Choosing The Simple Smoothing Parameter

What the Smoothing Parameter Controls

Simple exponential smoothing represents a series with a changing level but no explicit trend or seasonal component. Its smoothing parameter controls how much the latest observation changes that level. This creates a practical trade-off: a level that responds quickly to new information may also react sharply to temporary noise, while a steadier level may take longer to reflect a genuine change.

Definition: Suppose observations $y_1,y_2,\ldots$ arrive at equally spaced times, and $l_t$ is the smoothed level after observing $y_t$. The one-step-ahead forecast made before observing $y_t$ is $ \hat{y}_{t|t-1}=l_{t-1} $. The simple exponential smoothing update is $$ l_t=\alpha y_t+(1-\alpha)l_{t-1}, $$ where $0<\alpha\leq 1$ is the smoothing parameter. Thus, the new level is a weighted average of the latest observation and the previous level.

The interpretation depends on the weights: $\alpha$ is the weight on the new observation, and $1-\alpha$ is the weight on the previous level. A larger $\alpha$ gives the latest observation more influence. At $\alpha=1$, the updated level equals the latest observation. A value close to zero makes each update smaller; the limiting case $\alpha=0$, though excluded from the usual parameter range, would leave the level unchanged.

This interpretation assumes the series is being modeled as a level-only process: observations are taken at regular intervals, and there is no separate trend or seasonal update in this equation. If the data have a persistent trend or seasonality, a level-only smoother may not represent those features. Choosing $\alpha$ cannot, by itself, add components that the model does not contain.

Comparing Updates for Different Parameter Values

The following example isolates the arithmetic of the update. Both candidate smoothers start from the same level and receive the same observations, so any difference in the resulting levels is due to the parameter value.

Worked Example: Two Parameters on the Same Series

A hypothetical weekly service-volume series has an initial smoothed level $l_0=100$. The next three observed volumes are $104$, $98$, and $110$. Compare the updates for $\alpha=0.2$ and $\alpha=0.8$, and state the forecast for the week after the third observation.

Method. Apply the update in time order, using each newly calculated level as the previous level in the next update. The forecast for the next week is the final smoothed level.

For $\alpha=0.2$: the weight on the previous level is $1-0.2=0.8$.

At time 1: $$ l_1=0.2(104)+0.8(100)=20.8+80=100.8. $$ At time 2: $$ l_2=0.2(98)+0.8(100.8)=19.6+80.64=100.24. $$ At time 3: $$ l_3=0.2(110)+0.8(100.24)=22+80.192=102.192. $$

For $\alpha=0.8$: the weight on the previous level is $1-0.8=0.2$.

At time 1: $$ l_1=0.8(104)+0.2(100)=83.2+20=103.2. $$ At time 2: $$ l_2=0.8(98)+0.2(103.2)=78.4+20.64=99.04. $$ At time 3: $$ l_3=0.8(110)+0.2(99.04)=88+19.808=107.808. $$

The forecasts for the week after the third observation are $102.192$ for $\alpha=0.2$ and $107.808$ for $\alpha=0.8$. The larger parameter responds more to both the fall to $98$ and the rise to $110$, so its level moves more sharply. The check is that each new level lies between its latest observation and previous level, as a weighted average with weights between zero and one must.

Common mistake avoided: Do not interpret the result as a simple average of all three observations. Each update uses the current observation and the immediately preceding smoothed level; the starting level therefore continues to affect the calculation.

Response to a Sudden Persistent Change

A useful way to compare parameter values is to imagine that observations change suddenly and then remain at a new constant value. A small parameter moves toward the new value gradually. A large parameter moves toward it more quickly.

Derivation: Suppose the level before a change is $L$, and every subsequent observation equals a new constant value $C$. Define the remaining gap after update $k$ as $d_k=C-l_k$. From the smoothing update, $$ d_k=C-\{\alpha C+(1-\alpha)l_{k-1}\} =(1-\alpha)(C-l_{k-1})=(1-\alpha)d_{k-1}. $$ Since $d_0=C-L$, repeated substitution gives $$ d_k=(1-\alpha)^k(C-L). $$ Thus, after $k$ updates, the fraction of the original gap still remaining is $(1-\alpha)^k$. A larger $\alpha$ makes this fraction smaller for any fixed positive $k$.

This result describes response speed when the new observations persist at the same value. It does not say that a large parameter is always preferable: if an observation is only a temporary spike, fast movement toward it may be undesirable.

Worked Example: Measuring Response to a Step Change

A hypothetical claim-intake process has a smoothed level of $50$ claims per week. A process change causes the observed weekly count to become $70$, and it remains at $70$ for four weeks. Compare the levels for $\alpha=0.25$ and $\alpha=0.75$. Give the forecast for the week after the fourth updated observation.

Method. Starting from $l_0=50$, apply the same update four times for each parameter. The observation is $70$ at every update.

For $\alpha=0.25$: the previous-level weight is $0.75$.

$$ \begin{aligned} l_1&=0.25(70)+0.75(50)=17.5+37.5=55,\\ l_2&=0.25(70)+0.75(55)=17.5+41.25=58.75,\\ l_3&=0.25(70)+0.75(58.75)=17.5+44.0625=61.5625,\\ l_4&=0.25(70)+0.75(61.5625)=17.5+46.171875=63.671875. \end{aligned} $$

For $\alpha=0.75$: the previous-level weight is $0.25$.

$$ \begin{aligned} l_1&=0.75(70)+0.25(50)=52.5+12.5=65,\\ l_2&=0.75(70)+0.25(65)=52.5+16.25=68.75,\\ l_3&=0.75(70)+0.25(68.75)=52.5+17.1875=69.6875,\\ l_4&=0.75(70)+0.25(69.6875)=52.5+17.421875=69.921875. \end{aligned} $$

The respective forecasts for the following week are $63.671875$ and $69.921875$ claims. For $\alpha=0.25$, the remaining gap from $70$ is $70-63.671875=6.328125$. The derived formula gives $0.75^4(70-50)=0.31640625(20)=6.328125$, confirming the update calculations. For $\alpha=0.75$, the gap is $70-69.921875=0.078125$, also equal to $0.25^4(20)=0.078125$.

The larger parameter has nearly caught up after four observations, while the smaller parameter retains a substantial gap. This comparison concerns adjustment speed, not which parameter would produce more accurate forecasts on every data set.

Common mistake avoided: Do not calculate the response using the new observation alone. The remaining gap is multiplied by $1-\alpha$ at each update; it is not reduced by a fixed number of units each week.

Responsiveness Can Also Mean Sensitivity to Noise

A sudden change that persists is not the only reason a level moves. A single unusually high or low observation also enters the update with weight $\alpha$. If the following observation returns to its previous value, a high-$\alpha$ level can reverse sharply. The next example separates a one-time spike from a persistent change.

Worked Example: A One-Time Spike Followed by a Return

A hypothetical weekly workload has initial level $l_0=200$. The next three observations are $202$, $240$, and $201$. Compare the level after each observation for $\alpha=0.2$ and $\alpha=0.8$, then give each forecast after the third observation.

Method. Update in sequence. The value $240$ is a one-time spike in the data presented: the next observation is close to the initial level. Do not treat the spike as if it were known in advance to represent a lasting shift.

For $\alpha=0.2$: $$ \begin{aligned} l_1&=0.2(202)+0.8(200)=40.4+160=200.4,\\ l_2&=0.2(240)+0.8(200.4)=48+160.32=208.32,\\ l_3&=0.2(201)+0.8(208.32)=40.2+166.656=206.856. \end{aligned} $$

For $\alpha=0.8$: $$ \begin{aligned} l_1&=0.8(202)+0.2(200)=161.6+40=201.6,\\ l_2&=0.8(240)+0.2(201.6)=192+40.32=232.32,\\ l_3&=0.8(201)+0.2(232.32)=160.8+46.464=207.264. \end{aligned} $$

The forecasts after the third observation are $206.856$ and $207.264$, respectively. The high-$\alpha$ level jumped to $232.32$ when the spike arrived, then fell by $25.056$ when the observation returned to $201$. The low-$\alpha$ level rose more moderately and then fell by $1.464$. The final levels happen to be close, but their paths differ substantially; a forecast made immediately after the spike would have been much more different.

Check. For the high parameter, the third update is $0.8(201)+0.2(232.32)=207.264$, which lies between $201$ and $232.32$. This verifies that the update uses the prior level and the new observation with the stated weights.

Common mistake avoided: A sharp response to an observation is not proof that the underlying process has permanently changed. Interpret the observed pattern before treating a high parameter as an advantage.

Choosing Between Responsiveness and Stability

The parameter should be interpreted in the context of the forecasting task. If the process can change quickly and those changes are expected to persist, a larger parameter allows the level to adapt sooner. If observations contain substantial short-term noise and the underlying level is comparatively stable, a smaller parameter reduces the impact of each individual observation.

Parameter choice	Effect on the level	Potential advantage	Potential cost
Small $\alpha$	Each new observation has modest influence	Less movement in response to a single unusual value	More lag after a genuine persistent change
Large $\alpha$	Each new observation has strong influence	Faster response to a lasting change	Greater sensitivity to temporary spikes and dips
$\alpha=1$	The updated level equals the latest observation	Immediate tracking of the latest value	No smoothing of individual observations

These are tendencies, not guarantees about forecast accuracy. A parameter that tracks a persistent shift well may perform poorly when the same movement is random noise. In an applied comparison, forecast errors should be evaluated using an appropriate procedure and data set; the parameter with the largest value is not automatically the best. The earlier material on forecast errors provides measures for comparing forecasts, but the interpretation of $\alpha$ comes directly from its role in the update.

A Common Exam Trap: Reversing the Weights

The most direct calculation error is to place $\alpha$ on the previous level instead of the new observation. This reverses the meaning of small and large parameter values.

Common mistake: If $l_{t-1}=100$, $y_t=120$, and $\alpha=0.25$, a reversed calculation gives $0.25(100)+0.75(120)=115$. That is incorrect for the stated convention. The correct update is $$ l_t=0.25(120)+0.75(100)=30+75=105. $$ A small parameter moves the level only one quarter of the way from $100$ toward $120$, not three quarters of the way.

A useful check is to rewrite the update as $$ l_t-l_{t-1}=\alpha(y_t-l_{t-1}). $$ The change in the level is the fraction $\alpha$ of the current forecast error. If the new observation exceeds the old level, the updated level must rise; if it is below, the updated level must fall. For $0<\alpha\leq1$, it cannot move past the observation in a single update.

Check Your Understanding

For each problem, use the stated update convention and distinguish a response to a persistent change from a response to a single unusual observation.

A level is $80$, the next observation is $90$, and $\alpha=0.3$. Calculate the updated level and identify the weight applied to the previous level. (Answer: Updated level $83$; previous-level weight $0.7$.)
Two smoothers start at $120$ and observe $100$. Calculate the updated levels for $\alpha=0.2$ and $\alpha=0.6$. Which responds more to the new observation? (Answer: $116$ and $108$, respectively; $\alpha=0.6$ responds more.)
A process starts with level $40$, then records three consecutive observations of $60$. For $\alpha=0.5$, calculate each updated level and the forecast after the third observation. (Answer: Levels $50$, $55$, and $57.5$; next forecast $57.5$.)
A level is $30$, $\alpha=0.25$, and two observations arrive: $50$ followed by $30$. Calculate the two updated levels. What does the second update illustrate? (Answer: Levels $35$ and $33.75$; the level reverses partly toward the old value after the one-time high observation.)
A level starts at $0$, and two consecutive observations are both $100$. Compare the level after the second update for $\alpha=0.1$ and $\alpha=0.9$. Give the amount of the original gap from $100$ that remains in each case. (Answer: Levels $19$ and $99$, respectively; remaining gaps $81$ and $1$.)

Parameter choice	Effect on the level	Potential advantage	Potential cost
Small \(\alpha\)	Each new observation has modest influence	Less movement in response to a single unusual value	More lag after a genuine persistent change
Large \(\alpha\)	Each new observation has strong influence	Faster response to a lasting change	Greater sensitivity to temporary spikes and dips
\(\alpha=1\)	The updated level equals the latest observation	Immediate tracking of the latest value	No smoothing of individual observations

What You'll Learn