Introduction
A forecasting method is useful only when its assumptions fit the pattern in the data and its forecasts work at the horizons where decisions will be made. A method that tracks a changing level can be needlessly reactive for a stable series; a straight-line trend can be implausible far into the future; and a seasonal adjustment can be badly scaled if it is applied additively when the seasonal size grows with the level.
Choosing and checking a method therefore involves two distinct tasks: identify whether the series is mainly stable, trending, or seasonal, and evaluate forecasts on observations that were not used to fit or initialize the method. The evaluation must preserve time order and match each forecast to the observation it was intended to predict. These choices are aligned with the SOA Exam SRM (Statistics for Risk Modeling) syllabus.
Match the Method to the Pattern
Begin by describing the pattern in time order, not by choosing a method from the name of the variable. Look for a stable level, a sustained trend, and a repeating seasonal pattern. A pattern can contain more than one of these features. The method should represent the important features without extending them beyond what the data reasonably support.
| Pattern to forecast | Suitable method family | What to check |
|---|---|---|
| Approximately stable level, without clear trend or seasonality | Simple exponential smoothing or a level-based benchmark | Whether the level remains plausible at the required horizons |
| Changing level with a persistent trend | Trend smoothing, including damped-trend smoothing | Whether the trend continues and whether its long-run extrapolation is reasonable |
| Repeating seasonal pattern | Seasonal smoothing, with additive or multiplicative seasonality as appropriate | Whether seasonal effects stay roughly constant in size or grow with the level |
This is a guide, not a guarantee that one method is always best for a pattern. For example, short-term trend may be evident even when a long-range straight-line forecast is not credible. Damped-trend smoothing allows the contribution of the trend to diminish with the horizon. For level \(l_T\), trend \(b_T\), damping parameter \(0<\phi<1\), and horizon \(h\), its point forecast is
A stable-level method should not be forced to explain a clear trend, and a seasonal method should not be used just because the data fluctuate. A seasonal pattern must recur at a known period and be supported by the available history. The additive form treats the seasonal effect as an amount; the multiplicative form treats it as a factor.
Evaluate Forecasts on the Correct Observations
Let \(y_{T+h}\) be the actual value and \(\hat{y}_{T+h|T}\) its forecast. Define the forecast error as actual minus forecast:
A positive error means the forecast was too low; a negative error means it was too high. Two common measures over \(n\) evaluated forecasts are mean absolute error and mean squared error. Root mean squared error is the square root of mean squared error.
MAE is expressed in the same units as the observations. MSE gives relatively more weight to large errors because errors are squared; RMSE returns to the original units. Neither measure reveals whether errors tend to be positive or negative, so inspect the signed errors as well.
The distinction matters for smoothing methods. With simple exponential smoothing, the level updates according to \(l_t=\alpha y_t+(1-\alpha)l_{t-1}\), where \(0<\alpha\leq1\); its one-step forecast is \(l_t\). After the next actual value arrives, the level changes and so does the next rolling forecast. A fixed-origin forecast, by contrast, holds the origin fixed and predicts several future periods without using intervening actuals.
Worked Example: Rolling One-Step Forecasts for a Stable Level
A hypothetical claims-processing team wants to forecast weekly intake, measured in claims. Use simple exponential smoothing with \(\alpha=0.5\), initialize \(l_1=y_1=100\), and use \(y_1,\ldots,y_4=100,104,102,104\) as training observations. The next three actual values, kept out of the initial fit, are \(105,101,104\). Calculate rolling one-step forecasts and their MAE, MSE, and RMSE.
Method. This is a level-only method. For each training observation after the initial value, update the level using \(l_t=0.5y_t+0.5l_{t-1}\). The training updates are
The first holdout forecast is \(\hat{y}_{5|4}=l_4=103\), so \(e_5=105-103=2\). Now update with the observed value \(105\): \(l_5=0.5(105)+0.5(103)=104\). The second forecast is \(\hat{y}_{6|5}=104\), with \(e_6=101-104=-3\). Update again: \(l_6=0.5(101)+0.5(104)=102.5\). The third forecast is \(\hat{y}_{7|6}=102.5\), with \(e_7=104-102.5=1.5\).
Thus the absolute errors are \(2,3,1.5\), and the squared errors are \(4,9,2.25\). Therefore
The check is that the forecasts use only information available before their respective target weeks, and the squared errors sum to \(15.25\). The common mistake is to use \(103\) for all three targets: that would be a fixed-origin forecast, not the rolling one-step evaluation requested.
Check More Than One Horizon
Forecast performance can depend on the horizon. A trend method may beat a level method at short horizons but become less reliable as its trend compounds. Conversely, a model that performs well at one step may fail to capture the cumulative change over several periods. When decisions require forecasts at different horizons, report errors separately by horizon whenever enough validation forecasts are available.
For a fixed-origin evaluation, all forecasts in a block share one origin and are made without updating on the intervening actuals. For a rolling-origin evaluation, the origin advances as new actuals become available. Both are legitimate, but they answer different questions. In either case, preserve time order: validation observations must occur after the data used to fit or initialize the method.
Worked Example: Comparing Level and Damped-Trend Forecasts
A hypothetical monthly benefit-payment series has a training-end level \(l_T=50\) thousand dollars and trend \(b_T=4\) thousand dollars per month. A damped-trend method uses \(\phi=0.8\). For a three-month fixed-origin validation block, the actual values are \(54,57,58\) thousand dollars. Compare its forecasts with a level-only method forecasting \(50\) thousand dollars at every horizon.
Method. Keep the origin fixed for all three predictions. The damped-trend increments over the level are \(4(0.8)=3.2\), \(4(0.8+0.8^2)=5.76\), and \(4(0.8+0.8^2+0.8^3)=7.808\). Thus the damped forecasts are \(53.2,55.76,57.808\). The level forecasts are \(50,50,50\).
| Horizon | Actual | Damped forecast | Damped error | Level forecast | Level error |
|---|---|---|---|---|---|
| 1 | 54 | 53.2 | 0.8 | 50 | 4 |
| 2 | 57 | 55.76 | 1.24 | 50 | 7 |
| 3 | 58 | 57.808 | 0.192 | 50 | 8 |
For the damped forecasts, the absolute errors sum to \(0.8+1.24+0.192=2.232\), and the squared errors sum to \(0.64+1.5376+0.036864=2.214464\). Hence MAE is \(2.232/3=0.744\), MSE is \(2.214464/3=0.738155\), and RMSE is approximately \(0.8592\) thousand dollars. For the level forecasts, MAE is \((4+7+8)/3=6.3333\), MSE is \((16+49+64)/3=43\), and RMSE is \(\sqrt{43}\approx6.5574\) thousand dollars.
The check is that both methods are scored against the same three actual values and in the same units. In this block the damped trend has smaller errors at each horizon. That does not establish that it will always be superior; it is evidence from this particular validation block. The common mistake is to update the trend method with the first or second holdout actual while still calling the results fixed-origin forecasts.
Seasonality: Choose the Scale and Align the Cycle
For additive seasonality, the seasonal component is measured in the units of the series and is added to the underlying level or trend forecast. For multiplicative seasonality, the seasonal component is a dimensionless factor and multiplies the underlying forecast. Seasonal positions must also be aligned correctly: the factor for one quarter, month, or other seasonal position is applied only to forecasts for that position.
An additive pattern is appropriate when seasonal rises and falls are roughly constant in absolute size as the level changes. A multiplicative pattern is appropriate when those changes are better described as proportions of the level. Seasonal components supplied by a fitted method should be estimated from training data only; estimating them using the holdout values would leak evaluation information into the forecasts.
Worked Example: Checking Additive and Multiplicative Seasonal Forecasts
A hypothetical quarterly portfolio has underlying forecasts, in thousands of dollars, of \(200,210,220,230\) for the next four quarters. Seasonal components estimated from its training data are additive effects \(-10,0,10,0\), or multiplicative factors \(0.95,1.00,1.05,1.00\), in quarter order. The actual holdout values are \(192,208,235,227\). Compare the two seasonal forms.
Method. Apply each seasonal component to the underlying forecast in its quarter. The additive forecasts are \(200-10=190\), \(210+0=210\), \(220+10=230\), and \(230+0=230\). The multiplicative forecasts are \(200(0.95)=190\), \(210(1.00)=210\), \(220(1.05)=231\), and \(230(1.00)=230\).
For the additive forecasts, actual-minus-forecast errors are \(2,-2,5,-3\). The absolute errors sum to \(12\), and the squared errors sum to \(4+4+25+9=42\). Thus MAE is \(12/4=3\), MSE is \(42/4=10.5\), and RMSE is \(\sqrt{10.5}\approx3.2404\) thousand dollars.
For the multiplicative forecasts, the errors are \(2,-2,4,-3\). The absolute errors sum to \(11\), and the squared errors sum to \(4+4+16+9=33\). Thus MAE is \(11/4=2.75\), MSE is \(33/4=8.25\), and RMSE is \(\sqrt{8.25}\approx2.8723\) thousand dollars. The check is that the seasonal effects in the additive model are amounts, while the multiplicative components are factors; the units of both resulting forecasts are still thousands of dollars.
The multiplicative form scores better on this four-quarter block, but the evidence is limited to these observations. The common mistake is to add \(1.05\) to the underlying value, or to multiply by an additive effect such as \(10\). Either operation confuses the scale of the seasonal component.
Deriving the Exponential-Smoothing Forecast
The level update also explains why simple exponential smoothing produces a flat forecast beyond its origin. Starting with \(l_t=\alpha y_t+(1-\alpha)l_{t-1}\), substitute the same update for the previous level:
Repeating the substitution gives a weighted average of past observations and the initial level, with more recent observations receiving larger weights. The weights on observed values are \(\alpha,\alpha(1-\alpha),\alpha(1-\alpha)^2,\ldots\). Because this method has no trend or seasonal state, its forecast from origin \(T\) is the current estimated level for every future horizon:
That flat forecast is a property of the model, not a claim that the actual series cannot change. If the series has a persistent trend or recurring seasonality, a level-only forecast may systematically miss those features.
Common Forecasting Traps
- Using the wrong forecast origin. A rolling forecast must be updated after each newly observed value. A fixed-origin forecast must not use those future observations.
- Scoring training fits as if they were validation forecasts. A fitted value calculated using an observation is not a genuine forecast of that observation from an earlier origin.
- Mixing horizons. A pooled average can conceal poor performance at the horizon that matters. Preserve horizon information when possible.
- Ignoring extrapolation risk. A trend that fits recent data may not continue indefinitely. Damping limits, but does not eliminate, the effect of trend in long-range forecasts.
- Confusing seasonal scales or positions. Add seasonal amounts and multiply by seasonal factors; apply each component to the correct position in the cycle.
- Comparing errors on incompatible scales or periods. Compare methods on the same holdout observations and units. A percentage error is undefined when the actual value is zero, so it is not a universal replacement for MAE or RMSE.
A practical check is to state the forecast origin, horizon, and update rule before calculating errors. Then verify that every forecast uses only information available at that origin, that actual and forecast values share units and seasonal positions, and that the chosen accuracy measure answers the decision question.
Check Your Understanding
For each problem, identify the relevant method or calculation and check that the forecast timing and scale are consistent.
- Simple exponential smoothing has \(\alpha=0.25\), initial level \(l_1=40\), and subsequent training observations \(y_2=44\) and \(y_3=42\). What is the one-step forecast for \(y_4\)? (Answer: \(41.25\))
- A damped-trend forecast has \(l_T=80\), \(b_T=5\), and \(\phi=0.6\). Calculate the fixed-origin two-step forecast. (Answer: \(84.8\))
- Underlying forecasts for three successive seasonal positions are \(50,52,54\), and the corresponding additive seasonal effects are \(-4,1,3\). Give the three seasonal forecasts in order. (Answer: \(46,53,57\))
- Three actual values are \(30,34,31\), and their forecasts are \(32,33,29\). Calculate MAE and RMSE, using actual minus forecast for the errors. (Answer: MAE \(=5/3\approx1.6667\); RMSE \(=\sqrt{3}\approx1.7321\))
- A seasonal method has an underlying forecast of \(120\) for each of three positions and multiplicative factors \(0.9,1.0,1.1\). The holdout actuals are \(108,126,126\). Calculate the forecasts and their MAE. (Answer: Forecasts \(108,120,132\); MAE \(=4\))