Exponential Smoothing Exercises in R: 50 Practice Problems

Explanation: SES has one state, the level, and no trend or seasonal term, so every forecast horizon inherits the same final level estimate. That flat line is a feature, not a bug: it is the right answer when a series wanders around a slowly drifting mean with no persistent direction. If your series clearly trends, the flat forecast is your signal to move up to holt().

Explanation: ses() returns a forecast object whose $model slot is a fitted ets object, and $par holds every estimated parameter including the initial state. An alpha near 0.25 means each new observation shifts the level by only a quarter of its surprise, so the fit is fairly smooth. Alpha near 1 would make the model chase the latest value like a naive forecast.

Explanation: The optimal fit must have the smaller error sum, because that is the objective the optimizer minimizes. The gap here is tiny, roughly 0.2 percent, which tells you the SSE surface is flat near the optimum: a range of alpha values fits this series almost equally well. That flatness is common with short annual series and is a good reason not to over-interpret the exact alpha.

Explanation: Writing the recursion yourself proves there is no magic in ses(): it estimates two numbers, an initial level and alpha, then runs this one-line update through the data. The row states[1, ] is the pre-sample state $\ell_0$, not the first fitted level, which is why the loop starts there. Getting the same answer as the package is the cleanest possible check that you understand the model.

Explanation: This is why the method is called exponential smoothing: the weights decline geometrically, so the observation from six periods ago counts about a sixth as much as today's. The weights sum to one in the limit, and after 20 terms they already account for 99.9 percent of the total. A smaller alpha stretches this decay out and produces a smoother, more sluggish level.

Explanation: By default ses() estimates $\ell_0$ jointly with alpha by maximum likelihood, which is why the states matrix has one more row than the series has observations. Passing initial = "simple" instead fixes $\ell_0$ at the first observation and estimates only alpha. On short series the choice matters, because a badly placed starting level takes many periods to wash out.

Explanation: Alpha is pinned at its upper bound, so the level equation reduces to $\ell_t \approx y_t$ and the forecast is essentially the last observed value. That is the naive forecast wearing a disguise. An alpha of 1 is a diagnostic, not a result: it says the level-only model has no way to express the trend it can see, and you should fit holt() instead.

Explanation: Alpha collapses to its lower bound, which turns the level into an equally weighted average of the whole history. That is exactly right for a series with no drift: averaging 16 weeks beats chasing last week's noise. Alpha near 0 and alpha near 1 are the two informative extremes, and here the low value is telling you the series has no memory worth tracking.

Explanation: The forecast is $\ell_T + h b_T$, a straight line extrapolating the final level and slope. WWWusage was falling at the end of the sample, so the line slopes down by two users per minute per period. This is the method's biggest risk: nothing stops that line, so a long horizon can produce forecasts far outside anything the series has ever done.

Explanation: Alpha weights new information into the level and beta weights it into the slope. Both sit at their upper bound here, meaning the slope is re-estimated almost entirely from the most recent change, which makes the trend jumpy. Beta near 0 would give you an almost fixed slope estimated across the whole sample. Note beta in the forecast package is the $\beta^*$ of the standard textbook parameterisation.

Explanation: With damping the forecast becomes $\ell_T + (\phi + \phi^2 + \dots + \phi^h) b_T$, so each extra step contributes less than the one before. With $\phi = 0.81$ the increments shrink by about 19 percent per period and the path converges to an asymptote. Damped trend is the default recommendation for automated forecasting because it rarely blows up on long horizons.

Explanation: After 30 steps the undamped line has fallen 60 units while the damped path has flattened out near 211. The gap grows linearly with the horizon, which is why damping matters most for long-range planning and hardly at all for one-step forecasting. If a stakeholder needs a 12 month view, this single comparison usually settles the choice.

Explanation: Lower AICc wins, so the damped model is preferred by about 10.7 units, which is a decisive margin. AICc penalises the extra phi parameter and the damped fit still comes out ahead, meaning the improvement in likelihood is real rather than a reward for flexibility. Note that AICc is only comparable across models fitted to the same data on the same scale.

Explanation: The states matrix is the whole model laid out in the open, and its last row is everything the forecast function needs. Matching the package output by hand confirms that a Holt forecast contains no information beyond those two numbers. The tiny gap in the third decimal comes from rounding the states, not from a different formula.

Explanation: The order matters: the slope update needs the change in level, so the level must be updated first and the previous value stored. Getting these two lines in the wrong order is the classic bug when people implement Holt by hand. Notice the level equation smooths towards level_prev + b, the one-step forecast, not towards level_prev as in SES.

Explanation: Beta at its lower bound means the slope is treated as near constant, estimated across all 16 months instead of from the last few. Phi at 0.98 is the upper bound forecast allows, so the damping is mild and the six-month path still rises steadily. Reporting these three parameters alongside the numbers is what makes the projection defensible in a review.

Explanation: Holt-Winters adds a third state, a vector of seasonal terms, one per period of the seasonal cycle. Additive seasonality adds a fixed number of ppm to each month, which suits co2 because the May peak sits about the same distance above trend every year. The forecast keeps rising because the trend state carries on while the seasonal pattern repeats.

Explanation: Multiplicative seasonality multiplies the level by a factor per month, so a July index of about 1.16 means July runs 16 percent above the deseasonalized level whatever that level is. Fitting an additive model here would understate the summer peak in later years and overstate it in the early ones. The visual test is simple: fan-shaped seasonal swings mean multiplicative.

Explanation: Both measures point the same way and the margin is enormous: 160 AICc units and a 37 percent cut in RMSE. AICc and training RMSE do not always agree, because AICc penalises parameters while RMSE does not, so when they do agree you can stop worrying. Both models have the same parameter count here, which makes the AICc comparison a pure likelihood comparison.

Explanation: The two approaches are close but not identical, and the reason is the error term. Log-additive assumes additive errors on the log scale, which is multiplicative error on the original scale, while hw(seasonal = "multiplicative") keeps additive errors on the original scale. Back-transforming with exp() also returns the median rather than the mean, which biases the point forecast slightly low unless you apply a bias adjustment.

Explanation: Indices above 1 mark months that run above the deseasonalized level and indices below 1 mark months that run below it. Read them in reverse chronological order: s1 is the most recent month in the cycle, not January. That indexing convention trips up almost everyone the first time, so always check against a month you know is a peak before labelling a chart with these numbers.

Explanation: HoltWinters() initialises its states from a decomposition of the first few cycles and optimises squared error, while hw() estimates initial states and parameters jointly by maximum likelihood in a state space framework. The parameter estimates differ as a result. Prefer hw() or ets() in new work: only the state space version gives you AICc for model selection and analytic prediction intervals.

Explanation: Here damping loses by 14 AICc units, so the extra parameter does not earn its keep. Phi sits at the 0.98 ceiling that forecast imposes, another sign the data wants no damping at all: the trend in AirPassengers is genuine and persistent. Always fit both rather than assuming damped trend is the safer default on every series.

Explanation: If you forget frequency = 4 the series is treated as non-seasonal and hw() stops with an error about the data not being seasonal. Four years of quarterly data is also close to the practical minimum: seasonal exponential smoothing needs at least two full cycles to estimate the seasonal states, and more cycles make those estimates far steadier.

Explanation: The states tell the planner the story behind the numbers: an underlying level near 302 growing about 4.9 per quarter, with the most recent quarter in the cycle running 27 percent above that level. Reading s1 as Q4 rather than Q1 is what makes the interpretation correct. The planner can now stress test the forecast by asking what happens if the Q4 uplift softens to 1.15.

Explanation: Read the label as error Multiplicative, trend Additive damped, seasonal Multiplicative. Multiplicative error means the noise scales with the level, which is exactly what airline data does. ets() fits every admissible combination and keeps the lowest AICc, so the label is a summary of a search, not an assumption you imposed. Always print the label before you trust the forecast.

Explanation: The forced additive model is 170 AICc units worse, a gap so large that no reasonable tolerance would call them competitive. Rules of thumb put a difference above 10 in decisive territory. Forcing a model is still worth doing when you have a strong reason, for example when a downstream system requires additive components, but you should know what it costs.

Explanation: The ranking separates cleanly into multiplicative seasonal models on top and additive seasonal models at the bottom, which confirms the seasonal form matters far more here than the trend does. Note that MNM, with no trend at all, still beats both additive seasonal models. When you are unsure which component to fix first, a small grid like this answers it in one line.

Explanation: The search picks N for trend and N for seasonal, so this is simple exponential smoothing with a multiplicative error. ses() by default fits ETS(A,N,N), the additive-error version, which is why its alpha of 0.2457 differs from the 0.1514 here. Same forecast function, different noise assumption: the point forecasts stay flat either way but the prediction intervals do not.

Explanation: Parsing the method string with regular expressions is fragile, so use $components whenever code needs to branch on the selected form, for instance to decide whether a back-transformation is required. The fourth element is the damping flag stored as a character, so compare it with "TRUE" rather than the logical value. This is the safe way to log model choices across a batch of series.

Explanation: The search reaches the same conclusion the manual AICc comparison in Exercise 2.5 reached: damped trend wins. This is the argument for using ets() as the default entry point rather than hand-picking between ses(), holt() and hw(). Those three functions are convenience wrappers around fixed corners of the same model space that ets() searches for you.

Explanation: The transform stabilises the variance, so an additive model now fits about as well as the multiplicative one did, with an RMSE of 10.4 against 10.7. The AICc is negative and cannot be compared with the untransformed models, because the likelihood is computed on a different scale. That is the classic trap: only compare AICc across models fitted to the identically transformed data.

Explanation: These are in-sample one-step errors, so they flatter the model: it saw every observation while fitting. A MAPE of 2.9 percent looks excellent, but the honest number comes from a holdout, which Section 6 covers. Use training accuracy to compare models on the same data, never to promise a stakeholder how a forecast will perform next quarter.

Explanation: By default forecast() returns two levels, 80 and 95 percent, so column 1 is the 80 percent bound and column 2 is the 95 percent bound. Getting that column order backwards is a common reporting error that quietly narrows every published interval. Pass level = 95 if you only want one, and the object will then have a single column.

Explanation: The null hypothesis is that the residuals are independent, and a p-value of 0.001 rejects it. Some structure remains that the model has not captured, which is common with AirPassengers because its seasonal shape shifts slightly over the decade. A failed test does not make the forecast useless: it warns that the prediction intervals are probably too narrow.

Explanation: With 144 observations the rough significance bound is $2/\sqrt{144} = 0.167$, so the lag-4 value of -0.229 is clearly outside it while the first two are not. That single spike is what drove the Ljung-Box rejection in the previous exercise. Reading the individual lags tells you where the leftover structure sits, which a single portmanteau p-value hides.

Explanation: The interval nearly triples over a year, which is the honest cost of forecasting further out. Because this is a multiplicative model, the width also tracks the seasonal level, so it does not grow perfectly smoothly: high-season months carry wider intervals than low-season ones. A single point forecast without this context invites false confidence.

Explanation: The interval widens slowly here because alpha is only 0.25, so $(h-1)\alpha^2$ adds about 0.06 of extra variance per step. With alpha near 1 the same formula would widen almost like a random walk. Using qnorm(0.975) rather than 1.96 matters at the second decimal, and mismatches at that level are usually this rounding rather than a wrong formula.

Explanation: A 50 percent interval is the range the outcome falls inside half the time, which many stakeholders find easier to reason about than a 95 percent band that looks alarmingly wide. Showing both is good practice for a planning document: the inner band frames the base case and the outer band frames the risk. Column names carry the level, so you can subset by "95%" rather than by position.

Explanation: Innovation residuals are the relative errors the multiplicative model actually assumes to be white noise, so a value of 0.064 means the observation came in 6.4 percent above the one-step forecast. Response residuals are in passengers and grow with the level. Diagnostic tests belong on the innovation residuals, but accuracy reporting belongs on the response scale, and residuals() defaults to innovation.

Explanation: A p-value of 0.48 gives no evidence against normality, so the analytic intervals rest on a reasonable assumption here. When this test fails, switch to bootstrapped intervals with forecast(fit, bootstrap = TRUE), which resamples the observed residuals instead of assuming a shape. Note that normality matters for the intervals only: the point forecasts do not depend on it.

Explanation: Actual coverage of 87.5 percent falls short of the nominal 95 percent, the usual finding: analytic intervals ignore the uncertainty in the estimated parameters, so they run too narrow. Three misses out of 24 is a small sample, so treat this as a warning rather than a measurement. If a plan depends on the tail, widen the intervals by bootstrapping or hold back more data.

Explanation: The test MAPE of 13.3 percent dwarfs the 2.9 percent training MAPE from Exercise 4.8, which is the whole reason holdouts exist. RMSE punishes large misses more than MAE does, so a gap between them signals a few bad months rather than uniform drift. MAPE is scale free and easy to explain, but it breaks down on series that pass near zero.

Explanation: MASE scales errors by the average one-step seasonal naive error on the training data, so a value above 1 means the forecast is worse than that naive rule was in sample. Both models exceed 1 here because the test period grew faster than the training period did. The comparison that matters is relative: ETS beats seasonal naive by about 11 percent, a real but modest edge.

Explanation: SES sits between the two extremes it interpolates and beats both: the naive forecast reacts to every wobble, the mean forecast reacts to nothing, and SES tunes the balance through alpha. tsCV() returns a series of errors with NA where there was not enough history, hence na.rm = TRUE. This is a far more honest comparison than a single fixed split, because every observation gets forecast once.

Explanation: The ranking matches what AICc said in Exercise 2.5, which is reassuring: two independent criteria, one penalised likelihood and one out-of-sample error, pick the same model. Adding a trend cuts the error by a third and damping shaves a little more. When AICc and cross-validation disagree, trust cross-validation, because it measures the thing you actually care about.

Explanation: ETS wins by about 2 percent, which is close enough that either model is defensible on this series. The important discipline is that AICc from ets() and AICc from auto.arima() are not comparable, because the likelihoods are computed on differently differenced data. A holdout or tsCV() is the only fair referee between the two families.

Explanation: The combination lands between the two, slightly worse than the better model and better than the worse one, which is what happens when both models make errors in the same direction. Averaging pays off when errors are less correlated, and its real value is insurance: you avoid betting everything on the model that happened to win the last backtest. Note accuracy() accepts a bare ts of point forecasts, so no interval is produced for the combination.

Explanation: Naming the list once and mapping over it keeps the code honest: every model is scored on the same test window with the same function, so nobody can accuse the table of a rigged comparison. All three measures rank the models identically here, which is the easy case. When they disagree, say so in the report and name the measure you optimised.

Explanation: Running this check before fitting turns a cryptic error into a data problem you can act on: treat the outage month as missing and interpolate it, or fall back to an additive model. Never quietly replace the zero with a small number, because that inflates the estimated growth rate afterwards. The same guard belongs in any automated pipeline that lets ets() choose the model.