1. What Is Smoothing?

Time-series data often contains:

• a useful underlying signal,
• short-term random fluctuations,
• measurement noise,
• unusual events.

Smoothing tries to reduce the effect of short-term noise so that the underlying level, trend, or seasonal pattern becomes easier to identify.

Suppose daily demand is:

The values fluctuate, but the underlying level appears to be around 101.

Smoothing asks:

What is the current underlying level after ignoring some of the random movement?

2. Why Not Use the Latest Observation Directly?

If today’s demand is 130 because of an unusual event, predicting 130 for tomorrow may be too reactive.

On the other hand, using the average of the entire historical dataset may respond too slowly when demand genuinely changes.

We need a balance between:

• trusting the newest observation,
• retaining information from previous observations.

This is the central idea behind exponential smoothing.

3. Moving-Average Intuition

A simple approach is to average the most recent \(w\) observations.

For a three-day moving average:

If recent demand is:

then:

This reduces noise, but it has two limitations:

1. Every observation inside the window receives equal weight.
2. Observations suddenly receive zero weight when they leave the window.

Exponential smoothing uses weights that decrease gradually instead.

4. Simple Exponential Smoothing

Simple exponential smoothing maintains one quantity:

The update equation is:

where:

• \(Y_t\): new observation
• \(\ell_{t-1}\): previous level estimate
• \(\alpha\): smoothing parameter
• \(0\leq\alpha\leq1\)

5. Weighted-Average Intuition

The new level is a compromise between:

• what we just observed,
• what we previously believed.

If \(\alpha=0.3\), the new level uses:

• 30% of the new observation,
• 70% of the previous estimate.

6. Walk-Toward-the-Observation Intuition

The same equation can be rewritten as:

The quantity:

is the latest forecast error.

Therefore:

Start at the old level and move \(\alpha\) of the way toward the new observation.

7. Numerical Example

Suppose:

The error is:

Move 30% of that distance:

The new observation was 120, but the estimated level moves only from 100 to 106.

The model treats part of the jump as possible noise.

If the next observation is 90:

The estimate moves downward but does not immediately jump to 90.

8. Meaning of \(\alpha\)

Small \(\alpha\)

For example:

• trusts history more,
• produces a smoother line,
• reacts slowly,
• is less sensitive to noise,
• may miss sudden genuine changes.

Large \(\alpha\)

For example:

• trusts the latest observation,
• reacts quickly,
• produces less smoothing,
• is more sensitive to noise and outliers.

The choice is a bias-responsiveness tradeoff.

9. Why Is It Called “Exponential”?

Repeatedly substitute the previous level:

Since:

we obtain:

The weights are:

They decline geometrically, or exponentially, as observations become older.

Older observations are not discarded abruptly. Their influence gradually becomes smaller.

10. Forecasting With SES

Simple exponential smoothing has only a level state. It has no trend or seasonal state.

Therefore, all future forecasts are equal to the latest estimated level:

for:

If:

then:

SES produces a flat forecast path.

Use SES when the series has:

• a changing local level,
• no persistent trend,
• no seasonality.

11. Why SES Cannot Handle Trend

Suppose demand is:

SES updates the level, but its future forecasts remain flat.

It may estimate the current level near 130 or 135, but it does not know that demand is increasing by approximately 10 every period.

To model that movement, we need a separate trend state.

12. Holt’s Method

Holt’s method maintains two states:

The equations are:

The forecast is:

13. Holt’s Method in Small Steps

Suppose:

Step 1: Predict today before seeing it

The old level was 100 and the old trend was 5:

The model expected today’s value to be 105.

Step 2: Update the level

Blend the actual observation with the projected level:

Step 3: Observe the new level change

The newly observed growth is 7.8.

Step 4: Update the trend

Blend the new growth with the old trend:

Step 5: Forecast

One step ahead:

Two steps ahead:

Unlike SES, Holt’s method produces a trending future path.

14. Meaning of \(\alpha\) and \(\beta\)

• \(\alpha\) controls how quickly the level changes.
• \(\beta\) controls how quickly the estimated trend changes.

A large \(\beta\) allows the slope to change rapidly. A small \(\beta\) assumes that the trend evolves gradually.

15. Why Holt’s Method Cannot Handle Seasonality

Suppose bike demand has:

• moderate weekday demand,
• high Saturday demand,
• high Sunday demand.

Holt’s method can model the overall level and growth, but it has no memory of which day of the week is being forecast.

We therefore add a seasonal state.

16. Holt-Winters Method

Additive Holt-Winters maintains three components:

For season length \(m\):

The forecast is:

17. Holt-Winters Intuition

Before updating the level:

1. Remove the known seasonal effect.
2. Estimate the underlying level.
3. Update the trend.
4. Estimate whether this seasonal position was stronger or weaker than usual.
5. Store that updated seasonal effect for the next cycle.

For daily observations with weekly seasonality:

Each weekday has its own seasonal memory.

For hourly demand:

• \(m=24\): hour-of-day seasonality.
• \(m=168\): hour-and-weekday seasonality.

18. Small Holt-Winters Example

Suppose:

The seasonal effect for this weekday is:

Today’s observed demand is:

Use:

Step 1: Remove seasonality

The underlying nonseasonal value is approximately 105.

Step 2: Update level

Step 3: Update trend

Step 4: Update seasonality

The new seasonal evidence is:

Blend it with the previous seasonal effect:

The weekday effect changes from 15 to 15.9.

Step 5: Forecast the same weekday next week

19. Additive Versus Multiplicative Seasonality

Additive seasonality

Use it when seasonal effects remain approximately constant in units.

Example:

Every Saturday adds approximately 50 rentals.

Multiplicative seasonality

Use it when seasonal effects grow with the level.

Example:

Saturday demand is approximately 30% above normal.

If the seasonal fluctuations increase as the series grows, multiplicative seasonality may be more appropriate.

20. The Three Methods as a Story

The progression is:

Final Intuition

Exponential smoothing is a controlled memory system.

Each new observation asks:

• Should the estimated level move?
• Has the trend changed?
• Was this seasonal position stronger or weaker than usual?

The smoothing parameters determine how much the model changes its beliefs after receiving new evidence.