Holt Forecasts Across Multiple Horizons

From a Current Trend to Several Forecast Horizons

A one-period forecast answers only the nearest question: what value does the current level and trend imply for the next period? In many actuarial and business settings, decisions instead depend on several future periods. A staffing plan may need projected weekly claim counts for the next month; a hypothetical insurer may want a sequence of monthly service-volume forecasts. Ordinary Holt linear trend smoothing gives a direct way to extend its current level and trend to each horizon.

The central implication is simple but easy to misread: if the estimated trend is constant, each additional forecast period adds the same amount to the forecast. The total difference between a forecast and the current estimated level grows with the horizon, but the difference between adjacent forecast horizons does not.

Definition: At time $t$, let $l_t$ denote Holt’s estimated level and $b_t$ its estimated trend, measured in response units per time period. For an integer forecast horizon $h \geq 1$, the ordinary Holt point forecast is $$ \hat{y}_{t+h\mid t}=l_t+h b_t. $$ Here $h$ counts periods ahead from the same forecast origin $t$. The forecast at horizon one is $l_t+b_t$; the forecast at horizon two is $l_t+2b_t$.

The level and trend in this formula are the estimates available at the end of time $t$. They are held fixed while forecasts are generated from that origin. The trend is not re-estimated separately for each future horizon. This is an extrapolation rule for ordinary, undamped Holt smoothing: it assumes that the estimated linear trend continues at the same amount per period, with no seasonal adjustment.

The response and trend must use compatible units. If $l_t$ is measured in claims per week, then $b_t$ must be measured in claims per week of forecast horizon, and $h$ must count weeks. A six-month forecast from monthly data uses $h=6$, not $h=1$. Holt’s forecast is a point forecast, not a statement that the future observation must equal the forecast.

Why the Multi-Horizon Formula Holds

The formula follows by extending the current estimated level forward using the current trend at each step. In the forecast path, the estimated trend remains $b_t$, and the level advances by $b_t$ each period. Thus the first future level is $l_t+b_t$; the next adds $b_t$ again; and after $h$ steps the total addition is $h b_t$.

Start at the estimated level.
At the forecast origin, the current level is $l_t$.

Advance one period at a time.
Each forecast step adds the same estimated trend $b_t$, so after $h$ steps the total addition is $b_t+b_t+\cdots+b_t$, with $h$ terms.

Combine the additions.
The sum of $h$ identical additions is $h b_t$, giving $\hat{y}_{t+h\mid t}=l_t+h b_t$.

The behavior of forecast increments follows directly. For two successive horizons,

\hat{y}_{t+h+1\mid t}-\hat{y}_{t+h\mid t} =(l_t+(h+1)b_t)-(l_t+h b_t)=b_t.

Therefore, the forecast sequence has a constant first difference, equal to the estimated trend. Equivalently, its second differences are zero. The forecast’s displacement from the level is $h b_t$, which changes with $h$; the increment from one forecast horizon to the next is $b_t$, which does not. If $b_t>0$, forecasts rise by a fixed amount per period. If $b_t<0$, they fall by a fixed amount per period. If $b_t=0$, every horizon has the same forecast $l_t$.

Worked Forecasts at Multiple Horizons

Worked Example: A Positive Weekly Trend

A hypothetical claims unit has a Holt level estimate of $125$ claims per week and a trend estimate of $7$ claims per week per week at the end of week $t$. Calculate the forecasts one through four weeks ahead and interpret the increments.

Method. Use $\hat{y}_{t+h\mid t}=l_t+h b_t$ with $l_t=125$, $b_t=7$, and $h$ measured in weeks.

At horizon one, $$ \hat{y}_{t+1\mid t}=125+1(7)=132. $$ At horizon two, $$ \hat{y}_{t+2\mid t}=125+2(7)=125+14=139. $$ At horizon three, $$ \hat{y}_{t+3\mid t}=125+3(7)=125+21=146. $$ At horizon four, $$ \hat{y}_{t+4\mid t}=125+4(7)=125+28=153. $$

The sequence is $132,139,146,153$. Its successive increments are $139-132=7$, $146-139=7$, and $153-146=7$, matching the trend estimate. The four-week forecast is $28$ above the current level, because the trend is added four times; it is not $28$ above the one-week forecast.

Check. The four-week displacement is $4(7)=28$, and $125+28=153$, agreeing with the calculated forecast. The common mistake avoided here is treating the trend as the total increase over all horizons rather than an increase per period.

Worked Example: A Negative Monthly Trend

A hypothetical monthly service-volume series has current Holt estimates $l_t=68$ requests per month and $b_t=-1.5$ requests per month per month. Find the one-, four-, and eight-month forecasts, and state the change from the current level at each horizon.

Method. Apply the same ordinary Holt formula; a negative trend subtracts from the level at each future step.

For one month ahead, $$ \hat{y}_{t+1\mid t}=68+1(-1.5)=66.5. $$ For four months ahead, $$ \hat{y}_{t+4\mid t}=68+4(-1.5)=68-6=62. $$ For eight months ahead, $$ \hat{y}_{t+8\mid t}=68+8(-1.5)=68-12=56. $$

Relative to the current level of $68$, these forecasts are lower by $1.5$, $6$, and $12$ requests, respectively. The change between the one- and four-month forecasts is $62-66.5=-4.5$, which equals three additional periods times $-1.5$. From month four to month eight the change is $56-62=-6$, which equals four additional periods times $-1.5$.

Check. The sequence declines by $1.5$ for each single monthly step: the one-month forecast is $66.5$, and four more steps give $66.5+4(-1.5)=60.5$. That value is not the four-month forecast because this comparison starts at horizon one and advances four more periods to horizon five. The requested four-month forecast, $62$, is three steps after the one-month forecast: $66.5+3(-1.5)=62$. This indexing check prevents confusing a horizon with the number of increments between two horizons.

Worked Example: Finding the First Horizon Above a Capacity

A hypothetical weekly processing team has a Holt level estimate of $210$ cases per week and a trend estimate of $12$ cases per week per week. Its planned capacity is $300$ cases per week. Find the first whole-number forecast horizon at which the point forecast exceeds capacity.

Method. Set up the forecast inequality using $h$ weeks ahead: $$ 210+12h>300. $$ Subtracting $210$ gives $12h>90$, so $h>7.5$. Since forecast horizons are whole numbers of weeks, the first qualifying horizon is $h=8$.

At seven weeks, $$ \hat{y}_{t+7\mid t}=210+7(12)=210+84=294, $$ which is below capacity. At eight weeks, $$ \hat{y}_{t+8\mid t}=210+8(12)=210+96=306, $$ which exceeds capacity by $306-300=6$ cases per week.

Check. The forecast increases by $12$ per week, so the next forecast after $294$ is $306$. The first exceedance is therefore at week eight, not week seven. The common mistake avoided is rounding $7.5$ down or reporting a fractional week as the first whole-number forecast horizon.

Reading the Forecast Sequence Correctly

A multiple-horizon table is often the clearest way to communicate a Holt forecast. Each row should identify the horizon and the forecast measured from the same origin. For example, with $l_t=125$ and $b_t=7$, the one-week forecast is $132$, the two-week forecast is $139$, and the three-week forecast is $146$. These are all forecasts made at time $t$, not successive observed values or forecasts recalculated after new data arrive.

That distinction matters operationally. If a new observation becomes available at time $t+1$, Holt’s level and trend may be updated, and forecasts from the new origin can differ from the original forecast path. The formula here describes the forecasts made from a fixed origin using the estimates available at that origin; it does not assert that future updates will leave the estimates unchanged.

Quantity	Expression	Interpretation
Forecast at horizon $h$	$l_t+h b_t$	Point forecast $h$ periods after origin $t$
Displacement from current level	$h b_t$	Total trend-based change over $h$ periods
Increment between adjacent horizons	$b_t$	Change for one additional period

Ordinary Holt extrapolation can eventually produce implausible values in applications where the response is constrained, such as nonnegative counts or bounded rates. The multi-horizon formula itself still gives the specified point forecasts, but a user must judge whether continued linear extrapolation is appropriate for the decision and forecast range. It supplies no automatic cap and no prediction interval.

Common Horizon and Unit Errors

Using $h-1$ instead of $h$. The one-step forecast is $l_t+b_t$, not $l_t$. For the positive-trend example, the incorrect four-step calculation $125+3(7)=146$ is actually the three-step forecast. The correction is $125+4(7)=153$.
Confusing displacement with the adjacent increment. At horizon four, the forecast is $4b_t$ above the level, but it is only $b_t$ above the horizon-three forecast.
Changing the time unit without changing the horizon. If data are monthly, a forecast four periods ahead is four months ahead. Calling it a one-year forecast would require using $h=12$, assuming twelve monthly steps are intended.
Using a positive sign for a negative trend. The formula always uses $+h b_t$. When $b_t$ is negative, this addition reduces the level.
Re-estimating while holding the origin fixed. A fixed-origin forecast uses the same $l_t$ and $b_t$ at every horizon. Updating states after new observations describes forecasts from a later origin instead.

Key takeaway: For ordinary Holt smoothing, a forecast $h$ periods ahead is the current estimated level plus $h$ times the current estimated trend. Successive horizon forecasts differ by exactly one trend increment, while their total displacement from the level grows in proportion to the horizon.

Check Your Understanding

For each problem, use a fixed forecast origin and the ordinary Holt multi-horizon formula.

A hypothetical weekly series has level $34$ and trend $2.5$ units per week. Find the forecasts one, three, and six weeks ahead. (Answer: $36.5,\ 41.5,\ 49$, respectively.)
A monthly service series has level $510$ and trend $-14$ units per month. Find the forecasts two, five, and nine months ahead. (Answer: $482,\ 440,\ 384$, respectively.)
A weekly process has level $72.4$ and trend $0.8$. What is the first whole-number horizon at which the forecast is at least $80$? Verify the forecast one period earlier. (Answer: $h=10$; the forecast is $80.4$, while at $h=9$ it is $79.6$.)
A quarterly hypothetical portfolio count has level $240$ and trend $18$ units per quarter. Find the forecast two quarters ahead and one year ahead, then find the increase from the two-quarter forecast to the one-year forecast. (Answer: $276$, $312$, and an increase of $36$.)
From one fixed origin, a Holt forecast at horizon two is $96$, and at horizon seven it is $126$. Assuming the ordinary Holt forecast form, find the trend, current level, and forecast at horizon ten. (Answer: trend $6$ units per period, level $84$, and horizon-ten forecast $144$.)

Quantity	Expression	Interpretation
Forecast at horizon \(h\)	\(l_t+h b_t\)	Point forecast \(h\) periods after origin \(t\)
Displacement from current level	\(h b_t\)	Total trend-based change over \(h\) periods
Increment between adjacent horizons	\(b_t\)	Change for one additional period

What You'll Learn