Risk Modelling and Survival Analysis · Core concepts of time series models

Box-Jenkins Model Identification, Fitting and Forecasting

Q: What are the steps of the Box-Jenkins method?

There are four: identify the model order using the ACF and PACF after making the series stationary, estimate the parameters, check the residuals for white noise, and forecast. If the checks fail, return to identification and try another model.

Q: How do I identify an ARIMA model from the ACF and PACF?

Difference until the series is stationary and count the differences as d. Then, if the PACF cuts off after lag p, use AR(p). If the ACF cuts off after lag q, use MA(q). If both decay, try ARMA models and compare them.

Q: How do I forecast k steps ahead for an AR(1)?

Use x̂_t(k) = μ + α^k (x_t − μ). The forecast moves geometrically toward the mean as k increases. The error variance rises to the process variance σ² ÷ (1 − α²).

Q: What do residual diagnostics check?

They check that the residuals behave like white noise. You look at the residual ACF, a portmanteau test, a plot against time and, where relevant, a normality check. Any remaining pattern means the model is missing structure.

Updated 11 October 2026 · Fact-checked

The Box-Jenkins method is a loop for building ARIMA models: make the series stationary, identify p and q from the sample ACF and PACF, estimate parameters, check the residuals look like white noise, then forecast. For AR(1), the k-step forecast is μ + α^k (x_t − μ).

Understand Model Identification, Fitting and Forecasting

A time series model is only useful if it matches the data. The Box-Jenkins approach gives you a repeatable way to get there. It has four stages: identification, estimation, diagnostic checking and forecasting. If the diagnostics fail, you go back and try another model.

First you need a stationary series. If the plot shows a trend, or the sample ACF decays very slowly, difference the data. The number of differences is d in ARIMA(p, d, q). Usually d is 0, 1 or 2. Over-differencing adds unnecessary MA structure, so stop as soon as the series looks stationary.

Next you identify p and q from the sample ACF and sample PACF. An AR(p) process has a PACF that cuts off after lag p and an ACF that decays. An MA(q) process has an ACF that cuts off after lag q and a PACF that decays. An ARMA(p, q) process has both decaying, so you cannot read the orders off cleanly. Then you try small models and compare them.

Then you estimate the parameters, by least squares, method of moments (Yule-Walker for AR) or maximum likelihood. Choose between competing models using a criterion such as AIC, which penalises extra parameters. Check the residuals. If the model is right, residuals behave like white noise: no significant autocorrelation, roughly constant variance, and roughly normal if normality is assumed.

Finally you forecast. The forecast is the conditional expectation of the future value given the data to date. Forecasts of a stationary model move toward the mean as the horizon grows. The forecast error variance rises toward the variance of the process.

Key rules to remember

Identification pattern for AR(p): PACF cuts off after lag p; ACF decays; Sample values are inside ±2/√n roughly when the true value is zero, where n is the sample size.
Identification pattern for MA(q): ACF cuts off after lag q; PACF decays; For ARMA both decay, so compare candidate models using diagnostics and AIC.
Approximate significance band: ±1.96/√n (often quoted as ±2/√n); A sample autocorrelation outside the band is taken as significantly different from 0.
AR(1) model: X_t − μ = α(X_{t−1} − μ) + e_t, |α| < 1; e_t is white noise with variance σ².
AR(1) k-step forecast: x̂_t(k) = μ + α^k (x_t − μ); Replace α and μ by estimates when fitting from data.
AR(1) k-step forecast error variance: σ² (1 − α^{2k}) ÷ (1 − α²); Tends to σ² ÷ (1 − α²), the process variance, as k grows.
Yule-Walker for AR(1): α̂ = r₁ (sample lag-1 autocorrelation); For AR(1), ρ₁ = α.
Information criterion: AIC = −2 ln L + 2 × (number of parameters); Smaller is better. Compare models fitted to the same data.
Differencing: ∇X_t = X_t − X_{t−1}; ARIMA(p, d, q) means ∇^d X_t is a stationary ARMA(p, q).

How to solve Model Identification, Fitting and Forecasting questions

Use this order for any question on building, checking or using a time series model.

1Look at the plot and the sample ACF. A trend or a very slowly decaying ACF means the series is not stationary, so difference it and recheck.
2Read the sample ACF and PACF of the stationary series. Note which cuts off and where, using the ±2/√n band as a guide.
3Propose one or more candidate models. State the pattern that supports each choice.
4Estimate the parameters. For AR(1), use α̂ = r₁ and the sample mean for μ if the question asks for moment estimates. State the method used.
5Check the residuals. Look at the residual ACF, a portmanteau test such as Ljung-Box, a plot against time, and a normality check. Say what failure would mean.
6If the diagnostics fail, revise the model. If several fit, choose using AIC or parsimony.
7Forecast using the conditional expectation. Substitute known values, and set future errors to zero. Compute the forecast error variance if asked.
8State the answer with units and mention the assumptions, such as stationarity and constant parameters.

Quickest way: Quick read of ACF and PACF, then AR(1) forecast

When to use it: Use this for MCQs and short written parts that give you a plot or a few values and ask for the model type or a forecast.

Slow ACF decay that stays near 1 means difference first.
Find which plot cuts off sharply. PACF cut-off at p means AR(p). ACF cut-off at q means MA(q).
If neither cuts off, say ARMA and suggest comparing small models.
For AR(1) forecasts, compute the current deviation x_t − μ, multiply by α^k, and add μ.
For error variance, use σ²(1 − α^{2k}) ÷ (1 − α²). Check that k = 1 gives σ².

Common mistakes in Model Identification, Fitting and Forecasting

Fitting a model to a non-stationary series without differencing.
Students jump straight to the ACF and PACF patterns, which only apply to stationary series.
Fix: Check the plot and the ACF decay first. Difference and recheck before identifying p and q.
Swapping the cut-off rules for AR and MA.
Both patterns look similar and are easy to mix up under time pressure.
Fix: Remember: AR is identified by the PACF cutting off. MA is identified by the ACF cutting off.
Treating one sample value just outside the ±2/√n band as proof of a nonzero lag.
Students forget the band is approximate and that about one in twenty values will fall outside by chance.
Fix: Look for a clear pattern and consider isolated, high-lag spikes as likely noise. Support the choice with other diagnostics.
Forecasting AR(1) as α^k x_t instead of μ + α^k (x_t − μ).
Students forget that the model is written in deviations from the mean.
Fix: Subtract μ first, scale by α^k, then add μ back. Check that the forecast tends to μ as k grows.
Using residual diagnostics that test the wrong thing, such as saying a good fit means the residuals are large or small.
Students confuse fit quality with residual size rather than residual structure.
Fix: A good model leaves residuals that look like white noise: no autocorrelation, constant variance. Say this explicitly.
Choosing the model with the highest likelihood and ignoring the number of parameters.
Adding terms always raises the likelihood, so it looks better.
Fix: Use AIC or another penalised criterion and prefer the simpler model when fits are similar.

Worked examples

Example 1

A stationary series has n = 100 observations. The sample PACF has a large spike at lag 1 (0.72) and values below 0.1 in absolute value at all later lags. The sample ACF declines gradually: 0.72, 0.52, 0.37, 0.27. (a) Suggest a model and estimate α. (b) Is a lag-3 PACF value of 0.15 significant?

Show the solution

The PACF cuts off after lag 1 and the ACF decays gradually. This pattern fits AR(1).
For AR(1), ρ₁ = α, so the Yule-Walker estimate is α̂ = r₁ = 0.72.
Check the ACF: 0.72² = 0.5184 ≈ 0.52, and 0.72³ ≈ 0.373 ≈ 0.37. This is consistent with AR(1).
The approximate band is ±2/√100 = ±0.2.
A PACF value of 0.15 lies inside ±0.2, so it is not significant.

Answer: AR(1) with α̂ = 0.72. The lag-3 PACF value of 0.15 is not significant.

Example 2

An AR(1) model X_t − 50 = 0.8(X_{t−1} − 50) + e_t has σ² = 9. The latest observation is x_t = 60. Find the 2-step forecast and the variance of its forecast error.

Show the solution

The deviation from the mean is 60 − 50 = 10.
The forecast is μ + α^k (x_t − μ) with k = 2: α² = 0.64.
Forecast = 50 + 0.64 × 10 = 56.4.
The error variance is σ²(1 − α^{2k}) ÷ (1 − α²) with k = 2.
α⁴ = 0.64² = 0.4096, so 1 − α⁴ = 0.5904.
1 − α² = 0.36.
Variance = 9 × 0.5904 ÷ 0.36 = 9 × 1.64 = 14.76.
Check directly: error = e_{t+2} + 0.8 e_{t+1}, with variance 9(1 + 0.64) = 14.76.

Answer: The 2-step forecast is 56.4 and the forecast error variance is 14.76.

Exam tips

Write the pattern in words before naming the model, for example 'PACF cuts off after lag 2, ACF decays, so AR(2)'. Marks often go to the reasoning.
In diagnostics answers, name at least two checks: residual ACF with a portmanteau test, and a plot of residuals against time for constant variance.
For forecasts, show the formula, the substitution and the final number. State what happens as k grows.
In the computer-based paper, state the model order, the fitted coefficients and the residual check you ran, not only the code output.
Mention assumptions: stationarity, constant parameters and white noise errors with constant variance.

Practice questions from Core concepts of time series models

Model Identification, Fitting and Forecasting: frequently asked questions

What are the steps of the Box-Jenkins method?

There are four: identify the model order using the ACF and PACF after making the series stationary, estimate the parameters, check the residuals for white noise, and forecast. If the checks fail, return to identification and try another model.

How do I identify an ARIMA model from the ACF and PACF?

Difference until the series is stationary and count the differences as d. Then, if the PACF cuts off after lag p, use AR(p). If the ACF cuts off after lag q, use MA(q). If both decay, try ARMA models and compare them.

How do I forecast k steps ahead for an AR(1)?

Use x̂_t(k) = μ + α^k (x_t − μ). The forecast moves geometrically toward the mean as k increases. The error variance rises to the process variance σ² ÷ (1 − α²).

What do residual diagnostics check?

They check that the residuals behave like white noise. You look at the residual ACF, a portmanteau test, a plot against time and, where relevant, a normality check. Any remaining pattern means the model is missing structure.