Holt’s Linear Trend Smoothing

Why Add a Trend Component?

Simple exponential smoothing updates a level to follow observations but does not represent a persistent upward or downward movement separately. When a series changes steadily over time, a level-only forecast can lag: the latest observation may be above the smoothed level because the series has been rising. Holt’s linear trend method addresses this by tracking two quantities: a smoothed level and a smoothed trend.

The method applies to observations \(y_1,y_2,\ldots\) recorded at equally spaced times. Its additive trend represents a change in the series per time period, in the same units as the observations. The basic form does not include a seasonal component and is most naturally interpreted when a local linear trend is a reasonable short-term description. The update calculations themselves are deterministic once the observations, smoothing parameters, and initial components are specified.

Definition: At time \(t\), the level \(l_t\) is the current smoothed estimate of the series’ local level, and the trend \(b_t\) is the current smoothed estimate of its change per period. The level smoothing parameter \(\alpha\) and trend smoothing parameter \(\beta\) satisfy \(0\leq\alpha\leq1\) and \(0\leq\beta\leq1\). They control how strongly the updates respond to new information.

The Update Equations

Before observing \(y_t\), the previous level and trend together give the one-step forecast. The level update then combines the new observation with that forecast. The trend update combines the latest change in smoothed level with the previous trend.

\widehat{y}_{t\mid t-1}=l_{t-1}+b_{t-1}

l_t=\alpha y_t+(1-\alpha)(l_{t-1}+b_{t-1})

b_t=\beta(l_t-l_{t-1})+(1-\beta)b_{t-1}

Here \(\widehat{y}_{t\mid t-1}\) denotes the forecast of \(y_t\) made using information available through time \(t-1\). The observation \(y_t\) updates the components to \(l_t\) and \(b_t\). With those updated components, the forecast for the next observation is \(\widehat{y}_{t+1\mid t}=l_t+b_t\). A trend can be negative, zero, or positive.

The parameter \(\alpha\) controls the level’s response to the gap between the observed value and its one-step forecast. A larger \(\alpha\) gives more weight to the new observation. The parameter \(\beta\) controls how quickly the trend estimate responds to the latest change in level. A larger \(\beta\) gives more weight to \(l_t-l_{t-1}\); it does not directly give more weight to the raw difference \(y_t-y_{t-1}\).

Why the level update has this form: The one-step prediction before seeing \(y_t\) is \(l_{t-1}+b_{t-1}\). The forecast error is therefore \(e_t=y_t-(l_{t-1}+b_{t-1})\). Rewriting the level equation gives \(l_t=(l_{t-1}+b_{t-1})+\alpha e_t\): the level starts at the predicted value and moves a fraction \(\alpha\) of the way toward the new observation. The trend equation smooths the resulting level change against the previous trend, so a single unusual level movement need not become the entire trend estimate.

To apply the equations, initial values \(l_1\) and \(b_1\) must be supplied or obtained through an initialization procedure. For the calculations below, they are given as part of the problem. Keep track of the time index: at each period, first forecast using the previous components, then update the level, then update the trend.

Forecast the current observation.
Add the previous level and trend: \(l_{t-1}+b_{t-1}\).

Update the level.
Weight the observed value by \(\alpha\) and the one-step forecast by \(1-\alpha\).

Update the trend.
Weight the change in level by \(\beta\) and the previous trend by \(1-\beta\).

Worked Calculations

Worked Example: Updating Both Components During a Steady Rise

A hypothetical monthly portfolio index has observations \(y_1=100\), \(y_2=112\), and \(y_3=123\). Use the supplied initial components \(l_1=100\), \(b_1=10\), with \(\alpha=0.50\) and \(\beta=0.25\). Find the level and trend after months 2 and 3, then the one-step forecast for month 4.

Method. Apply Holt’s updates in order at each time. The inputs and both smoothing parameters are given, and all periods are one month long.

Month 2 forecast: \(\widehat{y}_{2\mid1}=l_1+b_1=100+10=110\).

Month 2 level:

$$ l_2=0.50(112)+0.50(110)=56+55=111 $$

Month 2 trend:

$$ b_2=0.25(111-100)+0.75(10)=0.25(11)+7.5=2.75+7.5=10.25 $$

Month 3 forecast: \(\widehat{y}_{3\mid2}=l_2+b_2=111+10.25=121.25\).

Month 3 level:

$$ l_3=0.50(123)+0.50(121.25)=61.5+60.625=122.125 $$

Month 3 trend:

$$ b_3=0.25(122.125-111)+0.75(10.25)=0.25(11.125)+7.6875=2.78125+7.6875=10.46875 $$

The month 4 one-step forecast is \(l_3+b_3=122.125+10.46875=132.59375\), or 132.59 to two decimal places. The updated trend remains close to the initial trend of 10, while incorporating the latest change in level. As a check, the month 3 level lies between its forecast, 121.25, and observation, 123, as expected for \(\alpha=0.50\).

Common mistake avoided: the month 3 level update uses the month 3 forecast of 121.25, not the previous level of 111 by itself. Omitting the trend would use the wrong comparison value.

Worked Example: A Larger Level Response Than Trend Response

A hypothetical monthly claim count is processed with initial components \(l_1=200\) and \(b_1=8\). The next observations are \(y_2=210\) and \(y_3=220\), with \(\alpha=0.30\) and \(\beta=0.60\). Calculate both components after month 3 and the forecast for month 4.

Month 2 forecast: \(\widehat{y}_{2\mid1}=200+8=208\).

Month 2 level: \(l_2=0.30(210)+0.70(208)=63+145.6=208.6\).

Month 2 trend: \(b_2=0.60(208.6-200)+0.40(8)=0.60(8.6)+3.2=5.16+3.2=8.36\).

Month 3 forecast: \(\widehat{y}_{3\mid2}=208.6+8.36=216.96\).

Month 3 level: \(l_3=0.30(220)+0.70(216.96)=66+151.872=217.872\).

Month 3 trend: \(b_3=0.60(217.872-208.6)+0.40(8.36)=0.60(9.272)+3.344=5.5632+3.344=8.9072\).

The forecast for month 4 is \(217.872+8.9072=226.7792\), or 226.78 to two decimal places. The level smoothing parameter is smaller than the trend smoothing parameter, so the new observation has a more restrained direct effect on the level, while the trend responds relatively quickly to the change in level. The calculations are internally consistent: the month 3 level, 217.872, is between the forecast 216.96 and observation 220.

Common mistake avoided: \(\beta=0.60\) is used in the trend update, not the level update. The month 3 level uses \(\alpha=0.30\).

Worked Example: Updating a Positive Trend After a Large Observation

For a hypothetical weekly service volume, the supplied initial components are \(l_1=50\) and \(b_1=4\). The next observations are \(y_2=55\), \(y_3=59\), and \(y_4=66\). Use \(\alpha=0.80\) and \(\beta=0.20\). Find the components after week 4 and the one-step forecast for week 5.

Week 2: The forecast is \(50+4=54\). The level is \(l_2=0.80(55)+0.20(54)=44+10.8=54.8\). The trend is \(b_2=0.20(54.8-50)+0.80(4)=0.96+3.2=4.16\).

Week 3: The forecast is \(54.8+4.16=58.96\). The level is \(l_3=0.80(59)+0.20(58.96)=47.2+11.792=58.992\). The trend is \(b_3=0.20(58.992-54.8)+0.80(4.16)=0.20(4.192)+3.328=0.8384+3.328=4.1664\).

Week 4: The forecast is \(58.992+4.1664=63.1584\). The level is \(l_4=0.80(66)+0.20(63.1584)=52.8+12.63168=65.43168\). The trend is \(b_4=0.20(65.43168-58.992)+0.80(4.1664)=0.20(6.43968)+3.33312=1.287936+3.33312=4.621056\).

The week 5 one-step forecast is \(65.43168+4.621056=70.052736\), or 70.05 to two decimal places. The large week 4 observation pulls the level upward, but the trend moves less sharply because \(\beta=0.20\). The week 4 level is between its forecast, 63.1584, and observation, 66, which checks the level update for \(\alpha=0.80\).

Common mistake avoided: the level is not set equal to 66 just because that is the latest observation. The update retains 20% of the prior one-step forecast.

Common Mistakes and Interpretation

Holt’s method is straightforward to calculate, but several details are easy to mix up under time pressure. The level and trend equations use different inputs and different parameters, so label each intermediate quantity before substituting.

Using the previous level alone in the forecast. The one-step forecast is \(l_{t-1}+b_{t-1}\), not \(l_{t-1}\). The trend is part of the prediction.
Using the raw observation difference as the trend update. The trend update uses \(l_t-l_{t-1}\), not \(y_t-y_{t-1}\). Observations are first incorporated into the smoothed level.
Swapping \(\alpha\) and \(\beta\). The level equation uses \(\alpha\); the trend equation uses \(\beta\).
Updating the trend before the level. Compute \(l_t\) first because \(b_t\) depends on the newly updated level.
Ignoring the time unit. The trend is measured in observation units per period. If observations are weekly, \(b_t\) is per week; do not interpret it as a monthly change without an appropriate conversion.

The method’s usefulness depends on the pattern being modeled. An additive level and trend are appropriate when the series has a changing level and a reasonably interpretable change per period. The method does not by itself represent recurring seasonal movements. A trend that persists indefinitely may also produce implausible long-range values, so a good short-term fit should not automatically be treated as a reliable long-term projection.

For an actuarial application, one might hypothetically track the weekly volume of a particular type of service request. A positive trend would indicate increasing smoothed volume per week, not a percentage growth rate. If the scale of fluctuations increases with the level, or if the data have a seasonal pattern, the basic additive form may not be a suitable description. Holt’s updates are therefore a forecasting tool, not a substitute for checking the time scale and structure of the series.

Key takeaway: Holt’s method keeps two linked summaries. First forecast with the previous level plus trend; then update the level using \(\alpha\), and update the trend using \(\beta\) and the change in smoothed level. Always preserve that order.

Check Your Understanding

For each calculation, use equally spaced observations and the supplied initial components. Check that you use the level update before the trend update.

At time \(t-1\), \(l_{t-1}=80\) and \(b_{t-1}=5\). The new observation is \(y_t=88\), with \(\alpha=0.40\) and \(\beta=0.20\). Find the one-step forecast, \(l_t\), and \(b_t\). (Answer: forecast \(85\), \(l_t=86.2\), \(b_t=5.24\).)
Given \(l_{t-1}=120\), \(b_{t-1}=3\), \(y_t=128\), \(\alpha=0.50\), and \(\beta=0.50\), find the updated components and the next one-step forecast. (Answer: \(l_t=125.5\), \(b_t=4.25\), next forecast \(129.75\).)
Initial components are \(l_1=40\), \(b_1=2\), with \(\alpha=0.25\) and \(\beta=0.50\). For observations \(y_2=45\) and \(y_3=49\), find \(l_3\), \(b_3\), and the next one-step forecast. (Answer: \(l_3=46.09375\), \(b_3=2.859375\), next forecast \(48.953125\).)
At time \(t-1\), \(l_{t-1}=300\) and \(b_{t-1}=-6\). Given \(y_t=288\), \(\alpha=0.60\), and \(\beta=0.25\), calculate \(l_t\), \(b_t\), and the next one-step forecast. (Answer: \(l_t=290.4\), \(b_t=-6.9\), next forecast \(283.5\).)
In Holt’s equations, what does a larger \(\alpha\) change directly, and what does a larger \(\beta\) change directly? (Answer: A larger \(\alpha\) gives more weight to the new observation in the level update; a larger \(\beta\) gives more weight to the latest change in smoothed level in the trend update.)

What You'll Learn