Implementing Decomposition in Python

Implementing time series decomposition in Python involves utilizing libraries and functions to analyze components such as trend, seasonality, and residuals.

We've discussed the concepts of time series decomposition and the distinction between additive and multiplicative models. Now, let's put this into practice using Python. The statsmodels library provides tools for statistical analysis, including time series decomposition.

The main function we'll employ is statsmodels.tsa.seasonal.seasonal_decompose. This function implements a classical decomposition method, typically based on moving averages, to separate a time series into its trend, seasonal, and residual components.

Using seasonal_decompose

The function seasonal_decompose requires the time series itself and information about the model type and seasonal period.

import pandas as pd
from statsmodels.tsa.seasonal import seasonal_decompose
import matplotlib.pyplot as plt # Using matplotlib for basic plotting example

# Assume 'series' is your pandas Series with a DatetimeIndex.
# It should contain the time series values you want to decompose.
# Example:
# series = pd.read_csv('your_data.csv',
#                     index_col='Date',
#                     parse_dates=True)['Value']

# --- Perform Decomposition ---
# model: Specify 'additive' or 'multiplicative'.
# period: The number of observations per seasonal cycle
#         (e.g., 12 for monthly data with yearly seasonality,
#          7 for daily data with weekly seasonality).

# Example for additive model with monthly data (period=12)
decomposition_result = seasonal_decompose(series, model='additive', period=12)

# The function returns a DecomposeResult object. You can access
# the individual components as attributes:
trend_component = decomposition_result.trend
seasonal_component = decomposition_result.seasonal
residual_component = decomposition_result.resid
observed_data = decomposition_result.observed # Original data

# --- Visualize the Components ---
# A common way to visualize is stacking the plots:
fig, axes = plt.subplots(4, 1, figsize=(10, 8), sharex=True)

observed_data.plot(ax=axes[0], legend=False, color='#1c7ed6')
axes[0].set_ylabel('Observed')

trend_component.plot(ax=axes[1], legend=False, color='#f76707')
axes[1].set_ylabel('Trend')

seasonal_component.plot(ax=axes[2], legend=False, color='#37b24d')
axes[2].set_ylabel('Seasonal')

residual_component.plot(ax=axes[3], legend=False, color='#adb5bd')
axes[3].set_ylabel('Residual')

plt.xlabel('Date')
plt.suptitle('Time Series Decomposition', y=0.92) # Add a title slightly above plots
plt.tight_layout(rect=[0, 0, 1, 0.9]) # Adjust layout to prevent title overlap
plt.show()

Choosing the Model Type:

• Set model='additive' when the seasonal variations appear relatively constant over time, regardless of the series' level. The magnitude of the peaks and troughs in the seasonal pattern remains roughly the same. This corresponds to the $y_t = T_t + S_t + R_t$ model.
• Set model='multiplicative' when the seasonal variations seem to scale with the level of the series. As the trend increases, the amplitude of the seasonal pattern also increases. This fits the $y_t = T_t \times S_t \times R_t$ model. If you suspect multiplicative effects, sometimes analyzing the logarithm of the series can make the pattern additive ($log(y_t) \approx log(T_t) + log(S_t) + log(R_t)$), which can simplify modeling.

Specifying the period:

This parameter is significant for correctly identifying the seasonal pattern. It represents the number of time steps in a full seasonal cycle. Common values include:

• period=12 for monthly data with an annual cycle.
• period=4 for quarterly data with an annual cycle.
• period=7 for daily data with a weekly cycle.
• period=52 for weekly data with an annual cycle (approximately).

Choosing the correct period based on your data's known frequency is essential for meaningful decomposition.

Example: Airline Passenger Data Decomposition

Let's apply this to the classic 'AirPassengers' dataset, known for its strong upward trend and increasing annual seasonality. This increasing seasonality suggests a multiplicative model might be more appropriate.

We'll perform both decompositions for comparison. (Assume air_passengers is a pandas Series containing the monthly passenger counts, indexed by date).

# Assume 'air_passengers' Series is loaded and prepared with a DatetimeIndex
# Example synthetic data generation for demonstration:
date_rng = pd.date_range(start='1949-01-01', end='1960-12-01', freq='MS')
passengers = (
    110 +  # Base level slightly higher
    (date_rng.year - 1949) * 22 +  # Slightly steeper trend
    (1 + (date_rng.year - 1949) * 0.12) * ( # Slightly stronger seasonality increase
        35 * (date_rng.month == 7) + 25 * (date_rng.month == 8) +
        -25 * (date_rng.month == 11) + -20 * (date_rng.month == 2)
    ) +
    pd.Series(np.random.normal(0, 12, size=len(date_rng)), index=date_rng) # Slightly more noise
).astype(int)
air_passengers = pd.Series(passengers, index=date_rng)

# Perform Multiplicative Decomposition (period=12 for monthly annual cycle)
result_mul = seasonal_decompose(air_passengers, model='multiplicative', period=12)

# Perform Additive Decomposition for comparison
result_add = seasonal_decompose(air_passengers, model='additive', period=12)

# -- Plotting with Plotly for better web display --
import plotly.graph_objects as go
from plotly.subplots import make_subplots
import numpy as np # Needed for np.nan comparison

fig = make_subplots(
    rows=4, cols=2,
    shared_xaxes=True,
    subplot_titles=('Multiplicative: Observed', 'Additive: Observed',
                    'Multiplicative: Trend', 'Additive: Trend',
                    'Multiplicative: Seasonal', 'Additive: Seasonal',
                    'Multiplicative: Residual', 'Additive: Residual'),
    vertical_spacing=0.06 # Slightly more spacing
)

# Function to add traces, handling potential NaNs at ends for trend/resid
def add_trace_safe(fig, data, name, color, row, col, showlegend=True):
    # Only plot where data is not NaN
    valid_index = data.dropna().index
    valid_data = data.dropna().values
    fig.add_trace(go.Scatter(x=valid_index, y=valid_data, mode='lines',
                             name=name, line=dict(color=color), showlegend=showlegend),
                  row=row, col=col)

# Add traces column by column
add_trace_safe(fig, result_mul.observed, 'Observed', '#1c7ed6', 1, 1)
add_trace_safe(fig, result_mul.trend, 'Trend', '#f76707', 2, 1)
add_trace_safe(fig, result_mul.seasonal, 'Seasonal', '#37b24d', 3, 1)
add_trace_safe(fig, result_mul.resid, 'Residual', '#adb5bd', 4, 1)

add_trace_safe(fig, result_add.observed, 'Observed', '#1c7ed6', 1, 2, showlegend=False)
add_trace_safe(fig, result_add.trend, 'Trend', '#f76707', 2, 2, showlegend=False)
add_trace_safe(fig, result_add.seasonal, 'Seasonal', '#37b24d', 3, 2, showlegend=False)
add_trace_safe(fig, result_add.resid, 'Residual', '#adb5bd', 4, 2, showlegend=False)

fig.update_layout(
    height=750, # Slightly taller
    title_text='Time Series Decomposition: Multiplicative vs. Additive (Air Passengers)',
    margin=dict(l=60, r=30, t=90, b=60),
    hovermode='x unified',
    legend_title_text='Components',
    plot_bgcolor='#e9ecef', # Light gray background
    paper_bgcolor='#ffffff' # White paper background
)
fig.update_xaxes(title_text="Date", row=4, col=1, showgrid=False)
fig.update_xaxes(title_text="Date", row=4, col=2, showgrid=False)
fig.update_yaxes(showgrid=True, gridwidth=1, gridcolor='#dee2e6') # Lighter grid lines

# To generate the JSON output for embedding:
# print(fig.to_json())

Comparison of multiplicative (left column) and additive (right column) decomposition applied to the air passenger dataset.

Interpreting the Decomposition Plots

• Observed: The original time series data as loaded.
• Trend: This component isolates the long-term direction or level change in the series. For the air passengers, it correctly captures the consistent increase over the years. Note the NaN values (missing segments) at the beginning and end of the trend line. This happens because the moving average calculation used internally requires data points both before and after the current point.
• Seasonal: This plot shows the pattern that repeats every period (annually in this case).
  • In the multiplicative decomposition, the seasonal component oscillates around 1.0. It represents factors that scale the trend (e.g., a value of 1.2 might mean 20% above the trend level for that month). Notice how this pattern appears relatively stable in shape across the years.
  • In the additive decomposition, the seasonal component oscillates around 0. It represents absolute values added to or subtracted from the trend. For the air passenger data, notice how the amplitude of this additive seasonal component seems to increase over time, mirroring the overall increase in passenger numbers. This suggests it doesn't capture the seasonality as effectively as the multiplicative model, where the seasonal factor remains more consistent.
• Residual: This is the remainder after removing the estimated trend and seasonal components. Ideally, the residuals should look like random noise without any obvious patterns or structures.
  • Comparing the two residual plots, the residuals from the multiplicative decomposition appear more random and evenly scattered around 1.0, with a relatively constant variance over time.
  • The residuals from the additive decomposition, however, show some fanning out (increasing variance) over time and might retain some underlying pattern, especially in later years. This further supports that the multiplicative model was a better choice for this dataset, as it left residuals that are closer to being random noise.

Approaches for seasonal_decompose

While seasonal_decompose is a useful exploratory tool, keep these points in mind:

1. Methodology: It typically uses centered moving averages. This approach can be less effective if the trend changes rapidly or if there are sharp spikes or dips in the data. The seasonal component is computed by averaging the detrended values for each season and then centering these factors.
2. End Point Issues: As mentioned, trend and residual estimates near the series' start and end points are often missing (NaN) or less reliable due to the moving average requiring a full window of data. "3. Static Seasonality Assumption: The method assumes that the seasonal pattern is either perfectly additive or perfectly multiplicative throughout the entire series. Seasonality might evolve in more complex ways over long periods."
3. Descriptive, Not Predictive: Decomposition is primarily for understanding the historical structure of the data. It doesn't directly generate forecasts, although understanding the components is important for selecting and building forecasting models. More advanced methods like STL (Seasonal and Trend decomposition using Loess) can offer more flexibility.

Decomposition provides valuable insights into the underlying structure of your time series. Examining the components, particularly the residuals, helps assess whether the original series exhibits clear patterns and whether removing them results in something closer to random noise. This assessment is a direct lead-in to the concept of stationarity, which we will investigate more formally using statistical tests in the following sections.