Exponential smoothing in Excel: simple, double & triple

Exponential smoothing is the workhorse of short-term forecasting — and you can do every flavour of it in Excel. This guide walks through single smoothing by hand, double smoothing (Holt) for trended data, triple smoothing (Holt-Winters) for seasonal data, the built-in Data Analysis ToolPak, and the one-cell FORECAST.ETS function, with a worked example throughout.

Dr Waqas Rafique · Founder & CTO · PhD, Statistical Machine Learning
LinkedIn · About

1.What exponential smoothing actually is

Exponential smoothing forecasts the next value as a weighted average of every past value, where the weights shrink geometrically the further back you go. The most recent observation gets weight α, the one before it α(1−α), the one before that α(1−α)2, and so on. Because the weights never quite reach zero, the whole history contributes — but recent data dominates.

That single idea comes in three sizes, depending on what your data is doing:

• Single (simple) exponential smoothing — tracks the level only. Use it when the series wobbles around a roughly flat mean with no trend or season.
• Double exponential smoothing (Holt’s method) — adds a trend term so the forecast can slope up or down. Use it for data that drifts but has no seasonal pattern.
• Triple exponential smoothing (Holt-Winters) — adds a seasonal term so a repeating monthly/quarterly fingerprint is carried forward. Use it for revenue, demand, energy — anything with both trend and season.

The closely related moving average in Excel also smooths noise, but it weights the last k points equally and ignores everything older. Exponential smoothing is usually the better forecaster because it never fully forgets and reacts faster to genuine change.

2.Step 1 — Lay out the data

Open Excel. Put the period label in column A and the observed value in column B, sorted oldest at the top with no missing rows. For single and double smoothing you only need a handful of points to start; for the seasonal (triple) model you want at least two full cycles — 24 months for monthly data.

Reserve a cell for each smoothing parameter so you can tune them in one place: put α in F1, β in F2, and γ in F3. We’ll reference those absolute cells in every formula.

3.Step 2 — Single (simple) smoothing by formula

The single exponential smoothing recursion is one line:

You have to seed the very first smoothed value. The simplest, most common seed is just the first observation itself: S1 = Y1. With values in B2:B… and the smoothed series in column C:

C2:  =B2                          ' seed: first smoothed value = first actual
C3:  =$F$1*B3 + (1-$F$1)*C2       ' then drag this down

The one-step-ahead forecast for the next period is simply the last smoothed value: Forecast = St. With α in F1 = 0.3 and our worked numbers, the smoothing fills in like this:

Notice the smoothed line lags the actuals and is far less jumpy — exactly what you want for a noisy but flat series. If your data clearly trends upwards like this one, single smoothing will systematically lag behind, which is the cue to move to double smoothing.

4.Step 3 — Double smoothing (Holt) for trend

Holt’s method runs two recursions side by side: one for the level and one for the trend. Each has its own smoothing parameter.

Seed the level with the first observation and the trend with the first difference. With actuals in column B, level in C, trend in D:

C2:  =B2                                  ' level seed
D2:  =B3-B2                               ' trend seed (first step)
C3:  =$F$1*B3 + (1-$F$1)*(C2+D2)          ' level update
D3:  =$F$2*(C3-C2) + (1-$F$2)*D2          ' trend update

Drag C3:D3 down. The h-step-ahead forecast projects the trend forward:

So one month ahead is L + T; three months ahead is L + 3·T. This is what stops the forecast lagging on trended data. If you only ever need a clean trend line and not a probabilistic forecast, a chart trendline or a weighted moving average may be enough — but Holt is barely more work and forecasts properly.

5.Step 4 — Triple smoothing (Holt-Winters) for seasonality

Triple exponential smoothing adds a third recursion for the seasonal index, using parameter γ and a season length m (12 for monthly, 4 for quarterly). For the multiplicative form, which suits most revenue and demand data:

The forecast multiplies the projected level-plus-trend by the matching seasonal index:

This is exactly the Holt-Winters method — the initialisation of the seasonal indices and the Solver-based parameter tuning are involved enough to deserve their own walk-through. We cover the full build, cell by cell, in the dedicated Holt-Winters in Excel tutorial. If your data is seasonal, start there. For seasonality detection and seasonal indices specifically, see seasonal forecasting in Excel.

6.Step 5 — The Data Analysis ToolPak shortcut

If you don’t want to type the recursion at all, Excel ships a point-and-click version for single smoothing. First enable it: File → Options → Add-ins → Manage: Excel Add-ins → Go → tick “Analysis ToolPak”. Then:

1. Go to Data → Data Analysis → Exponential Smoothing.
2. Set Input Range to your value column (B2:B…).
3. Set the Damping factor. This is the trap: the ToolPak asks for the damping factor, which is 1−α. For α=0.3, enter 0.7.
4. Tick Chart Output for an instant actual-vs-smoothed chart, and Standard Errors if you want them.

The ToolPak only does single smoothing — it has no trend or seasonal option. For those you either build the recursions above or jump to FORECAST.ETS.

7.Step 6 — FORECAST.ETS in one cell

Excel 2016 and later (and Microsoft 365) include FORECAST.ETS, which fits a full triple exponential smoothing model — level, trend and automatically-detected seasonality — and returns the forecast for a future date in a single cell. No recursion, no parameter cells.

With dates in A2:A40, values in B2:B40, and the date you want to forecast in A41:

=FORECAST.ETS(A41, $B$2:$B$40, $A$2:$A$40)         ' point forecast
=FORECAST.ETS.CONFINT(A41, $B$2:$B$40, $A$2:$A$40) ' +/- prediction interval
=FORECAST.ETS.SEASONALITY($B$2:$B$40, $A$2:$A$40)  ' detected season length

Leave the seasonality argument blank and Excel detects the cycle for you; set it to 1 to switch seasonality off (which degrades it to Holt-style double smoothing). Wrap the point forecast and the confidence interval together to draw a fan chart. This is the fastest honest forecast Excel can give you.

Choosing alpha, beta and gamma

The smoothing parameters all live between 0 and 1 and control how fast each component reacts to new data. Small values smooth heavily and respond slowly; large values track the latest point closely but pass more noise through. There are three sensible ways to set them:

Option 1 — Solver. Build a cell that computes mean absolute percentage error (MAPE) of the in-sample fit, then run Data → Solver to minimise it by changing F1:F3, with constraints 0 < α, β, γ < 1. Best accuracy, two clicks once set up.

Option 2 — grid search. Try each parameter over {0.1, 0.3, 0.5, 0.7, 0.9} and keep the combination with the lowest error on a held-out tail of the data. This is essentially what auto-tuners do internally.

Option 3 — defaults. α=0.3 for single; α=0.3, β=0.1 for Holt; α=0.3, β=0.1, γ=0.3 for Holt-Winters. Serviceable when you’re in a hurry.

Whichever you pick, always hold out the last 10–20% of the series to test on. A model tuned to fit the entire past usually over-fits and forecasts worse.

Common mistakes

1. Using single smoothing on trended data. If the series clearly slopes, single smoothing lags permanently behind. Move up to Holt’s double smoothing.

2. Forgetting the damping-factor flip. The ToolPak wants 1−α, not α. Enter 0.7 for an alpha of 0.3, or your smoothing will be far heavier than you intended.

3. Multiplicative seasonality with zeros. The multiplicative Holt-Winters form divides by the seasonal index and level — if your data has zeros or negatives, switch to the additive form or it returns nonsense.

4. Re-tuning parameters every period. Fix α/β/γ for a given series. Re-optimising them every month gives the illusion of a perfect fit and forecasts that wobble.

5. Forecasting through a structural break. A relaunch, acquisition or pricing change breaks the pattern. Smoothing will keep projecting the old shape; model the break or forecast only the post-break period.

The 30-second shortcut

If you’d rather not build any of these recursions, our free forecasting calculator runs auto-tuned exponential smoothing on any pasted column of values, in your browser, no signup. Same method, same maths, none of the cell-dragging.

Frequently asked questions