5.2 Simple exponential smoothing

Simple Exponential Smoothing

Simple exponential smoothing (SES) is a forecasting technique that weights recent observations more heavily than older ones. It works well for short-term forecasts when your data has no clear trend or seasonal pattern, essentially smoothing out noise to reveal the underlying level of a series.

The method relies on a single parameter, alpha ($\alpha$), to control how much influence new data has versus old data. By adjusting $\alpha$, you can make forecasts more reactive to recent changes or more stable over time.

Concept of Exponential Smoothing

The core idea is that not all past observations deserve equal attention. SES assigns exponentially decreasing weights to older data points, so the most recent observation has the strongest influence on your forecast, the one before it has less, the one before that even less, and so on.

For example, if $\alpha = 0.3$, the most recent observation gets a weight of 0.3, the one before it gets roughly $0.3 \times 0.7 = 0.21$, the next gets about $0.3 \times 0.7^2 = 0.147$, and so on. The weights shrink quickly, which is why it's called exponential smoothing.

This approach is best suited for stationary time series, meaning data that fluctuates around a roughly constant level without trending upward/downward or showing seasonal cycles. Common use cases include daily sales figures or weekly website traffic where you just need a near-term forecast.

Application of Simple Exponential Smoothing

The SES forecast formula is:

$\hat{y}_{t+1} = \alpha \, y_t + (1 - \alpha) \, \hat{y}_t$

where:

• $\hat{y}_{t+1}$ = forecast for the next period
• $y_t$ = actual observed value at time $t$
• $\hat{y}_t$ = forecast for time $t$ (your previous forecast)
• $\alpha$ = smoothing parameter, a value between 0 and 1

In plain terms: your next forecast is a blend of what actually just happened ($y_t$) and what you previously predicted ($\hat{y}_t$). The parameter $\alpha$ controls the mix.

How to apply SES step by step:

1. Initialize. Set the first forecast equal to the first observed value: $\hat{y}_1 = y_1$.
2. Choose a value for $\alpha$ (or optimize it later using error measures).
3. Compute each forecast by plugging the current actual value and the previous forecast into the formula.
4. Repeat for each time period in your series.

Suppose your first three observations are 100, 110, and 105, and $\alpha = 0.4$:

• $\hat{y}_1 = 100$ (initialization)
• $\hat{y}_2 = 0.4(100) + 0.6(100) = 100$
• $\hat{y}_3 = 0.4(110) + 0.6(100) = 104$
• $\hat{y}_4 = 0.4(105) + 0.6(104) = 104.4$

Notice how the forecast adjusts gradually toward each new observation rather than jumping to it.

Impact of the Smoothing Parameter

The value of $\alpha$ is the single most important choice you make with SES.

• High $\alpha$ (closer to 1): The forecast leans heavily on the most recent observation. This makes it responsive to sudden shifts in level, but also more sensitive to random noise.
• Low $\alpha$ (closer to 0): The forecast relies more on the accumulated history. This produces smoother, more stable forecasts, but it reacts slowly when the true level of the series changes.

A practical way to think about it: volatile data where the level genuinely shifts often may benefit from a higher $\alpha$, while stable data with lots of random noise usually calls for a lower $\alpha$.

Finding the optimal $\alpha$: Rather than guessing, you can try many values of $\alpha$ (from 0 to 1) and pick the one that minimizes a forecast error measure like the sum of squared errors (SSE) over your historical data. Most time series software does this automatically.

Accuracy of Exponential Smoothing Forecasts

To judge how well SES performs, you compare its forecasts against actual values using error measures:

1. Mean Absolute Error (MAE): The average of the absolute differences between actual values and forecasts. Easy to interpret since it's in the same units as your data.
2. Mean Squared Error (MSE): The average of the squared differences. Squaring penalizes large errors more heavily than small ones, so MSE is useful when big misses are especially costly.
3. Mean Absolute Percentage Error (MAPE): The average of the absolute percentage errors. Useful for comparing accuracy across series with different scales (e.g., a series in the hundreds vs. one in the thousands).

For all three, lower values mean better accuracy. You can also use these measures to compare SES against other methods like a simple moving average. If SES produces a lower MAE than a 5-period moving average on the same data, that's evidence SES is the better fit for that particular series.