Forecasting Horizon and Evaluation Windows
What you'll learn: How to define the timeframe for predictions and properly evaluate forecast quality across different temporal windows.
The Core Idea
When forecasting time series, you must decide how far into the future you want to predict (the forecasting horizon) and how you'll measure accuracy across that period (the evaluation window). These choices fundamentally shape your model architecture, training strategy, and performance metrics.
Single-Step vs Multi-Step Forecasting
Single-step forecasting predicts just the next time point. Think of predicting tomorrow's temperature: you forecast one day ahead, then retrain or update with new data before the next prediction.
Multi-step forecasting predicts multiple future points at once. For example, forecasting the next 7 days of temperature simultaneously. This is harder because errors can compound—a mistake in day 2's prediction might corrupt day 3, 4, 5...
Forecasting Horizon Constraints
Your horizon (how many steps ahead) determines:
- Model complexity: Longer horizons may need more sophisticated features
- Prediction confidence: Accuracy typically degrades as you predict further out
- Business utility: A 1-hour-ahead forecast serves different needs than a 1-month-ahead forecast
Evaluation Windows
Instead of testing on a single period, use sliding or expanding windows to assess robustness:
- Fixed window: Always use the last N observations to predict H steps ahead
- Expanding window: Gradually include more historical data as you move forward
- Rolling forecast: Predict H steps, move forward 1 step, predict again—mimics real deployment
This matches the Train-Test Splitting for Time Series concept you learned: respecting temporal order while testing generalization.
Key Takeaway: Your forecasting horizon defines how far you predict, while evaluation windows ensure you test your model across multiple realistic scenarios—critical for understanding whether your forecasts will hold up in production.