Box–Jenkins Methodology
Prof. Dr. Sohail Chand
ARIMA Models
ARIMA stands for:
• AR: Autoregressive component
• I: Integrated (differencing)
• MA: Moving Average component
Used for:
• Univariate time series
• Short- to medium-term forecasting
2
Box-Jenkins Model Building
ARIMA modeling follows an iterative procedure:
1. Identification
2. Estimation
3. Diagnostic Checking
Feedback loop until an adequate model is
obtained.
3
Step-by-Step Procedure
1. Data screening and visualization
2. Stationarity and variance checks
3. Transformation and differencing
4. ACF/PACF-based identification
5. Model estimation
6. Diagnostic checking (if model fails go to step-4)
7. Model selection
8. Forecasting and validation
4
Screening
Before modeling, ask:
• What does the series represent?
• Frequency: daily, monthly, quarterly?
• Presence of seasonality or trends?
• Structural breaks or interventions?
Contextual understanding is critical.
5
Visual Exploration of Data
Key plots:
• Time series plot
• Rolling mean and variance
• Seasonal subseries plot
Look for:
• Trend
• Seasonality
• Outliers
• Variance instability
6
Handling Missing Values and Outliers
Missing values:
1. Interpolation
2. Kalman smoothing
3. Truncation (if minimal)
Outliers:
• Investigate source
• Adjust cautiously or model explicitly
7
Assumptions of ARIMA
Key assumptions:
1. Weak stationarity (after differencing if
required)
2. Constant variance
3. Uncorrelated residuals
4. Approximately normal errors (for inference)
8
Checking Stationarity – Visual Tools
Indicators of non-stationarity:
1. Persistent trend
2. Changing variance
3. Slowly decaying ACF
Visual tools:
1. Time plot
2. ACF plot
9
Stationarity – Formal Tests
Common tests:
1. Augmented Dickey–Fuller (ADF)
– H₀: Unit root (non-stationary)
2. KPSS Test
– H₀: Stationary
Use tests jointly for robustness.
10
Variance Stabilization
If variance varies with level, we can use the
following transformations:
• Log transformation
• Square root transformation
• Box–Cox transformation (can address non-
normality)
11
Over-differencing
Symptoms:
1. Negative autocorrelation at lag 1
2. Increased variance
3. Poor forecasts
Rule: Difference only as much as needed.
12
Model Identification
After stationarity:
• Use ACF and PACF plots
• Identify AR and MA orders
Model form: ARIMA(p, d, q)
13
ACF and PACF
Typical ACF patterns:
• Sharp cutoff after lag q → MA(q)
• Slow decay → AR or mixed model
Typical PACF patterns:
• Sharp cutoff after lag p → AR(p)
• Slow decay → MA or mixed model
14
Parsimony Principle
• Prefer simpler models
• Avoid unnecessary parameters
• Theory + diagnostics guide final choice
15
Model Estimation
MA model
• Method of least squares
AR model
• Maximum Likelihood Method
• Yule-Walker Method
16
Model Selection
Compare candidate models using:
• AIC (Akaike Information Criterion)
• BIC (Bayesian Information Criterion)
Lower value → better balance of fit and
complexity.
17
Diagnostic Checking
Goal: verify that residuals behave like white
noise.
Residuals should be:
1. Uncorrelated
2. Zero mean
3. Constant variance
4. Approximately normal
18
Residuals Analysis
• Residuals time plot
• ACF of residuals
No significant autocorrelation should remain.
19
Portmanteau Test
Tests joint autocorrelation of residuals
H₀: Residuals are independent
Failure to reject H₀ → model is adequate.
• Box-Pierce Test
• Ljung-Box Test
• Monti’s Test
• Mixed Portmantau Test
20
Normality of Residuals
Check using:
• Histogram
• Q–Q plot
• Shapiro–Wilk test (optional)
Normality is important for inference and
prediction intervals.
21
Parameters Significance
• Examine standard errors and t-statistics
• Insignificant parameters may indicate
overfitting
Refine model if necessary.
22
Forecasting
Once diagnostics are satisfactory:
• Generate point forecasts
• Construct prediction intervals
Remember to transform back, if required, the
forecasts to original variable.
23
Forecast Evaluation
Preferably for the validation data, compare the
competing models using the following measures
based on forecast error:
• MAE
• RMSE
• MAPE
24