ETS Smoothing Parameters: Theory, Intuition, and Excel Verification

Part 1 — What Are Smoothing Parameters in an ETS Model?

From First-Order Exponential Smoothing to ETS

The McGuigan textbook (Chapter 5, Equation 5.15) introduces first-order exponential smoothing as:

\[\hat{Y}_{t+1} = wY_t + (1 - w)\hat{Y}_t\]

where \(w\) is a weight between 0 and 1 placed on the most recent observation, and \((1 - w)\) is the weight placed on the previous forecast. A large \(w\) means the model reacts quickly to recent data; a small \(w\) produces smoother, slower-moving forecasts.

The ETS (Error, Trend, Seasonality) framework generalizes this idea into a complete family of exponential smoothing models. Every ETS model has up to four smoothing parameters, each controlling how quickly the model adapts to a different component of the series.

---

The Four Smoothing Parameters

Alpha (\(\alpha\)) — Level Smoothing

Alpha is the direct extension of McGuigan’s \(w\). It controls how quickly the estimated level of the series updates in response to new observations.

\[\ell_t = \alpha y_t + (1 - \alpha)\ell_{t-1}\]

• \(\alpha\) close to 1: the level jumps quickly to match the latest observation. The model has a short memory and reacts fast to shocks.
• \(\alpha\) close to 0: the level changes slowly, giving heavy weight to the historical average. The model has a long memory and is resistant to outliers.
• All ETS models have alpha. It is the only parameter in ETS(A,N,N) — simple exponential smoothing with no trend and no seasonality.

Beta (\(\beta\)) — Trend Smoothing

Beta controls how quickly the estimated trend (slope) updates over time. It only appears in models with a trend component (ETS models where the middle letter is A for additive trend or M for multiplicative trend).

\[b_t = \beta(\ell_t - \ell_{t-1}) + (1 - \beta)b_{t-1}\]

• \(\beta\) close to 1: the slope is recalculated aggressively from the most recent change in level. The trend can reverse direction rapidly.
• \(\beta\) close to 0: the slope is very stable and changes slowly. This is appropriate when the underlying trend is gradual and consistent.
• For most real economic series, the estimated \(\beta\) is small because trend direction rarely changes abruptly.

Gamma (\(\gamma\)) — Seasonal Smoothing

Gamma controls how quickly the estimated seasonal indices update. It only appears in models with a seasonal component (ETS models where the last letter is A or M).

\[s_t = \gamma(y_t - \ell_{t-1} - b_{t-1}) + (1 - \gamma)s_{t-m}\]

where \(m\) is the seasonal period (12 for monthly data, 4 for quarterly).

• \(\gamma\) close to 1: seasonal patterns are updated aggressively each period. Appropriate when seasonal patterns are evolving over time.
• \(\gamma\) close to 0: seasonal patterns are treated as fixed and stable. Appropriate when the same seasonal cycle repeats reliably year after year.

Phi (\(\phi\)) — Damping Parameter

Phi is unique — it does not smooth a component but instead controls whether a trend is allowed to persist indefinitely or is gradually dampened toward zero.

• \(\phi = 1\): the trend continues at its current slope forever. This produces unrealistically explosive long-range forecasts for most series.
• \(0 < \phi < 1\): the trend is damped. As the forecast horizon grows, the trend contribution shrinks. At long horizons the forecast flattens.
• Phi only appears in damped trend models: ETS(A,Ad,N) and ETS(A,Ad,A).
• Research shows that damped trend models outperform undamped ones on average across many series because most real trends do not continue indefinitely.

---

Summary Table

Parameter	Component	Range	Low value means	High value means
\(\alpha\)	Level	(0,1)	Long memory, slow adaptation	Short memory, fast adaptation
\(\beta\)	Trend	(0,1)	Stable slope	Volatile slope
\(\gamma\)	Seasonality	(0,1)	Fixed seasonal pattern	Evolving seasonal pattern
\(\phi\)	Damping	(0,1)	Trend dies quickly	Trend persists longer

---

Connecting to McGuigan’s Framework

McGuigan’s \(w\) in Equation 5.15 is exactly \(\alpha\) in ETS(A,N,N). His observation that “a large \(w\) indicates that a heavy weight is being placed on the most recent observation” is the intuition behind alpha across all ETS variants. The ETS framework simply extends this one parameter into a full system where the level, trend, and seasonal components each have their own adaptive smoothing rate.

The key insight from Equation 5.18 in the textbook is that exponential smoothing is a weighted average of all past observations with geometrically declining weights — the most recent observation gets weight \(w\), the one before gets \(w(1-w)\), the one before that gets \(w(1-w)^2\), and so on. This geometric decay is what distinguishes exponential smoothing from simple moving averages which give equal weight to the last \(N\) observations and zero weight to everything before.

---

Part 2 — Two Time Series: Visualization and Parameter Intuition

library(fpp3)
library(tidyverse)
library(fredr)
library(kableExtra)
library(patchwork)

fredr_set_key("cdbfb0bbfdb75085e9eea36ef1806968")

Series 1: US Retail Sales (RSAFS) — Trended, Seasonal

retail_raw <- fredr(
  series_id         = "RSAFS",
  observation_start = as.Date("2010-01-01"),
  observation_end   = as.Date("2023-12-01")
)

retail_ts <- retail_raw |>
  select(date, value) |>
  mutate(Month = yearmonth(date)) |>
  as_tsibble(index = Month) |>
  rename(Sales = value) |>
  select(Month, Sales)

p1 <- retail_ts |>
  autoplot(Sales) +
  labs(
    title    = "US Advance Retail Sales (RSAFS)",
    subtitle = "Monthly | Jan 2010 to Dec 2023",
    y        = "Millions of Dollars",
    x        = NULL
  ) +
  theme_minimal()

p2 <- retail_ts |>
  gg_season(Sales, labels = "both") +
  labs(title = "Seasonal Plot", y = NULL, x = NULL) +
  theme_minimal()

p1 / p2

Series 1: US Advance Retail Sales — Strong trend with seasonal pattern

Eyeballed parameter expectations for Retail Sales:

Looking at the plot, retail sales shows a strong upward trend from 2010 to 2023 with a sharp COVID dip in early 2020 followed by an unusually rapid recovery. The seasonal pattern is consistent every year with a December peak and a February trough. Based on this visual inspection:

• Alpha: I expect a moderate alpha around 0.3 to 0.5. The series has a strong trend and consistent seasonal pattern, so the level does not need to update too aggressively. However the COVID shock in 2020 was dramatic enough that the model needs some responsiveness.
• Beta: I expect a small beta around 0.05 to 0.15. The underlying trend is steady and gradual. Trend direction rarely reverses, so slope should update slowly.
• Gamma: I expect a small gamma around 0.05 to 0.15. The seasonal pattern is extremely consistent across all 14 years. December is always the peak and February is always the trough by roughly the same proportion. Seasonal indices should be treated as stable.
• Phi: If a damped model is selected I would expect phi around 0.90 to 0.98. The trend is real and persistent, so damping should be mild rather than aggressive.

---

Series 2: US Unemployment Rate (UNRATE) — No Trend, Irregular Spikes

unemp_raw <- fredr(
  series_id         = "UNRATE",
  observation_start = as.Date("2010-01-01"),
  observation_end   = as.Date("2023-12-01")
)

unemp_ts <- unemp_raw |>
  select(date, value) |>
  mutate(Month = yearmonth(date)) |>
  as_tsibble(index = Month) |>
  rename(Unemployment = value) |>
  select(Month, Unemployment)

p3 <- unemp_ts |>
  autoplot(Unemployment) +
  labs(
    title    = "US Unemployment Rate (UNRATE)",
    subtitle = "Monthly | Jan 2010 to Dec 2023",
    y        = "Percent",
    x        = NULL
  ) +
  theme_minimal()

p4 <- unemp_ts |>
  gg_subseries(Unemployment) +
  labs(title = "Subseries Plot by Month", y = NULL, x = NULL) +
  theme_minimal()

p3 / p4

Series 2: US Unemployment Rate — Declining trend, dramatic COVID spike

Eyeballed parameter expectations for Unemployment:

The unemployment rate starts around 9.8% in early 2010, declines steadily to about 3.5% by 2019, spikes dramatically to 14.7% in April 2020 due to COVID, then recovers rapidly back to around 3.5% by late 2022. The subseries plot shows very little consistent seasonal pattern — unemployment is not strongly seasonal in the US.

• Alpha: I expect a high alpha around 0.6 to 0.9. Unemployment reacts quickly to macroeconomic shocks and the model needs to track rapid level changes. The COVID spike and recovery happened within months, requiring fast adaptation.
• Beta: I expect a very small beta around 0.01 to 0.10. The underlying trend is gradual even though the level moves significantly. The slope from 2010 to 2019 was steady and slow.
• Gamma: I expect gamma close to 0 or the model will not include seasonality at all since the subseries plot shows no consistent seasonal pattern.
• Phi: If damped, I expect phi around 0.80 to 0.95. The trend in unemployment is not permanent — it eventually stabilizes at a structural rate. Damping is appropriate.

---

Part 3 — Fit Best ETS Model and Reconcile Intuition

# Fit best ETS model to both series using AIC selection
fit_retail <- retail_ts |>
  model(ETS = ETS(Sales))

fit_unemp <- unemp_ts |>
  model(ETS = ETS(Unemployment))

# Extract model reports
report_retail <- fit_retail |> report()

Series: Sales 
Model: ETS(M,A,N) 
  Smoothing parameters:
    alpha = 0.7109986 
    beta  = 0.0001000185 

  Initial states:
     l[0]     b[0]
 339200.4 1887.068

  sigma^2:  6e-04

     AIC     AICc      BIC 
3996.143 3996.513 4011.763 

report_unemp  <- fit_unemp  |> report()

Series: Unemployment 
Model: ETS(M,A,N) 
  Smoothing parameters:
    alpha = 0.9999 
    beta  = 0.9999 

  Initial states:
     l[0]       b[0]
 9.517881 -0.6905423

  sigma^2:  0.0238

     AIC     AICc      BIC 
811.6534 812.0238 827.2733 

Retail Sales: Best ETS Model

report_retail

ETS
<ETS(M,A,N)>

# Extract smoothing parameters
retail_params <- fit_retail |>
  tidy() |>
  filter(term %in% c("alpha", "beta", "gamma", "phi")) |>
  select(term, estimate) |>
  rename(Parameter = term, Estimate = estimate)

retail_params |>
  kbl(
    digits  = 4,
    caption = "ETS Smoothing Parameters — US Retail Sales"
  ) |>
  kable_styling(
    bootstrap_options = c("striped", "hover"),
    full_width        = FALSE
  )

ETS Smoothing Parameters — US Retail Sales

Parameter	Estimate
alpha	0.7110
beta	0.0001

fit_retail |>
  forecast(h = 24) |>
  autoplot(retail_ts, level = c(80, 95)) +
  labs(
    title    = "ETS Forecast: US Retail Sales",
    subtitle = paste("Model:", fit_retail |> select(ETS) |> as.character()),
    y        = "Millions of Dollars",
    x        = NULL
  ) +
  theme_minimal()

Best ETS Model Forecast — US Retail Sales (24 months ahead)

Reconciling Retail Sales Results:

Insert your actual parameter values from the model output above and compare to your predictions. Example reconciliation:

The best-selected ETS model for retail sales is typically ETS(M,Ad,M) or ETS(A,Ad,A) reflecting the multiplicative seasonal pattern growing with the level and a damped trend. Comparing to my predictions:

• Alpha: If the estimated alpha is around 0.3-0.5 this confirms the intuition that the level updates at a moderate pace. The COVID shock temporarily forced faster adaptation but the overall stable trend keeps alpha moderate.
• Beta: A small estimated beta confirms that the underlying trend is gradual and consistent. The model does not need to relearn the slope frequently.
• Gamma: A small estimated gamma confirms that seasonal patterns are stable across the 14-year sample. The December-February pattern barely changes.
• Phi: If the model selected a damped variant and phi is close to 0.98 this suggests the trend is real but expected to eventually slow, which is economically reasonable for retail sales.

---

Unemployment Rate: Best ETS Model

report_unemp

ETS
<ETS(M,A,N)>

unemp_params <- fit_unemp |>
  tidy() |>
  filter(term %in% c("alpha", "beta", "gamma", "phi")) |>
  select(term, estimate) |>
  rename(Parameter = term, Estimate = estimate)

unemp_params |>
  kbl(
    digits  = 4,
    caption = "ETS Smoothing Parameters — US Unemployment Rate"
  ) |>
  kable_styling(
    bootstrap_options = c("striped", "hover"),
    full_width        = FALSE
  )

ETS Smoothing Parameters — US Unemployment Rate

Parameter	Estimate
alpha	0.9999
beta	0.9999

fit_unemp |>
  forecast(h = 24) |>
  autoplot(unemp_ts, level = c(80, 95)) +
  labs(
    title    = "ETS Forecast: US Unemployment Rate",
    subtitle = paste("Model:", fit_unemp |> select(ETS) |> as.character()),
    y        = "Percent",
    x        = NULL
  ) +
  theme_minimal()

Best ETS Model Forecast — US Unemployment Rate (24 months ahead)

Reconciling Unemployment Results:

The best ETS model for unemployment is likely to be ETS(A,Ad,N) or ETS(A,N,N) reflecting the absence of meaningful seasonality and the presence of a declining trend that needs damping.

• Alpha: If the estimated alpha is higher than for retail sales this confirms the intuition that unemployment reacts faster to economic shocks. The COVID spike required rapid level adjustment.
• Beta: A small or near-zero beta confirms the trend changes slowly. The long steady decline from 2010 to 2019 barely changes slope month to month.
• Gamma: Either absent from the model entirely or very small, confirming the visual inspection that US unemployment has no meaningful seasonal pattern at monthly frequency.
• Phi: If phi is selected and is below 0.95 this suggests the model expects the trend to stabilize — consistent with unemployment mean reverting toward a structural equilibrium rate.

---

Part 4 — ETS(ANN): Finding Alpha with Excel Solver

ETS(A,N,N) is simple exponential smoothing — exactly McGuigan’s first-order model with additive errors. The forecast equation is:

\[\hat{y}_{t+1|t} = \ell_t = \alpha y_t + (1-\alpha)\ell_{t-1}\]

The model has a single parameter: \(\alpha\). We find it by minimizing the sum of squared errors (SSE) across the training period.

R Implementation for Reference

# Fit ETS(ANN) explicitly — forces simple exponential smoothing
fit_ann <- retail_ts |>
  model(ETS_ANN = ETS(Sales ~ error("A") + trend("N") + season("N")))

report_ann <- fit_ann |> report()

Series: Sales 
Model: ETS(A,N,N) 
  Smoothing parameters:
    alpha = 0.9998986 

  Initial states:
     l[0]
 347406.6

  sigma^2:  132474569

     AIC     AICc      BIC 
4006.733 4006.880 4016.105 

report_ann

ETS_ANN
<ETS(A,N,N)>

alpha_r <- fit_ann |>
  tidy() |>
  filter(term == "alpha") |>
  pull(estimate)

cat("Alpha estimated by R (ETS ANN):", round(alpha_r, 6), "\n")

Alpha estimated by R (ETS ANN): 0.999899 

fit_ann |>
  augment() |>
  ggplot(aes(x = Month)) +
  geom_line(aes(y = Sales, colour = "Actual")) +
  geom_line(aes(y = .fitted, colour = "Fitted"), linetype = "dashed") +
  scale_colour_manual(values = c("Actual" = "#2c3e50", "Fitted" = "#e74c3c")) +
  labs(
    title    = "ETS(ANN) Fitted Values vs Actual",
    subtitle = paste0("Alpha = ", round(alpha_r, 4)),
    y        = "Sales (Millions of Dollars)",
    x        = NULL,
    colour   = NULL
  ) +
  theme_minimal() +
  theme(legend.position = "bottom")

ETS(ANN) Fitted Values vs Actual — Retail Sales

Excel Solver Setup

To replicate the alpha estimate in Excel, set up the spreadsheet as follows:

Column structure:

Column	Content	Formula
A	Month (1, 2, 3, …)
B	Actual Sales \(Y_t\)	Data values
C	Level \(\ell_t\)	C2 = B2 (initialize to first observation)
		C3 = =$B$1*B2 + (1-$B$1)*C2 (drag down)
D	Forecast \(\hat{Y}_t\)	D3 = C2 (one-step-ahead forecast = lagged level)
E	Error \(e_t = Y_t - \hat{Y}_t\)	E3 = =B3-D3
F	Squared Error \(e_t^2\)	F3 = =E3^2
B1	Alpha parameter	Start with any value between 0 and 1
SSE	Sum of squared errors	=SUM(F3:F168)

Solver settings:

• Set Objective: SSE cell → Minimize
• By Changing Variable: Cell B1 (alpha)
• Subject to Constraints: B1 >= 0.0001 and B1 <= 0.9999
• Method: GRG Nonlinear

The alpha value Solver produces should match the R estimate to at least 3 decimal places. Any small difference reflects rounding in initialization or convergence tolerance rather than a methodological error.

(Attach Excel screenshot here showing Solver dialog and matching alpha value)

# Compute SSE for reference to verify Excel calculation
sse_r <- fit_ann |>
  augment() |>
  summarise(SSE = sum(.resid^2, na.rm = TRUE)) |>
  pull(SSE)

cat("SSE from R ETS(ANN):", round(sse_r, 2), "\n")

SSE from R ETS(ANN): 69116767 236349 54671965 8385365 9064354 117172.8 1201140 2484126 8947037 16291743 10467872 6397500 5833463 9494070 10892061 4183393 99094.44 9498527 20934.51 174711.7 17489478 4689061 1476758 317107.7 12574521 22842277 1927889 2640168 486038.6 12076116 3184968 23407994 9969746 34849.41 4901712 1918847 11163219 21943029 8378276 3969233 5981928 1548153 5509001 332654.5 124567.7 1836122 1350563 5067531 10567517 26777215 20200750 18631789 500467.8 1106855 599241.1 10890518 1415304 2937383 1464520 9017272 3618762 1186341 42613343 678418.5 11526592 7455.31 15928151 8573.92 3928361 1378748 2591717 3408318 7197485 13327815 1589188 5749791 1528297 22535199 200273.1 24634.76 9672199 626179.8 2032.23 29300618 26087053 16251.74 372115.8 2566206 6359665 5644161 1661.31 866753.3 94169447 147443.4 23445341 6956358 4716423 6847537 6282.98 1936.71 60933706 2997084 8490373 91988.06 3755725 24096352 5946270 94882266 1758310 1719073 56868685 2679.55 29138461 260658.4 14745997 12043601 10749534 1865048 12724477 2726995 557759 44489.06 2209562026 4529392898 5992955029 1567910262 68393123 19641190 111670983 4782658 26762222 21219618 427764437 205546806 3444345529 29008442 13531024 44457252 102663993 13801024 17126160 79253072 39713973 24357791 133876465 43469927 167098753 84057913 119764.9 39262316 43064633 16122914 2929551 46469124 84866638 59227396 811635653 65579424 50152186 29401144 11618210 10338447 1709101 40083239 21609870 10591958 672941.1 900443.2 

cat("Alpha from R ETS(ANN):", round(alpha_r, 6), "\n")

Alpha from R ETS(ANN): 0.999899 

cat("These values should match your Excel Solver output.\n")

These values should match your Excel Solver output.

---

Summary

	Retail Sales	Unemployment
Best ETS model	insert from output	insert from output
Alpha (predicted)	0.3 to 0.5	0.6 to 0.9
Alpha (actual)	insert	insert
Beta (predicted)	0.05 to 0.15	0.01 to 0.10
Beta (actual)	insert	insert
Gamma (predicted)	0.05 to 0.15	Near zero or absent
Gamma (actual)	insert	insert
Intuition correct?	assess	assess

The key insight connecting McGuigan’s Chapter 5 to ETS theory is that all exponential smoothing models are governed by the same fundamental tradeoff: a high smoothing parameter produces a reactive model that tracks recent data closely but is susceptible to noise; a low smoothing parameter produces a stable model that resists short-term fluctuations but may be slow to detect genuine structural changes. The ETS framework simply applies this tradeoff separately and independently to the level, trend, and seasonal components of a series.