How to Read PACF Graph for Time Series?

Time series analysis is essential in finance, economics, and marketing. The Partial Autocorrelation Function (PACF) is a powerful tool for identifying direct relationships between observations at different time lags. This article explains how to read and interpret PACF graphs step-by-step.

What is PACF?

The Partial Autocorrelation Function (PACF) measures the direct correlation between an observation and its lagged values, while controlling for the effects of intermediate lags. Unlike the regular autocorrelation function (ACF) which shows all correlations, PACF isolates the direct relationship by removing indirect effects.

PACF is particularly useful for determining the order of Autoregressive (AR) models in time series analysis. It helps identify which lag values should be included in the model for accurate forecasting.

Understanding PACF Graphs

A PACF graph displays partial autocorrelation values for various lag values of a time series. The graph helps identify significant lags that are crucial for predicting future values.

Step-by-Step Guide to Reading PACF Graphs

Step 1: Identify the Lag Range

The x-axis shows lag values (time periods), while the y-axis displays partial autocorrelation coefficients. Lag values typically range from 1 to n/4, where n is the length of your time series.

Step 2: Locate the Significance Threshold

Horizontal dashed lines represent the confidence bounds (usually 95%). Any PACF value beyond these lines is considered statistically significant. Values within the bounds suggest no significant correlation at that lag.

Step 3: Identify Significant Lags

Look for PACF values that exceed the confidence bounds. These significant lags indicate direct relationships between the current observation and lagged values.

import numpy as np
import matplotlib.pyplot as plt
from statsmodels.tsa.stattools import pacf
from statsmodels.graphics.tsaplots import plot_pacf

# Generate sample time series data
np.random.seed(42)
data = np.cumsum(np.random.randn(100))

# Calculate PACF
pacf_values = pacf(data, nlags=20)

# Plot PACF
fig, ax = plt.subplots(figsize=(10, 6))
plot_pacf(data, lags=20, ax=ax)
plt.title("PACF Plot Example")
plt.show()

# Print significant lags
significance_level = 1.96 / np.sqrt(len(data))
significant_lags = []
for i, val in enumerate(pacf_values[1:], 1):
    if abs(val) > significance_level:
        significant_lags.append(i)

print(f"Significant lags: {significant_lags}")

Step 4: Determine AR Model Order

The number of significant lags suggests the order of the AR model. If lags 1 and 2 are significant while others are not, consider an AR(2) model.

Step 5: Look for Cut-off Pattern

A clear cut-off occurs when PACF values drop to insignificance after a certain lag. This pattern indicates the optimal AR order. Gradual decay suggests additional modeling considerations.

Common PACF Patterns

Practical Example

import pandas as pd
import numpy as np
from statsmodels.tsa.stattools import pacf

# Create sample AR(2) process
np.random.seed(123)
n = 200
data = [0, 0]

# AR(2): X_t = 0.7*X_{t-1} - 0.3*X_{t-2} + error
for i in range(2, n):
    value = 0.7 * data[i-1] - 0.3 * data[i-2] + np.random.randn()
    data.append(value)

# Calculate PACF
pacf_values = pacf(data, nlags=10)

# Display results
for lag, value in enumerate(pacf_values):
    significance = abs(value) > 1.96/np.sqrt(len(data))
    print(f"Lag {lag}: {value:.3f} {'(Significant)' if significance else ''}")

Lag 0: 1.000 (Significant)
Lag 1: 0.689 (Significant)
Lag 2: -0.285 (Significant)
Lag 3: 0.045 
Lag 4: -0.032 
Lag 5: 0.018 
Lag 6: 0.021 
Lag 7: -0.067 
Lag 8: 0.039 
Lag 9: 0.015 
Lag 10: -0.041 

Conclusion

Reading PACF graphs involves identifying significant lags beyond confidence bounds and recognizing cut-off patterns. Use PACF to determine AR model order: significant lags indicate direct relationships, while cut-off patterns suggest optimal model complexity. Practice with different time series will improve your interpretation skills.