TapCub is live — analytics, insights and live chat, free on one platform
Glossary

Exponential smoothing: definition, formula and a worked example

Exponential smoothing weights recent data more heavily to smooth noisy series and make short-term forecasts. The formula, a six-day example recomputed at two alphas, and how to pick α.

Lin Zhou Reviewed by Su He Also called SES Last updated Nov 16, 2026 Glossary

Quick answer

Exponential smoothing is a time-series method that averages past observations with exponentially decaying weights, so recent data counts most. One constant, α (0–1), sets the trade-off: high α follows new data fast, low α stays steady. The smoothed value doubles as a one-step forecast.

Formula: S_t = α·x_t + (1−α)·S_{t−1}, 0 < α < 1

What is exponential smoothing?

In forecasting, exponential smoothing is a time-series technique that computes a running average in which every past observation keeps a weight, but the weights shrink exponentially as data gets older. Recent values dominate; old values fade instead of being dropped. One number, the smoothing constant α (alpha, between 0 and 1), sets the trade-off: high α follows the latest data closely, low α holds a steadier line. That is the whole idea — α decides whether the curve "follows fast" or "stays steady".

Two jobs come out of the same recursion. As a smoother, it turns a jagged daily series — say, website sessions that bounce between 95 and 180 — into a readable trend line. As a forecaster, the current smoothed value doubles as the one-step-ahead prediction for tomorrow. The NIST/SEMATECH e-Handbook presents it exactly this way: past observations get exponentially decreasing weights to forecast future values (NIST §6.4.3.1). It is the workhorse baseline of time series analysis because it needs one parameter, one stored value, and one line of arithmetic per period.

How the formula works

The simple exponential smoothing (SES) recursion is:

S_t = α · x_t + (1 − α) · S_{t−1}        0 < α < 1
  • x_t — the observation at time t (e.g. today's sessions).
  • S_{t−1} — yesterday's smoothed value, which already summarizes everything before it.
  • S_t — the new smoothed value: an interpolation between what you just saw and what you believed before.
  • α — how much of today's observation you let in. α = 0.3 means "30% new data, 70% history".

Two equivalent readings make the behaviour obvious. Rearranged as S_t = S_{t−1} + α·(x_t − S_{t−1}), it is error correction: nudge the old estimate toward the new observation by a fraction α of the gap. Expanded by repeated substitution, it is a weighted sum of all past data with weights α, α(1−α), α(1−α)², … — a geometric decay, which is where "exponential" comes from. With α = 0.3 the five most recent observations carry weights 0.3, 0.21, 0.147, 0.1029, 0.07203 (NIST); with α = 0.8 they carry 0.8, 0.16, 0.032, 0.0064, 0.00128 — history is practically forgotten after three periods.

One notation trap: the NIST e-Handbook writes S_t = α·y_{t−1} + (1−α)·S_{t−1} (Hunter's convention, where S_t is the forecast for period t), while Hyndman's Forecasting: Principles and Practice and the EWMA control-chart literature use the current-observation form above (Roberts, 1959). They are the same model shifted by one index — check which convention a tool uses before comparing outputs.

Worked example: six days, two alphas

Sample data: daily sessions for a small site over six days, with one spike on day 4 (a link shared in a newsletter). We initialize with the common convention S₁ = x₁ = 100.

Day123456
Sessions x_t10011095180105100

α = 0.3, every step shown (each row is 0.3·x_t + 0.7·S_{t−1}):

DayArithmeticS_t
1S₁ = x₁100
20.3×110 + 0.7×100 = 33 + 70103
30.3×95 + 0.7×103 = 28.5 + 72.1100.6
40.3×180 + 0.7×100.6 = 54 + 70.42124.42
50.3×105 + 0.7×124.42 = 31.5 + 87.094118.594
60.3×100 + 0.7×118.594 = 30 + 83.0158113.016

α = 0.8 side by side (same data, same initialization):

Dayx_tS_t (α=0.3)S_t (α=0.8)
1100100100
2110103108
395100.697.6
4180124.42163.52
5105118.59116.70
6100113.02103.34

Read day 4 onward. The spike pulls α = 0.8 up to 163.52 — the "trend" now claims the site runs at ~164 sessions, which is noise-chasing: by day 6 it has crashed back to 103.34, overshooting in both directions. α = 0.3 absorbs the spike to 124.42 and settles more slowly (113.02 on day 6) — steadier, but it lags: part of that 113 is still spike residue, not the true ~100–110 level.

Score it. Treat each S_{t−1} as the one-step forecast of x_t and take the mean absolute error (MAE) over days 2–6:

  • α = 0.3 errors: |110−100| = 10, |95−103| = 8, |180−100.6| = 79.4, |105−124.42| = 19.42, |100−118.594| = 18.594 → MAE = 135.414 ÷ 5 = 27.08
  • α = 0.8 errors: 10, 13, 82.4, 58.52, 16.704 → MAE = 180.624 ÷ 5 = 36.12

On this series the steady setting wins by 9 sessions of average error, and the gap comes almost entirely from the day after the spike (19.42 vs 58.52). That is the thesis in numbers: α decides whether you follow fast or stay steady, and on noisy data with one-off spikes, following fast costs you. If your series genuinely steps to a new level and stays there, the ranking flips — which is why you score α on your own data instead of copying a default.

Choosing alpha and the starting value

Rules that follow from the arithmetic above:

  • Score, don't guess. Compute one-step errors for a grid of α (e.g. 0.1 to 0.9) and pick the smallest MSE or MAE — the procedure NIST demonstrates on its 12-observation example (NIST §6.4.3.1). Software optimizes α the same way (FPP3 §8.1).
  • Translate α into lag. The average age of the data inside an SES forecast is 1/α periods (Nau, Duke): α = 0.3 → ~3.3 days, α = 0.1 → 10 days. If you need the curve to react within about three days, α below 0.3 cannot deliver it.
  • Initialize deliberately. S₁ = x₁ is common but injects day-1 noise into every later value. Alternatives: the mean of the first 4–5 observations, or a known target level. The smaller α is, the longer a bad initialization persists (NIST) — with only six days of data, treat the first two or three smoothed values as warm-up.
  • Know the corners. α = 1 reduces to "tomorrow = today" (random walk); α = 0 freezes at the initial value. Neither smooths anything.

Simple vs Holt vs Holt-Winters

SES only tracks a level, so its forecasts are flat: every future period gets the last smoothed value (FPP3 §8.1). Two extensions add components, each with its own smoothing constant:

MethodComponentsExtra parametersUse whenTypical data
Simple (SES)LevelαNo trend, no seasonalityDaily sessions of a stable site
Holt's linear (double)Level + trendα, βSustained growth or decline, no seasonalitySignups compounding month over month
Holt-Winters (triple)Level + trend + seasonα, β, γ, period mRepeating weekly/yearly cycleB2B traffic with weekend dips (seasonality)

Holt-Winters comes in additive (seasonal swing is a fixed amount) and multiplicative (swing scales with the level) versions (FPP3 §8.3). Decision rule: if SES residuals still show a slope, move to Holt; if they show a repeating cycle, move to Holt-Winters. Applying SES to strongly seasonal data — say, weekday-heavy B2B sessions — makes every Monday look like an anomaly; that is a model mismatch, not a traffic problem.

Exponential smoothing vs moving average

A moving average weights the last m observations equally and drops everything older; exponential smoothing discounts gradually and never fully forgets. Duke's notes give the clean comparison: a moving average of m terms has average data age (m+1)/2, SES has 1/α — so SES with α = 2/(m+1) matches an m-term average's lag while putting more weight on the newest values (Nau).

Same sample data, forecast MAE over the shared window days 4–6: a 3-day moving average forecasts 101.67, 128.33, 126.67 (errors 78.33, 23.33, 26.67; MAE 42.78), while SES α = 0.3 scores 39.14 on those same three days. The moving average also carries the spike for exactly three days and then drops it off a cliff; SES fades it geometrically. Practical rule: for a chart overlay or a one-parameter alert baseline, SES stores one number per series and updates in one line — an m-term average needs the last m raw values.

Using a smoothed baseline for anomaly detection

The error-correction form doubles as an anomaly detection residual: e_t = x_t − S_{t−1} is "today versus what the recent past led you to expect". In the sample data, day 4's residual is 180 − 100.6 = 79.4 sessions, roughly 79% above the smoothed baseline — flag-worthy; day 5's residual against the updated baseline is only −19.4.

The α trade-off returns here with teeth. High α lets a spike inflate the baseline, so a genuinely elevated second day looks normal (α = 0.8 expects 163.52 on day 5 and sees 105 as a drop). Low α keeps the baseline honest but reacts slowly to real level shifts, generating repeated false alarms after a legitimate step change. Two guardrails: score residuals on a holdout window before trusting thresholds, and remember SES has no seasonality — for weekly cycles, either compare like weekdays or use Holt-Winters residuals instead.

Common mistakes

  • Copying α = 0.3 from a tutorial. It is a teaching default, not an optimum. If your one-step MAE at α = 0.6 is lower on your data, use 0.6.
  • Reading the smoothed curve as a multi-step forecast. SES forecasts are flat. If you need next quarter's trajectory, that is a trend/seasonality model or a proper marketing forecasting workflow, not SES stretched further.
  • Letting the spike teach the baseline. If you smooth and alert on the same series, a one-off spike raises the baseline and masks the next day's real problem. Freeze or winsorize known one-off events before updating.
  • Ignoring warm-up. With S₁ = x₁, early smoothed values are mostly initialization noise. Skip the first few periods before computing errors or setting thresholds.
  • Mixing conventions. Hunter-style (y_{t−1}) and Roberts-style (y_t) formulas differ by one period; comparing outputs across tools without aligning them makes one look "a day late".

Smoothing your TapCub data

Take the six-day example above and run it on your real sessions. The boundary first: TapCub has no exponential-smoothing function or forecast, and none of its monitor conditions uses an exponentially weighted baseline. The smoothing is arithmetic you do on TapCub data, in four steps:

  1. Pull a clean daily series. In Web analytics › Reports, set the bucket to day, keep the sessions column and export the table (CSV or XLSX). Bots and AI crawlers are counted separately by default (Web analytics › Bots), so the series you smooth is human sessions and a crawler burst does not pose as a traffic spike.
  2. Smooth and score α in a spreadsheet. With observations in column B and smoothed values in column C, =0.3*B3+0.7*C2 is the whole model, exactly as in the worked example. Add a column of one-step errors and compare MAE across α = 0.1 to 0.9 before you settle on one.
  3. Alert with the baselines monitors actually have. In All-platform analytics › Operations › Monitors, the built-in sessions measure can use a forecast band (the mean of the last 7 days and the same weekday, skipped until 14 days of history) or an anomaly condition (mean and spread of the same weekday over 28 days). Both are equal-weight averages, closer to a moving average than to SES, and the same-weekday comparison already handles the weekly cycle that trips SES. A dry run reports how many times a rule would have fired over the past 30 days; compare that with the days your own SES residuals flag, such as the 79.4-session spike. Monitors only notify; they take no action.
  4. Explain the flagged day before it teaches the baseline. Mark the newsletter day in Configuration › Annotations so the spike is not mistaken for a new level, and open Why did it change in All-platform analytics › Operations › AI Assistant to see how much each channel, referrer domain, country, device or entry page contributed to the change.

So the division of labour is: TapCub supplies clean daily data, exports, equal-weight alert baselines and change attribution; the smoothing recursion, and any forecast built on it, stays in your spreadsheet. The α you pick there still decides whether your baseline follows fast or stays steady.

Sources

Was this explanation clear?

We use your answers to fix and expand this page.

Lin Zhou

Product Experience Specialist

Writes about product experience, practical workflows, and growth, turning complex processes into clear steps for everyday decisions.

FAQ

Questions people ask

Short answers to the follow-up questions that come up most.

Still have a question?

Chat with the team behind TapCub — we usually reply within a few hours on business days.

What is a good alpha value for exponential smoothing?

There is no universal value. Score one-step forecast errors for a grid of α on your own data and pick the lowest MSE or MAE — the NIST handbook procedure. As a starting intuition, α = 0.1–0.3 suits noisy series where you want stability; α = 0.5+ suits series that genuinely shift level and must be tracked quickly.

What is the difference between exponential smoothing and a moving average?

A moving average weights the last m observations equally and drops older ones; exponential smoothing discounts all history geometrically and never fully forgets. Matching average lag (SES 1/α vs SMA (m+1)/2), SES puts more weight on the newest data and needs only one stored value per series.

When should I use Holt or Holt-Winters instead of simple exponential smoothing?

Use Holt's linear method when the series has a sustained trend but no seasonality, and Holt-Winters when a weekly or yearly cycle repeats. Simple exponential smoothing only tracks a level, so its forecasts are flat and it misreads strong seasonal swings as anomalies.

Can exponential smoothing detect anomalies in website traffic?

Yes, as a residual: compare each observation with the previous smoothed value and flag large gaps. In our sample data the spike day sits 79.4 sessions (~79%) above its baseline. Watch the α trade-off — high α lets spikes inflate the baseline; and SES has no seasonality, so compare like weekdays.

See this number on your own site

Add one script tag for web and apps, and the first reports fill in within minutes.

See your data clearly. Find your growth.

Every click, backed by data. Install one line of code and see your first numbers in a minute.

No credit card · Free plans for analytics and chat · Cookieless analytics