, \quad k \geq 1

with variance:

\sigma^2 = \gamma_0 - \sum_{j=1}^p a_j \gamma_j

Proof:

Start from AR definition: Multiply both sides by $X_{t-k}$ :
$X_t X_{t-k} = \sum_{j=1}^p a_j X_{t-j} X_{t-k} + \epsilon_t X_{t-k}$
Take expectations: Using linearity of expectation:
$E[X_t X_{t-k}] = \sum_{j=1}^p a_j E[X_{t-j} X_{t-k}] + E[\epsilon_t X_{t-k}]$
Apply white noise orthogonality: For $k \geq 1$ , $\epsilon_t$ is uncorrelated with $X_{t-k}$ :
$E[\epsilon_t X_{t-k}] = E[\epsilon_t] E[X_{t-k}] = 0$
since $\epsilon_t$ is independent of all past $X$ values.
Substitute autocovariance: By definition $\gamma_k = E[X_t X_{t-k}]$ :
$\gamma_k = \sum_{j=1}^p a_j \gamma_{k-j}, \quad k \geq 1$
Variance equation (k=0 case): When $k=0$ , $E[\epsilon_t X_t] \neq 0$ :
$\gamma_0 = \sum_{j=1}^p a_j \gamma_j + E[\epsilon_t X_t]$
Since $X_t = \sum a_j X_{t-j} + \epsilon_t$ , we have $E[\epsilon_t X_t] = \sigma^2$ .
Final variance formula: Rearranging:
$\sigma^2 = \gamma_0 - \sum_{j=1}^p a_j \gamma_j$

Matrix Formulation

For $k = 1, \ldots, p$ , the Yule-Walker system in matrix form:

\begin{bmatrix} \gamma_0 & \gamma_1 & \cdots & \gamma_{p-1} \\ \gamma_1 & \gamma_0 & \cdots & \gamma_{p-2} \\ \vdots & \vdots & \ddots & \vdots \\ \gamma_{p-1} & \gamma_{p-2} & \cdots & \gamma_0 \end{bmatrix} \begin{bmatrix} a_1 \\ a_2 \\ \vdots \\ a_p \end{bmatrix} = \begin{bmatrix} \gamma_1 \\ \gamma_2 \\ \vdots \\ \gamma_p \end{bmatrix}

The coefficient matrix $\Gamma_p$ is a Toeplitz matrix (constant diagonals), guaranteed positive definite for stationary processes.

Theorem 3: PACF Truncation Property

Diagnostic signature for AR model order

Theorem Statement:

For an AR(p) process, the partial autocorrelation function (PACF) satisfies:

\phi_{kk} = \begin{cases} a_k & k \leq p \\ 0 & k > p \end{cases}

Proof Sketch:

PACF definition: $\phi_{kk}$ is the last coefficient in the regression of $X_t$ on $X_{t-1}, \ldots, X_{t-k}$ .
For k ≤ p: The optimal k-step predictor includes lags 1 through k, giving $\phi_{kk} = a_k$ .
For k > p: Lags beyond p add no predictive power (already captured by lags 1 to p), thus $\phi_{kk} = 0$ .

Model Identification Rule:

Plot sample PACF with 95% confidence bands $\pm 1.96/\sqrt{n}$ . The order p is identified where PACF last exceeds the bands before remaining within them.

Worked Examples

Step-by-step solutions for model analysis and parameter estimation

Example 1: Complete Analysis of AR(1) Process

Fundamental

Deriving properties for

X_t = 0.8 X_{t-1} + \epsilon_t

Problem:

Consider the process $X_t = 0.8 X_{t-1} + \epsilon_t$ where $\epsilon_t \sim WN(0, 1)$ . Determine stationarity, calculate the mean, variance, and the first 3 values of the ACF. Find the spectral density.

1
Stationarity Check

The characteristic polynomial is $A(z) = 1 - 0.8z$ .

Root: $1 - 0.8z = 0 \implies z = 1/0.8 = 1.25$ .

Since $|z| = 1.25 > 1$ , the root lies outside the unit circle.

The process is weakly stationary.

2
Moments Calculation

Mean

E[X_t] = 0.8 E[X_{t-1}] + E[\epsilon_t]

\mu = 0.8 \mu + 0 \implies 0.2\mu = 0 \implies \mu = 0

Variance

\text{Var}(X_t) = 0.8^2 \text{Var}(X_{t-1}) + \text{Var}(\epsilon_t)

\gamma_0 = 0.64 \gamma_0 + 1 \implies 0.36 \gamma_0 = 1

\gamma_0 = 1/0.36 \approx 2.778

3
Autocorrelation Function (ACF)

For AR(1), $\rho_k = \phi^k = 0.8^k$ .

Lag 1

\rho_1 = 0.8^1 = 0.8

Lag 2

\rho_2 = 0.8^2 = 0.64

Lag 3

\rho_3 = 0.8^3 = 0.512

Note the geometric decay, characteristic of AR(1) processes.

4
Spectral Density

Using the formula $f(\lambda) = \frac{\sigma^2}{2\pi} \frac{1}{|1 - \phi e^{-i\lambda}|^2}$ :

f(\lambda) = \frac{1}{2\pi} \frac{1}{|1 - 0.8 e^{-i\lambda}|^2} = \frac{1}{2\pi(1 + 0.8^2 - 2(0.8)\cos\lambda)}

f(\lambda) = \frac{1}{2\pi(1.64 - 1.6\cos\lambda)}

Analysis: The denominator is minimized when $\cos\lambda = 1$ (i.e., $\lambda = 0$ ), maximizing $f(\lambda)$ . This indicates a "low-pass" filter effect, dominating low frequencies (smooth trends).

Example 2: AR(2) with Real Roots

Intermediate

Analyzing

X_t = 0.7X_{t-1} - 0.1X_{t-2} + \epsilon_t

Problem:

For the model $X_t - 0.7X_{t-1} + 0.1X_{t-2} = \epsilon_t$ with $\sigma^2=1$ , find the characteristic roots, check stationarity, and calculate $\gamma_0, \gamma_1$ .

1. Roots Calculation

Characteristic equation: $1 - 0.7z + 0.1z^2 = 0$

Using quadratic formula for $0.1z^2 - 0.7z + 1 = 0$ :

z = \frac{0.7 \pm \sqrt{0.49 - 0.4}}{0.2} = \frac{0.7 \pm 0.3}{0.2}

z_1 = 1.0/0.2 = 5

z_2 = 0.4/0.2 = 2

Both $|z_1| > 1$ and $|z_2| > 1$ . The process is stationary.

2. Yule-Walker System

The Yule-Walker equations for AR(2):

\gamma_0 = 0.7\gamma_1 - 0.1\gamma_2 + \sigma^2

\gamma_1 = 0.7\gamma_0 - 0.1\gamma_1

\gamma_2 = 0.7\gamma_1 - 0.1\gamma_0

3. Solving for Autocovariances

From eq (2): $\gamma_1(1 + 0.1) = 0.7\gamma_0 \implies 1.1\gamma_1 = 0.7\gamma_0 \implies \rho_1 = \frac{0.7}{1.1} \approx 0.636$

Substitute $\gamma_1$ and $\gamma_2$ into eq (1):

\gamma_0 = 0.7(0.636\gamma_0) - 0.1(0.7(0.636\gamma_0) - 0.1\gamma_0) + 1

Solving this linear system yields:

\gamma_0 \approx 1.88, \quad \gamma_1 \approx 1.20

Example 3: AR(2) with Complex Roots (Pseudo-Cyclic)

Advanced

Analyzing

X_t = 1.5X_{t-1} - 0.75X_{t-2} + \epsilon_t

Problem:

Investigate the behavior of $X_t = 1.5X_{t-1} - 0.75X_{t-2} + \epsilon_t$ . Determine the period of the pseudo-cycles.

1. Characteristic Roots

1 - 1.5z + 0.75z^2 = 0

Discriminant $\Delta = 2.25 - 3 = -0.75 < 0$ . Roots are complex.

z = \frac{1.5 \pm i\sqrt{0.75}}{1.5} = 1 \pm i\frac{\sqrt{3}}{3}

Modulus $|z| = \sqrt{1 + 1/3} = \sqrt{4/3} \approx 1.15 > 1$ . Stationary.

2. Cycle Analysis

The ACF exhibits damped cosine waves: $\rho_k \approx A r^k \cos(k\lambda_0 + \phi)$ .

Frequency $\lambda_0$ is determined by the argument of the roots $z^{-1}$ .

\cos \lambda_0 = \frac{a_1}{2\sqrt{-a_2}} = \frac{1.5}{2\sqrt{0.75}} = \frac{1.5}{2(0.866)} \approx 0.866

\lambda_0 = \arccos(0.866) = \pi/6

Period: $P = 2\pi / \lambda_0 = 12$ .

The process exhibits pseudo-cycles of length 12.

Example 4: Yule-Walker Parameter Estimation

Practical

Estimating AR(2) coefficients from sample data

Problem:

A time series yields sample autocorrelations $\hat{\rho}_1 = 0.8$ and $\hat{\rho}_2 = 0.5$ . Estimate the parameters $\hat{a}_1, \hat{a}_2$ for an AR(2) model.

1. Setup Yule-Walker System

For AR(2), the normalized Yule-Walker equations (using correlations) are:

\begin{cases} \rho_1 = a_1 + a_2 \rho_1 \\ \rho_2 = a_1 \rho_1 + a_2 \end{cases}

2. Substitute Sample Values

\begin{cases} 0.8 = \hat{a}_1 + 0.8 \hat{a}_2 \\ 0.5 = 0.8 \hat{a}_1 + \hat{a}_2 \end{cases}

3. Solve Linear System

From eq 1: $\hat{a}_1 = 0.8 - 0.8\hat{a}_2$ . Substitute into eq 2:

0.5 = 0.8(0.8 - 0.8\hat{a}_2) + \hat{a}_2

0.5 = 0.64 - 0.64\hat{a}_2 + \hat{a}_2

0.5 - 0.64 = 0.36\hat{a}_2 \implies -0.14 = 0.36\hat{a}_2

\hat{a}_2 = -0.14/0.36 \approx -0.389

Now find $\hat{a}_1$ :

\hat{a}_1 = 0.8 - 0.8(-0.389) = 0.8 + 0.311 = 1.111

Estimated Model:

X_t = 1.11 X_{t-1} - 0.39 X_{t-2} + \epsilon_t

Check stationarity: $\hat{a}_1 + \hat{a}_2 = 0.72 < 1$ , $\hat{a}_2 - \hat{a}_1 = -1.5 < 1$ , $|\hat{a}_2| < 1$ . Valid stationary model.

Spectral Density Analysis

Understanding AR processes in the frequency domain

The Spectral Density Function

The spectral density $f(\lambda)$ describes how the variance of the time series is distributed across different frequencies $\lambda \in [-\pi, \pi]$ .

General Formula for AR(p):

f(\lambda) = \frac{\sigma^2}{2\pi} \frac{1}{|A(e^{-i\lambda})|^2} = \frac{\sigma^2}{2\pi} \frac{1}{|1 - \sum_{j=1}^p a_j e^{-ij\lambda}|^2}

Key Insights:

Peaks: Occur at frequencies where $|A(e^{-i\lambda})|$ is small (roots close to unit circle).
Low Frequency: Positive AR coefficients (e.g., $a_1 > 0$ ) boost low frequencies (trends).
High Frequency: Negative AR coefficients (e.g., $a_1 < 0$ ) boost high frequencies (rapid oscillation).

Frequency Domain Signatures

AR(1) Spectra

Positive φ: Peak at $\lambda=0$ . Looks like "red noise" (smooth).
Negative φ: Peak at $\lambda=\pi$ . Looks like "blue noise" (jagged).

AR(2) Pseudo-Cyclic Spectra

With complex roots $re^{\pm i\lambda_0}$ , the spectrum has a peak near $\lambda_0$ .
The closer $r$ is to 1, the sharper the peak (stronger periodicity).

"The spectral density is the Fourier transform of the autocovariance function. They contain equivalent information but offer different perspectives."

Model Diagnostics & Validation

Ensuring model adequacy through residual analysis and order selection

Order Selection Criteria

PACF Method

Look for the lag $p$ where PACF cuts off.
Rule: $|\phi_{kk}| > 1.96/\sqrt{n}$ for $k \le p$ and $\approx 0$ for $k > p$ .

AIC (Akaike)

AIC(p) = \ln \hat{\sigma}_p^2 + \frac{2p}{n}

Penalizes complexity. Good for prediction, tends to overestimate order slightly.

BIC (Bayesian)

BIC(p) = \ln \hat{\sigma}_p^2 + \frac{p \ln n}{n}

Stronger penalty. Consistent estimator of true order $p$ as $n \to \infty$ .

Residual Whiteness Test

If the model is adequate, residuals $\hat{\epsilon}_t$ should behave like white noise.

ACF of Residuals: Should be within bounds $\pm 1.96/\sqrt{n}$ for all lags $k > 0$ .
Ljung-Box Test: $Q = n(n+2)\sum_{k=1}^m \frac{\hat{\rho}_k^2(\hat{\epsilon})}{n-k}$ .

Common Pitfalls

Overfitting: Including too many lags increases variance of estimates.
Underfitting: Missing lags leaves serial correlation in residuals (biased estimates).
Unit Roots: Applying AR to non-stationary data yields spurious results. Always difference first if needed.

Advanced Extensions

From heteroskedasticity to multivariate systems and deep learning

ARCH & GARCH

Standard AR models assume constant variance ( $\sigma^2$ ). Financial data often exhibits volatility clustering.

GARCH(1,1) Model:

\sigma_t^2 = \alpha_0 + \alpha_1 \epsilon_{t-1}^2 + \beta_1 \sigma_{t-1}^2

Used extensively in risk management (Value-at-Risk) and option pricing.

Vector AR (VAR)

Extends AR to $k$ variables. $\mathbf{X}_t$ is a vector, $\mathbf{A}_j$ are matrices.

VAR(p) Model:

\mathbf{X}_t = \sum_{j=1}^p \mathbf{A}_j \mathbf{X}_{t-j} + \mathbf{\epsilon}_t

Captures feedback loops between variables (e.g., GDP, Inflation, Interest Rates).

Deep Autoregression

Modern approaches replace linear combinations with non-linear neural networks.

Neural AR:

X_t = f_{NN}(X_{t-1}, \ldots, X_{t-p}) + \epsilon_t

Includes RNNs, LSTMs, and Transformers (e.g., Temporal Fusion Transformer).

Levinson-Durbin Recursion

Efficient recursive algorithm for solving Yule-Walker equations

Computational Efficiency

Solving the Yule-Walker system $\Gamma_p \mathbf{a} = \mathbf{\gamma}_p$ directly via matrix inversion (Gaussian elimination) requires $O(p^3)$ operations.

The Levinson Advantage:

By exploiting the Toeplitz structure of $\Gamma_p$ , the Levinson-Durbin algorithm reduces complexity to $O(p^2)$ .

This is crucial for fitting high-order AR models (e.g., in speech processing or geophysical signal analysis).

The Algorithm

Initialization:

\phi_{11} = \rho_1, \quad \sigma_1^2 = \gamma_0(1 - \phi_{11}^2)

Recursion (for k = 2 to p):

\phi_{kk} = \frac{\rho_k - \sum_{j=1}^{k-1} \phi_{k-1,j} \rho_{k-j}}{1 - \sum_{j=1}^{k-1} \phi_{k-1,j} \rho_j}

\phi_{kj} = \phi_{k-1,j} - \phi_{kk} \phi_{k-1,k-j}

\sigma_k^2 = \sigma_{k-1}^2 (1 - \phi_{kk}^2)

Note: $\phi_{kk}$ is the partial autocorrelation at lag k.

Forecasting with AR Models

Optimal linear prediction and mean squared error analysis

One-Step Ahead Prediction

The optimal predictor (minimizing MSE) of $X_{n+1}$ given history $\mathcal{F}_n$ is the conditional expectation:

\hat{X}_{n+1} = E[X_{n+1} | X_n, X_{n-1}, \ldots] = \sum_{j=1}^p a_j X_{n+1-j}

The prediction error is simply the white noise term $\epsilon_{n+1}$ , with variance $\sigma^2$ .

Multi-Step Ahead Prediction

For $h$ -step ahead forecast $\hat{X}_{n+h}$ , we iterate the AR equation, replacing future values with their forecasts:

\hat{X}_{n+h} = a_1 \hat{X}_{n+h-1} + \cdots + a_p \hat{X}_{n+h-p}

where $\hat{X}_{n+k} = X_{n+k}$ if $k \le 0$ (observed values).

Forecast Error Variance

The error $e_n(h) = X_{n+h} - \hat{X}_{n+h}$ can be written using the MA( $\infty$ ) representation:

e_n(h) = \sum_{j=0}^{h-1} \psi_j \epsilon_{n+h-j}

MSE: $\sigma^2(h) = \sigma^2 \sum_{j=0}^{h-1} \psi_j^2$

Convergence

As $h \to \infty$ , the forecast converges to the process mean (0), and the variance converges to the process variance $\gamma_0$ . This reflects the "short memory" of stationary AR processes.

Practice Quiz

Questions

Correct

Accuracy

What is the general form of an AR(p) model?

Not attempted

For an AR(1) process

X_t = \phi X_{t-1} + \varepsilon_t

to be stationary, what condition must

\phi

satisfy?

Not attempted

What do the Yule-Walker equations allow us to estimate?

Not attempted

For an AR(p) process, what is the behavior of the PACF?

Not attempted

What is the characteristic polynomial of an AR(2) process

X_t = \phi_1 X_{t-1} + \phi_2 X_{t-2} + \varepsilon_t

Not attempted

For AR(1) with

\phi = 0.5

, what is the autocorrelation at lag k?

Not attempted

What is the variance of an AR(1) process

X_t = \phi X_{t-1} + \varepsilon_t

with

|\phi| < 1

Not attempted

What is the lag operator B defined as?

Not attempted

Can an AR(p) process be represented as an infinite MA process?

Not attempted

What is the relationship between AIC and model complexity?

Not attempted

Frequently Asked Questions

Why must AR characteristic roots lie outside the unit circle?

Roots outside the unit circle (> 1 in absolute value) ensure exponential decay of the homogeneous solution, guaranteeing stationarity. If a root lies on or inside the unit circle, the process becomes non-stationary with unbounded variance.

What is the intuition behind the Yule-Walker equations?

They express a fundamental consistency relationship: in a stationary AR process, the autocovariance structure must satisfy the same linear relationship as the process itself. This allows us to estimate AR coefficients directly from sample autocovariances.

How does the PACF help identify AR model order?

For an AR(p) process, the PACF truncates after lag p—i.e., φ_kk = 0 for k > p. This diagnostic property allows model order selection: the last significant PACF lag indicates the appropriate AR order. This contrasts with the ACF, which decays gradually.

Can an AR(p) model be represented as an MA(∞) process?

Yes, via the Wold decomposition. When all characteristic roots lie outside the unit circle, we can invert A(B) to get X_t = A⁻¹(B)ε_t = ∑ψ_jε_{t-j}. The ψ_j coefficients decay exponentially, ensuring the infinite sum converges.

What makes the Levinson-Durbin algorithm efficient?

It exploits the Toeplitz structure of the autocovariance matrix, reducing computation from O(p³) to O(p²). The algorithm recursively builds solutions for orders 1 through p, computing PACF coefficients as byproducts—essential for large-scale time series analysis.

What is the difference between ACF and PACF?

ACF measures total correlation (direct + indirect) at each lag. PACF measures only the direct correlation after removing effects of intermediate lags. For AR(p): ACF decays, PACF cuts off after lag p. This pattern is crucial for model identification.

Autoregressive Models

Essential Definitions

Formal Definition

Interpretation

Key Property

Operator Definition

Properties

AR(p) in Operator Form

Critical Requirement

|z| > 1

|z| = 1

|z| < 1

Key Theorems & Proofs

Matrix Formulation

Worked Examples

1Stationarity Check

2Moments Calculation

3Autocorrelation Function (ACF)

4Spectral Density

1. Roots Calculation

2. Yule-Walker System

3. Solving for Autocovariances

1. Characteristic Roots

2. Cycle Analysis

1. Setup Yule-Walker System

2. Substitute Sample Values

3. Solve Linear System

Spectral Density Analysis

Key Insights:

AR(1) Spectra

AR(2) Pseudo-Cyclic Spectra

Model Diagnostics & Validation

PACF Method

AIC (Akaike)

BIC (Bayesian)

Advanced Extensions

Levinson-Durbin Recursion

Forecasting with AR Models

Forecast Error Variance

Convergence

Frequently Asked Questions

1
Stationarity Check

2
Moments Calculation

3
Autocorrelation Function (ACF)

4
Spectral Density