Course contentsShow
← Data Science
Lesson 749 of 2,14519. Time Series FoundationsFree lesson

Using Decomposition for Forecasting

Applying decomposed components to improve forecast accuracy and model selection.

Using Decomposition for Forecasting

What you'll learn: How to leverage decomposed time series components to build more accurate forecasts and choose better models.

Why Decomposition Improves Forecasting

Once you've separated your time series into trend, seasonal, and irregular components (through methods like STL decomposition), you can forecast each piece independently and then recombine them. This is often more accurate than trying to forecast the raw series directly—like solving smaller, simpler puzzles instead of one complex one.

The Decomposition Forecasting Strategy

Step 1: Decompose your historical data into trend, seasonal, and remainder components using your chosen method (classical or STL).

Step 2: Forecast the trend separately using appropriate methods for smooth patterns (like linear models or exponential smoothing).

Step 3: Project the seasonal pattern forward—since seasonality repeats, you simply extend the seasonal component from your last cycle.

Step 4: Handle the irregular component carefully—this is random noise, so typically you assume it averages to zero in forecasts (though you can use it to estimate forecast uncertainty).

Step 5: Recombine using your model type:

  • Additive model: Forecast = Trend forecast + Seasonal component
  • Multiplicative model: Forecast = Trend forecast × Seasonal component

Model Selection Benefits

Decomposition helps you identify what type of model you need. If you see strong seasonality, you know seasonal models are essential. If the trend is clearly linear, simpler models may suffice. If seasonality's strength changes with the level, a multiplicative model is better than additive.

Key Takeaway: Decomposition transforms complex forecasting into manageable parts—forecast simpler components separately, then recombine them according to your additive or multiplicative model structure for improved accuracy and clearer model selection.