Skip to content
Français

Exponential smoothing (ETS, Holt-Winters)

  • Lesson 11
  • 40 min
  • Module 4: Statistical forecasters

Where we are

The backtest table has its first rows: the naive forecasts. They give every past day the same weight, or only the last one. In this lesson, we meet exponential smoothing, where recent days count more, and we add a trend and a season to it.

Plays from YouTube. Nothing loads from YouTube until you press play. Privacy

What you will be able to do

Make a forecast where recent days count more than old days, and add a weekly season to it. This is exponential smoothing (ETS). It beats every naive forecast in the backtest table, and it fits 50 products in under a minute.

NotebookDo this lesson in a notebookColab · Kaggle · GitHub

The notebook has every code cell, task and check of this lesson. It computes the fading weights and fits ETS to the parasol. Then it runs ETS in the course backtest and compares it with the moving average, product by product. Last, it shows what ETS does with two seasons. Its setup cells install the packages and download the Larkfield data, so it needs no local setup. About 40 minutes, with about 1 minute of waiting.

Colab and Kaggle open your own copy. They need a free Google or Kaggle account. Stuck? See the solution notebook.

Before you startWhat to know and to have readyAbout 40 minutes

What you need to know

What you need on your computer

Nothing: the notebook runs in Colab or Kaggle. To run it on your own computer, use Python 3.12 with Jupyter or VS Code. The notebook’s first cell installs the packages.

Tested with Python 3.12, pandas 2.3.3 and statsforecast 2.1.1, in Jupyter on macOS.

Think about your trip to work. How long will it take tomorrow? You do not take the mean of every trip of the last year. You also do not copy only today’s trip. You think: “This week it took about 35 minutes. Last month it was nearer 30, but the road works started since then.”

So yesterday’s trip counts most. Last week’s trips count a little less. Trips from months ago count almost nothing, but they are not forgotten. Your memory fades. That is the idea of this lesson.

The problem: two baselines, two different mistakes

Section titled “The problem: two baselines, two different mistakes”

In the last module, two baselines did best, each in its own way:

  • The 28-day moving average had the lowest WAPE, 43.5%. But it gives the same weight to every one of its 28 days, and no weight at all to day 29. It also gives the same number for every day ahead, so it forgets the weekly season.
  • The seasonal naive forecast (the same weekday last week) follows the weekly season. But it copies one single week, with all of its accidents. Its WAPE was 51.3%.

Kenji, the stock planner, wants both at once: “Use many days, so that one strange day does not fool you. Count the recent days more. And remember that Saturdays sell more than Wednesdays.”

Simple exponential smoothing keeps one number, the level: the forecaster’s best guess of a normal day now. Every day, it moves the level a part of the way toward the new value. The part is the smoothing value α (say “alpha”), between 0 and 1:

  • new level = old level + α × (today’s value − old level)

The forecast for every day ahead is the last level. “Exponential” means that the weight of a past day shrinks by the same factor, 1 − α, every day back. “Smoothing” means that the noise is averaged away.

A product sold 10, 12, 9, 15 and 11 units in five days. Take α = 0.3 and start the level at the first value, 10. Work out each step:

Day Sold Level before Miss (sold − level) Level after = before + 0.3 × miss
1 10 10.00 (the start)
2 12 10.00 +2.00 10.60
3 9 10.60 −1.60 10.12
4 15 10.12 +4.88 11.58
5 11 11.58 −0.58 11.41

The forecast for day 6, and for every day after it, is 11.41. statsforecast’s SimpleExponentialSmoothing(alpha=0.3) gives the same 11.41 on these five days.

Day 4 sold 15, but the level only moved to 11.58. One busy day moves the forecast only 30% of the way.

Write the level out, and each past day gets a weight: α for yesterday, α × (1 − α) for the day before, then α × (1 − α)², and so on. With α = 0.3, the weights are 0.3, 0.21, 0.147, 0.103 and 0.072. The last 7 days get 0.918 of the weight together.

  • α = 0.6
  • α = 0.3
  • α = 0.1
  • moving average (28 days)
The weight of each past day in the forecast. Exponential smoothing's weights fade into the past; a larger α fades faster. The 28-day moving average gives each of its days 1/28 = 0.036, and then nothing.
Text description of the chart

The chart shows the weight of each past day, from yesterday (day 1) to 28 days back. With α = 0.6, yesterday gets 0.6, the day before 0.24, and the weights are almost 0 after 6 days. The last 7 days get 0.998 of the weight. With α = 0.3, the weights start at 0.3 and fall to almost 0 after about 14 days. The last 7 days get 0.918. With α = 0.1, the weights start at 0.1 and fall slowly, to 0.006 on day 28. The last 7 days get 0.522. The 28-day moving average is a flat dashed line at 0.036 for all 28 days.

A large α (0.6) follows changes fast, but it also follows the noise. A small α (0.1) averages many days, but it reacts late when the level really changes. The forecaster finds the α that fits the series best, as you see below.

The level alone gives a flat forecast. Two more parts fix this, each with its own smoothing value:

  • Holt’s method adds a trend: how much the level grows or falls per day. The forecast becomes a sloped line.
  • Holt-Winters also adds a season: how much each weekday is above or below the level. The forecast goes up on Saturdays and down on Wednesdays.

The whole family is called ETS: Error, Trend and Season. Each part can be absent (N, for none), added (A) or multiplied (M). “Multiplied” means that the part grows with the level: a season of +20% instead of +5 units. An ETS model is named with three letters, such as ETS(A,N,A).

You do not need to choose the letters yourself. AutoETS from statsforecast fits the possible models and keeps the one with the best AICc. AICc is a fit score: it rewards a model that fits the history closely, and it punishes each extra part. So a trend is kept only if it helps more than it costs.

In the notebook, Priya fits AutoETS to the parasol’s sales up to Sunday 22 June 2025, the week before the heat wave:

from statsforecast import StatsForecast
from statsforecast.models import AutoETS
parasol = load_sales(["P025"])
origin = pd.Timestamp("2025-06-22")
train = parasol[parasol["ds"] <= origin][["unique_id", "ds", "y"]]
sf = StatsForecast(models=[AutoETS(season_length=7)], freq="D")
fc = sf.forecast(df=train, h=14)
sf.fit(train)
print("model chosen:", sf.fitted_[0, 0].model_["method"])
real = parasol[(parasol["ds"] > origin) & (parasol["ds"] <= origin + pd.Timedelta(days=14))]
print("14-day total: ETS", round(fc["AutoETS"].sum(), 1), "| real", int(real["y"].sum()))

season_length=7 tells it that the season repeats every 7 days. The output:

model chosen: ETS(A,N,A)
14-day total: ETS 176.3 | real 197
Three boxes for the letters of ETS(A,N,A). Error: A, added. Trend: N, none. Season: A, added, shown in orange. A note says that each letter is N, A or M, and that AutoETS keeps the best model.
What the three letters of the parasol's model mean. AutoETS found no trend in the parasol's history, and a weekly season that adds or takes away units.
Text description of the diagram

The diagram has three columns. Error, A: added, so misses go up and down by about the same number of units. Trend, N: none. Season, in orange, A: added, so each weekday adds or takes away some units. A note says that each letter is N (none), A (added) or M (multiplied), and that AutoETS keeps the best model.

For the parasol, AutoETS also chose its smoothing values: α = 0.41 for the level, and 0.08 for the season. So the level moves 41% of the way to each new day, and the weekly pattern changes only slowly.

  • Sold
  • ETS
  • seasonal naive
Six weeks of parasol sales, then the ETS forecast and the seasonal naive forecast for the 14 days after 22 June 2025. ETS repeats a smooth weekly pattern; the seasonal naive forecast copies last week, with its noise. Neither can know the heat wave from 30 June.
Text description of the chart

The chart shows the parasol’s daily sales from 12 May to 22 June 2025 as a grey line, mostly between 1 and 25 a day. After the origin, the real sales are grey dots: 50 in the first week, 147 in the heat-wave week. The ETS forecast is a smooth blue line that repeats the same week twice. It is between 10.7 and 15.5 units a day, 176.3 units in total. The seasonal naive forecast is a dashed orange line that repeats the last week twice. It is between 3 and 20 a day, 134 units in total. Both stay far below the heat-wave week.

ETS forecasts 176.3 parasols, closer to the real 197 than every naive forecast. But it is still too low, for the same reason: the heat wave was not in the parasol’s past.

The notebook runs AutoETS at the 13 origins of the course backtest, for the 50 products: 650 fits. The forecaster is a small function that returns the column yhat, as backtest.py expects:

def ets(train, h, future):
sf = StatsForecast(models=[AutoETS(season_length=7)], freq="D")
out = sf.forecast(df=train[["unique_id", "ds", "y"]], h=h)
return out.rename(columns={"AutoETS": "yhat"}).assign(yhat=lambda d: d["yhat"].clip(lower=0))
sales = load_sales(BACKTEST_PRODUCTS)
fc_ets = bt.backtest(sales, ets, h=14, origins=ORIGINS)
fc_snaive = bt.backtest(sales, bt.seasonal_naive(7), h=14, origins=ORIGINS)
fc_ma = bt.backtest(sales, bt.moving_average(28), h=14, origins=ORIGINS)
rows = [bt.score(f, sales, name, season=7, cost=bt.stock_cost(PRICES))
for name, f in [("seasonal naive", fc_snaive), ("moving average (28 days)", fc_ma), ("ETS (AutoETS)", fc_ets)]]
table = bt.show(rows)

clip(lower=0) turns a negative forecast into 0: a shop cannot sell −2 parasols. The output (the fit times change from run to run):

method horizon MASE WAPE % bias % pinball fit s cost $
ETS (AutoETS) 14 0.832 41.3 0.1 53.69 16578
moving average (28 days) 14 0.875 43.5 -0.1 0.04 19444
seasonal naive 14 1.034 51.3 -0.5 0.04 16739

ETS is the new best row of the table. Its MASE is 0.832 and its WAPE 41.3%, against 0.875 and 43.5% for the moving average. Its bias is +0.1%, so it is not too high or too low on average. Its order cost, 16,578 dollars, is a little below the seasonal naive’s 16,739 and far below the moving average’s 19,444. It is the first method of the course that is good on error and on cost.

The fit took about 54 seconds for 650 fits in this run, and 46 seconds in the course’s own measurement. That is much more than the naive methods, and still very little.

The explorer shows the parasol at the 13 origins of the backtest. Choose ETS and the moving average, and move the origin.

Explore

ETS and the baselines on the parasol

Sold in the 14 days
101
Forecast for the 14 days
92
WAPE at this origin
45.5%
WAPE, all 13 origins
65.1%

At the origin 8 Jun 2025, ETS forecasts 92 units for the next 14 days. 101 were sold. Its WAPE here is 45.5%: it is too low. Over all 13 origins, ETS has a WAPE of 65.1%: number 2 of 5 methods. It beats the baseline, seasonal naive (66.1%).

13 Apr 202511 May 20258 Jun 20256 Jul 20253 Aug 202531 Aug 202528 Sep 202526 Oct 202523 Nov 202521 Dec 202518 Jan 202615 Feb 202615 Mar 2026
13 Apr 202511 May 20258 Jun 20256 Jul 20253 Aug 202531 Aug 202528 Sep 202526 Oct 202523 Nov 202521 Dec 202518 Jan 202615 Feb 202615 Mar 2026
  • Units sold
  • Forecast of the chosen method
  • Baseline: seasonal naive
0102030Units sold per day12 May19 May26 May2 Jun8 Jun15 Jun22 Junorigintest: 14 daysSold before the originSold 9 Jun 2025: 1Sold 10 Jun 2025: 5Sold 11 Jun 2025: 5Sold 12 Jun 2025: 4Sold 13 Jun 2025: 3Sold 14 Jun 2025: 9Sold 15 Jun 2025: 7Sold 16 Jun 2025: 4Sold 17 Jun 2025: 10Sold 18 Jun 2025: 5Sold 19 Jun 2025: 3Sold 20 Jun 2025: 12Sold 21 Jun 2025: 20Sold 22 Jun 2025: 13seasonal naive: 118ETS: 92
0102030Units a day18 May8 Jun22 Junorigintest: 14 daysSold before the originSold 9 Jun 2025: 1Sold 10 Jun 2025: 5Sold 11 Jun 2025: 5Sold 12 Jun 2025: 4Sold 13 Jun 2025: 3Sold 14 Jun 2025: 9Sold 15 Jun 2025: 7Sold 16 Jun 2025: 4Sold 17 Jun 2025: 10Sold 18 Jun 2025: 5Sold 19 Jun 2025: 3Sold 20 Jun 2025: 12Sold 21 Jun 2025: 20Sold 22 Jun 2025: 13seasonal naive: 118ETS: 92
Show every method, ranked
Every method, scored on all 13 origins (lower WAPE is better)
RankMethodFamilyWAPE at this originWAPE, all 13 origins
1moving average (28 days)baseline101.3%64.9%
2ETSstatistical45.5%65.1%
3seasonal naivebaseline74.3%66.1%
4naivebaseline50.5%90.9%
5driftbaseline50.3%92.5%
The parasol P025 at the 13 origins of the course backtest, 14 days each: the real sales, the chosen method's forecast and the seasonal naive baseline (dashed). Every forecast is measured; the WAPE adds up all 182 forecast days.
Text description of the explorer

The explorer shows the parasol’s daily sales. The four weeks before the chosen origin are a grey line, and the 14 days after it are grey dots. The chosen method’s forecast is a blue line, and the seasonal naive forecast a dashed orange line. The controls choose the origin and one of five methods: naive, seasonal naive, the 28-day moving average, drift and ETS. The readout shows the units sold, the forecast and the WAPE. At the origin of 8 June 2025, 101 parasols were sold and ETS forecast 92. The WAPE over all 13 origins is 64.9% for the moving average (28 days) and 65.1% for ETS. It is 66.1% for seasonal naive, 90.9% for naive and 92.5% for drift.

On the parasol, ETS (65.1%) and the moving average (64.9%) are almost the same. The parasol’s sales change with the weather, and a model of its own past cannot know the weather. Over 50 products, ETS wins clearly. On one hard product, it does not.

The table adds up all 50 products. Per product, the result is:

  • Against the seasonal naive forecast, ETS wins on 50 of the 50 products. ETS keeps the weekly season, but it averages many weeks instead of copying one.
  • Against the 28-day moving average, ETS wins on 33 of 50. The moving average wins on 17.

We looked for a simple rule for those 17 products, and did not find one. They are not slower sellers: the median product sells 5.8 units a day, against 5.9 where ETS wins. Their weekly pattern is only a little weaker: on a median product, a weekday differs from the weekly mean by 7.2%, against 8.3%. A likely reason: when the weekly pattern is weak and the noise is large, the season that ETS estimates is partly noise. Then a flat mean does as well. The backtest, not a rule, tells you which products these are.

Use ETS as the first model after the baselines. It is fast, it needs only the series itself, and its parts (level, trend, season) can be explained to Kenji in a sentence each. It is a strong default for daily or weekly business series with one season.

Know its limits:

  • One season only. ETS takes one season_length. Hourly tickets have a daily and a weekly season. The failure case below shows what happens. Several seasons and long series fixes it.
  • Only the series’ own past. ETS cannot use a planned promotion, a holiday or the weather. Module 5 adds them as inputs.
  • Short histories. A weekly season needs several weeks of history before the season part means anything. A new product with two weeks of sales gets a poor season. Module 6 handles new products.
  • Its range assumes normal errors. ETS also gives a range around its forecast. An 80% range should hold the real value 8 times out of 10. On the 50 products, it held the real value 83.5% of the time: close, but Module 7 checks this properly.

What it costs: about 46 seconds for 650 fits on a laptop. The forecasting job in Module 8 fixes the model to ETS(A,N,A), so AutoETS does not search. Then it fits 600 series in about 5 seconds instead of 40.

Common mistake

One season only: ETS on hourly tickets

Symptom: Grace, the support lead, asks for next week’s tickets per hour. Priya fits AutoETS with a daily season (24 hours) to the last 12 weeks of hourly tickets, and forecasts the last week of the history. Added up per day, the forecast and the real tickets are:

Day Mon Tue Wed Thu Fri Sat Sun
Real tickets 473 474 393 409 294 337 243
ETS forecast 243 243 243 243 243 243 243

The forecast is the same 243 tickets every day. The real tickets fall from 473 on Monday to 243 on Sunday.

Cause: a daily season repeats the same 24 hours seven times. The model cannot know that Monday is busy and Sunday is quiet. Here it repeats the level of the last day of the history, a quiet Sunday, for the whole week. With season_length=168 (one week of hours) instead, ETS follows the week (381 tickets on Monday, 226 on Sunday). But then it learns each hour of the week from only 12 examples in 12 weeks.

Fix: use a method that takes several seasons, such as MSTL, in Several seasons and long series.

Consequence in AI work: a model’s options limit what it can learn, whatever the data says. Before you compare models, check that each one can represent the patterns you already see in the series. Otherwise you compare a model with a handicap.

  1. Open the notebook from the box Do this lesson in a notebook at the top of this page.
  2. Run the setup cells and the helper cell. The data cell ends with the Ready: line.
  3. In section 1, predict how much weight the last 7 days get with α = 0.3. Then run the cell. You see 0.918.
  4. In section 2, run the parasol cell. You see model chosen: ETS(A,N,A) and real 197.
  5. In section 3, predict if ETS beats the moving average. Then run the backtest cell. It takes about a minute.
  6. In section 4, write the product-by-product count in the Your turn cell, and run it.
  7. Run the check cell. It prints Check passed.
  8. In section 5, run the hourly tickets cell. You see the same 243 tickets on every day.

Your turn: ETS against the moving average, product by product

Section titled “Your turn: ETS against the moving average, product by product”

fc_ets and fc_ma hold every forecast of ETS and of the 28-day moving average, with the real value y. Before you compute anything, predict: on how many of the 50 products does ETS have the smaller error? For each product, add up the absolute errors of each method over all origins and days. Count the products where the ETS sum is smaller, and put the count in ets_wins.

Show the answer
e_ets = (fc_ets["y"] - fc_ets["yhat"]).abs().groupby(fc_ets["unique_id"]).sum()
e_ma = (fc_ma["y"] - fc_ma["yhat"]).abs().groupby(fc_ma["unique_id"]).sum()
ets_wins = int((e_ets < e_ma).sum())
print(ets_wins, "of", len(e_ets))
33 of 50

ETS wins on 33 products, and the moving average on 17. The sum of the absolute errors is each product’s WAPE without the division. Both methods divide by the same sales, so the division does not change the winner.

Check your understanding

With α = 0.3, the level is 10.60 and the product sells 9 units today. What is the new level, and what is the forecast for the next three days?

Show the answer

The miss is 9 − 10.60 = −1.60. The new level is 10.60 + 0.3 × (−1.60) = 10.60 − 0.48 = 10.12. Simple exponential smoothing forecasts the last level for every day: 10.12, 10.12 and 10.12. A trend (Holt) or a season (Holt-Winters) would make the days differ.

Check your understanding

AutoETS chose ETS(A,N,A) for the parasol. What does each letter mean, and why can the model still not forecast the heat wave?

Show the answer
  • A (error): the misses are added; they go up and down by about the same number of units.
  • N (trend): no trend; the level does not grow or fall steadily.
  • A (season): an added weekly season; each weekday adds or takes away some units.

ETS learns only from the parasol’s own past sales. The heat wave came from the weather, which is not in the model. So it forecast 176.3 parasols for the two weeks, and 197 were sold.

Check your understanding

Use the backtest table and the product-by-product counts. On which of Larkfield's 50 products did ETS beat the seasonal naive forecast, on which did it not, and why?

Show the answer

ETS beat the seasonal naive forecast on all 50 products: MASE 0.832 against 1.034, WAPE 41.3% against 51.3%. The seasonal naive forecast copies one single week, with its noise and accidents. ETS also follows the weekly season, but it averages it over many weeks with weights that fade. Against the 28-day moving average, ETS won on 33 products and lost on 17. We found no simple rule for these 17 products: they are not slower sellers. A likely reason is a weak weekly pattern and much noise. Then the season that ETS estimates is partly noise, and a flat 28-day mean does as well.

Recap

  • Exponential smoothing keeps a level and moves it a part α of the way to each new value. The weights of past days fade: with α = 0.3, the last 7 days get 0.918 of the weight.
  • Holt adds a trend, Holt-Winters adds a season. The family is ETS (error, trend, season), each part none (N), added (A) or multiplied (M).
  • AutoETS chooses the model by its AICc fit score. For the parasol, it chose ETS(A,N,A) with α = 0.41.
  • In the course backtest, ETS has MASE 0.832, WAPE 41.3% and a cost of 16,578 dollars. It is the best row so far, and it took about 46 seconds for 650 fits.
  • ETS beats the seasonal naive forecast on 50 of 50 products, and the 28-day moving average on 33 of 50.
  • ETS takes one season only. On hourly tickets, it forecast the same 243 tickets every day.

You can now fit ETS and read its letters. ETS describes a series by its level, trend and season. Next, you meet a different way to describe a series: from its past values and its past errors. Its range also shows clearly how uncertainty grows with the horizon.

Next: ARIMA and SARIMA