Learn→Time Series Forecasting→Statistical Models (ARIMA, Exponential Smoothing)→2 of 6

Time Series Forecasting • Statistical Models (ARIMA, Exponential Smoothing)Easy⏱️ ~3 min

Exponential Smoothing: Weighted Averages of the Past

Core Idea
Exponential smoothing forecasts using a weighted average of past values—recent values count more, weights decrease exponentially going back in time.
THE INTUITION
Predicting tomorrow temperature: yesterday matters most, last week some, last month barely. Formula: forecast = α × latest + (1-α) × previous_forecast. Alpha (0-1) controls weight on latest. α=0.3 means 30% on latest, 70% on accumulated past.
CHOOSING ALPHA
High α (0.7-0.9): Reacts quickly, forgets fast. Good for volatile data. Low α (0.1-0.3): Stable forecasts, slow to react. Good for steady data. Optimize by testing which minimizes forecast error on historical data.
💡 Insight: Simple smoothing predicts flat forecasts—tomorrow equals today, forever. Works when data fluctuates around stable mean, fails for trending or seasonal data.
ADDING TREND
Real data often trends. Holt method adds a component tracking rate of change. Forecast extends current level along slope. Damped variant gradually flattens—useful because most trends do not continue forever.
ADDING SEASONALITY
Many series repeat: retail spikes in December, traffic drops on weekends. Holt-Winters learns seasonal factors (12 monthly multipliers for yearly patterns). Needs 2+ complete cycles of history.
⚠️ Trade-off: More components = more parameters = more data needed. Start simple, add trend if data grows, add seasonality only with clear patterns and sufficient history.

💡 Key Takeaways

✓Exponential smoothing forecasts using weighted average of past values—recent values weighted more

✓Alpha parameter (0-1) controls how fast old data is forgotten; optimize by minimizing historical error

✓Simple smoothing produces flat forecasts; add trend component for growing/shrinking data

✓Holt-Winters adds seasonal factors for repeating patterns (requires 2+ cycles of history)

✓More components need more parameters and more data—start simple, add complexity only if needed

📌 Interview Tips

1Explain the intuition: recent data matters more than old data, weights decay exponentially

2Know when to add components: flat data = simple, trending = Holt, seasonal = Holt-Winters

← Back to Statistical Models (ARIMA, Exponential Smoothing) Overview