Time Series Analysis: Interpretation of ACF and PACF Plots

Autocorrelation (ACF) and Partial Autocorrelation (PACF) plots are powerful tools for uncovering hidden patterns in time series data, guiding us towards accurate ARIMA model selection. However, their effectiveness hinges on a fundamental assumption: stationarity. This article provides a thorough exploration of ACF/PACF analysis, addressing crucial questions about stationarity, differencing, interpretation, and the underlying AR and MA components, ultimately equipping you with the knowledge to navigate the intricacies of time series modeling.

1. The Crucial Role of Stationarity

ACF and PACF plots are designed to reveal the autoregressive (AR) and moving average (MA) components within a time series. These components rely on the principle of stationarity, which assumes that the statistical properties of the data (e.g., mean, variance) remain constant over time.

What if We Ignore Non-Stationarity?

• Misleading Correlations: Non-stationary data, often exhibiting trends or seasonality, can produce spurious correlations in ACF/PACF plots. These correlations are artifacts of the non-stationarity, not true underlying patterns.
• Incorrect Model Identification: Misinterpreting ACF/PACF plots due to non-stationarity leads to selecting inappropriate ARIMA models, resulting in poor forecasts and unreliable insights.

2. Detecting Non-Stationarity: A Two-Pronged Approach

Before even plotting ACF/PACF, we must rigorously test for stationarity:

Visual Inspection: Start by plotting the time series. Look for:

• Trends: Upward or downward slopes in the data over time.
• Seasonality: Repeating patterns at regular intervals (e.g., peaks in summer, dips in winter).
• Changing Variance: Increasing or decreasing spread of the data over time.
• Statistical Tests: Visual inspection provides initial clues, but formal tests provide more robust evidence:
• Augmented Dickey-Fuller (ADF) Test: Tests for the presence of a unit root (indicating non-stationarity). A low p-value (< 0.05) suggests stationarity.
• Kwiatkowski-Phillips-Schmidt-Shin (KPSS) Test: Directly tests for stationarity. A high p-value (> 0.05) suggests stationarity.

3. Achieving Stationarity: The Art of Differencing

If non-stationarity is detected, differencing comes to the rescue:

First Differencing: Subtracts each observation from its predecessor, effectively de-trending the series.

series_diff = series.diff().dropna()

Seasonal Differencing: Subtracts the observation from the same season in the previous cycle, mitigating seasonality.

series_diff = series.diff(periods=12).dropna() # For monthly data with yearly seasonality

Applies differencing twice, often when both trend and seasonality are strong.

Should We Test for Stationarity After Differencing?

• Yes, Absolutely! After each differencing step, re-run the stationarity tests (ADF or KPSS) and visually inspect the differenced series. This iterative approach ensures you achieve stationarity without over-differencing, which can introduce artificial patterns.

When to Use Seasonal Differencing?

• Apply seasonal differencing when you observe seasonality in your data and the seasonality is not adequately addressed by first differencing alone. Look for repeating peaks or troughs at regular intervals in the time series plot.

4. Understanding AR and MA Components

• Autoregressive (AR) Component: An AR model predicts future values based on a linear combination of its own past values. The order (p) of the AR component indicates how many past values to consider. For example, an AR(2) model uses the values from two time steps ago to predict the current value.
• Moving Average (MA) Component: An MA model predicts future values based on past forecast errors (the difference between the actual value and the predicted value). The order (q) of the MA component indicates how many past errors to consider. For instance, an MA(1) model uses the error from one time step ago.

5. Interpreting ACF and PACF Plots: Unveiling the ARIMA Structure

Once stationarity is achieved, we can confidently interpret the ACF and PACF plots of the differenced series:

ACF Plot (for MA component):

• Gradual Decay: Suggests an autoregressive (AR) component.
• Sharp Cut-off: Suggests a moving average (MA) component. The lag at the cut-off indicates the MA order.
• Significant Peaks: May indicate remaining seasonality or other patterns.

PACF Plot (for AR component):

• Gradual Decay: Suggests an MA component.
• Sharp Cut-off: Suggests an AR component. The lag at the cut-off indicates the AR order.
• Significant Peaks: Can provide clues about the AR order.

6. Python Implementation: Bringing it All Together

import pandas as pd
import numpy as np
from statsmodels.tsa.stattools import adfuller, kpss
from statsmodels.graphics.tsaplots import plot_acf, plot_pacf
import matplotlib.pyplot as plt
# Generate dummy data with trend and seasonality
np.random.seed(42)
time = pd.date_range("2023–01–01", periods=100, freq="M")
data = np.random.randn(100) + np.arange(100) * 0.1 + 2*np.sin(2*np.pi*np.arange(100)/12)
series = pd.Series(data, index=time)
# Stationarity Tests (before differencing)
print(f"ADF p-value (Original): {adfuller(series)[1]:.3f}")
print(f"KPSS p-value (Original): {kpss(series)[1]:.3f}")

Output:

ADF p-value (Original): 0.977

KPSS p-value (Original): 0.010

# Differencing
series_diff = series.diff().diff(periods=12).dropna()
# Stationarity Tests (after differencing)
print(f"ADF p-value (Differenced): {adfuller(series_diff)[1]:.3f}")
print(f"KPSS p-value (Differenced): {kpss(series_diff)[1]:.3f}")

Output:

ADF p-value (Differenced): 0.005

KPSS p-value (Differenced): 0.100

# ACF and PACF Plots
plt.figure(figsize=(12, 6))
plt.subplot(211)
plot_acf(series_diff, lags=20, ax=plt.gca())
plt.title("ACF of Differenced Series")
plt.subplot(212)
plot_pacf(series_diff, lags=20, ax=plt.gca())
plt.title("PACF of Differenced Series")
plt.tight_layout()
plt.show()

7. From Plots to ARIMA: Guiding Model Selection

The ACF and PACF plots of the differenced series provide crucial insights for choosing an appropriate ARIMA model:

• AR(p): The order (p) of the autoregressive component is suggested by the cut-off in the PACF plot.
• I(d): The order (d) of the integrated component is determined by the number of differencing steps taken to achieve stationarity.
• MA(q): The order (q) of the moving average component is suggested by the cut-off in the ACF plot.

Key Takeaways

• Stationarity is paramount for reliable ACF/PACF analysis and ARIMA model selection.
• Use a combination of visual inspection and statistical tests to diagnose non-stationarity.
• Apply differencing strategically to achieve stationarity without over-differencing.
• Understand the AR and MA components and how they are reflected in ACF/PACF plots.
• Interpret the ACF and PACF plots of the differenced series to guide your choice of ARIMA model orders (p, d, q).
• Remember that model selection is an iterative process, and ACF/PACF plots provide valuable starting points for further analysis and refinement.

Stackademic 🎓

Thank you for reading until the end. Before you go:

• Please consider clapping and following the writer! 👏
• Follow us **X | [LinkedIn](https://www.linkedin.com/company/stackademic) | [YouTube](https://www.youtube.com/c/stackademic) | [Discord](https://discord.gg/in-plain-english-709094664682340443)**
• Visit our other platforms: **In Plain English | [CoFeed](https://cofeed.app/) | [Differ](https://differ.blog/)**
• More content at **Stackademic.com**