Interpreting Forecast Intervals Across Horizons

Reading Forecasts at Different Horizons

A forecast is rarely presented as a single number alone. A forecast of next month’s claim count may be fairly precise, while a forecast several months ahead has more opportunity for new, unanticipated events to affect the outcome. A prediction interval makes this uncertainty visible by giving a range of plausible values for a specified future observation.

When comparing horizons, read the horizon carefully: a forecast for one step ahead and a forecast for six steps ahead concern different future observations. Their point forecasts may differ because the series is expected to change, and their intervals may differ because forecast uncertainty changes with lead time. A larger point forecast does not by itself mean greater uncertainty; the interval width and, when useful, the width relative to the forecast provide that information.

Definition: Let \(Y_{T+h}\) be the observation \(h\) time steps after the end of the observed series, and let \(\hat{Y}_{T+h|T}\) be its point forecast using information available through time \(T\). A prediction interval for \(Y_{T+h}\) is an interval intended to contain that future observation with a stated probability under the forecasting model and its assumptions. Its endpoints depend on the forecast and the uncertainty at horizon \(h\).

A central prediction interval is designed to leave equal probability in the lower and upper tails under the assumed forecast-error distribution. For a normal-based interval, suppose the forecast error \(e_{T+h|T}=Y_{T+h}-\hat{Y}_{T+h|T}\), conditional on information through \(T\), is normally distributed with mean zero and standard deviation \(s_h\). This assumes the point forecast is the conditional mean and that the forecast-error distribution and its standard deviation are appropriately specified. Under these assumptions, the central 95% interval is

\left[ \hat{Y}_{T+h|T}-z_{0.975}s_h,\; \hat{Y}_{T+h|T}+z_{0.975}s_h \right], \qquad z_{0.975}\approx 1.959964.

Using \(1.96\) instead of the exact standard-normal quantile gives the familiar approximate 95% interval. The probability between \(-1.96\) and \(1.96\) is approximately 0.95, not exactly 0.95. The forecast-error standard deviation \(s_h\) is horizon-specific: it describes the spread of errors for forecasts made \(h\) steps ahead. Its calculation depends on the forecasting method and assumptions. Do not apply a particular method’s formula for \(s_h\) to a different method without justification.

Why an Interval Has This Form

The normal-based interval follows by standardizing the forecast error. If \(e_{T+h|T}\) is normally distributed with mean zero and standard deviation \(s_h\), then dividing it by \(s_h\) gives a standard normal variable \(Z\). The central probability between the standard-normal quantiles \(-z_{0.975}\) and \(z_{0.975}\) is 0.95. Multiplying the resulting inequalities by the positive value \(s_h\), then adding the point forecast, gives the prediction interval for the future observation.

P\left(-z_{0.975}\leq \frac{Y_{T+h}-\hat{Y}_{T+h|T}}{s_h} \leq z_{0.975}\right)=0.95

P\left(\hat{Y}_{T+h|T}-z_{0.975}s_h \leq Y_{T+h}\leq \hat{Y}_{T+h|T}+z_{0.975}s_h\right)=0.95.

This derivation depends on the stated normal forecast-error assumption and on the standard deviation being the one appropriate to horizon \(h\). If errors are not normal, the same endpoints need not have 95% coverage; other methods of constructing intervals may be needed. Even with a normal assumption, an estimated \(s_h\) and a fitted model make the claimed coverage conditional on the model being a reasonable description of the process.

Identify the horizon and target.
Check which future observation the forecast and interval describe. An interval for \(Y_{T+4}\) does not directly describe the sum of observations over the next four steps.

Read the point forecast and endpoints.
The point forecast gives the central predicted value; the endpoints describe a range for the future observation under the stated interval procedure.

Compare uncertainty on a suitable scale.
Use interval width for absolute uncertainty. If forecast levels differ substantially, also compare the half-width relative to the forecast, provided that ratio is meaningful for the measurement scale.

State the qualification.
Describe the interval as model-based probabilistic uncertainty, not as a guaranteed range for the future observation.

How Interval Width Changes Across Horizons

The full width of a symmetric interval is twice its margin of error. For a central normal-based interval with multiplier \(z\), the margin is \(zs_h\), and the full width is \(2zs_h\). Thus, for the same interval level, a larger forecast-error standard deviation means a wider interval. In many forecasting settings, uncertainty accumulates with horizon, so longer-horizon intervals are wider. This is a common pattern, not a rule that can be asserted without considering the model: the forecast-error uncertainty is method-specific and may not increase in the same way for every forecasting procedure.

Absolute width is not the only useful comparison. A margin of 10 units is large relative to a forecast of 20, but small relative to a forecast of 1,000. For a positive point forecast, a convenient relative margin is the interval half-width divided by the point forecast. It expresses the uncertainty scale as a proportion of the forecast. This ratio can be misleading if the forecast is near zero, negative, or measured on a scale where ratios do not have a meaningful interpretation. In those cases, report the absolute width and explain the context instead.

Worked Example: Comparing Short- and Long-Horizon Intervals

A hypothetical analyst forecasts a monthly service index. At horizon \(h=1\), the point forecast is 100 index units and the forecast-error standard deviation is 4 units. At horizon \(h=4\), the point forecast is 112 units and the forecast-error standard deviation is 10 units. Construct approximate central 95% intervals using \(z=1.96\), compare their widths, and compare their relative margins.

Method. Use the normal-based interval \(\hat{Y}_{T+h|T}\pm 1.96s_h\). This calculation assumes a suitably centered normal forecast error at each horizon and treats the supplied standard deviations as appropriate for those horizons.

Horizon 1. The margin is \(1.96(4)=7.84\) units. The interval is \([100-7.84,\;100+7.84]=[92.16,\;107.84]\). Its full width is \(107.84-92.16=15.68\) units. Its relative margin is \(7.84/100=0.0784\), or 7.84%.

Horizon 4. The margin is \(1.96(10)=19.60\) units. The interval is \([112-19.60,\;112+19.60]=[92.40,\;131.60]\). Its full width is \(131.60-92.40=39.20\) units. Its relative margin is \(19.60/112=0.175\), or 17.50%.

The horizon-4 interval is wider in absolute terms, 39.20 units compared with 15.68, and its relative margin is also larger. The intervals overlap, but overlap does not mean that the two horizons have the same point forecast or the same uncertainty. Each interval concerns a different future observation.

Common mistake avoided. Do not compare only the point forecasts, 100 and 112, and conclude that the larger forecast is less certain. The supplied forecast-error standard deviations and the resulting interval widths are what quantify uncertainty here.

Prediction Intervals Are Not Guaranteed Bounds

A 95% prediction interval is a probabilistic statement under a model, not a promise that every future value will fall inside the endpoints. If the interval procedure is correctly calibrated under repeated applications of its assumptions, about 95% of the corresponding future observations are expected to be covered in the long run. Some observations can still fall outside. An observation outside an interval may be unusual under the model, but it is possible; it does not, on its own, establish why the observation occurred or prove that the model is wrong.

A prediction interval for a future observation must also be distinguished from an interval expressing uncertainty about an estimated mean or other model quantity. A future observation includes both uncertainty in its expected level and the variation of the observation around that level. An interval for the mean concerns the mean itself and generally does not include all of that future-observation variation. Read the target in the question before interpreting the endpoints.

Worked Example: Checking an Observation Against an Interval

For a hypothetical monthly workload, a model gives a point forecast of 72 units at a specified horizon and a forecast-error standard deviation of 8 units. Assume a normal forecast error and use \(1.96\) for an approximate central 95% prediction interval. The observed workload is 90 units. Determine whether it is inside the interval and interpret the result.

Method. Construct the interval first, then standardize the observed forecast error. The normality and standard-deviation assumptions are needed for the interval and probability interpretation.

The margin is \(1.96(8)=15.68\). Therefore the interval is \([72-15.68,\;72+15.68]=[56.32,\;87.68]\) units. The observed value, 90, is 2.32 units above the upper endpoint, so it is outside the interval.

The standardized forecast error is \((90-72)/8=18/8=2.25\). Under the assumed standard normal error distribution, the two-sided probability of an error at least this far from zero is approximately \(P(|Z|\geq 2.25)=0.0244\). Thus the result is relatively unusual under the assumed model, but it is not impossible and does not make the interval a failed guarantee. It may prompt a closer look at the model, the data, or unusual conditions in that month.

Common mistake avoided. Do not say that a 95% interval must contain the next observation. The 95% describes the intended model-based coverage across repeated forecasts, not certainty about this particular outcome.

Interpreting Width and Coverage Carefully

When a report presents forecasts at several horizons, use consistent language. State the forecast value, the horizon, the interval level, and the interval endpoints. Then explain whether the interval is wider or narrower and what that means on the measurement scale. For example, “The longer-horizon interval is wider, indicating greater forecast uncertainty in absolute units under this model” is more precise than “the long-term forecast is less accurate.” A wider interval signals greater modeled uncertainty; it does not by itself establish the realized accuracy of a forecast.

For a symmetric interval, the point forecast is its midpoint. That property is a consequence of the symmetric construction, not a guarantee that the underlying outcome distribution is symmetric. When the distribution or interval method produces asymmetric endpoints, do not force a symmetric interpretation or assume the point forecast is the midpoint. Follow the procedure and assumptions that produced the interval.

Use the correct horizon. The uncertainty measure for one step ahead need not apply to a longer lead time.
Separate level from spread. A changing point forecast and a changing interval width answer different questions.
Distinguish absolute and relative uncertainty. Report the scale used, and avoid relative margins when the denominator makes them unstable or meaningless.
Keep the coverage qualification. A stated coverage level depends on the interval method and its assumptions; it is not a guaranteed bound for an individual future value.

These interpretations are aligned with the SOA Exam SRM (Statistics for Risk Modeling) syllabus. A strong response identifies what is being forecast, compares the interval at the relevant horizons, and interprets uncertainty without turning a probabilistic statement into certainty.

Key takeaway: A forecast interval describes model-based uncertainty for a specified future observation at a specified horizon. Compare its width on an appropriate scale, and interpret its coverage as probabilistic—not as a promise that future values cannot fall outside it.

Check Your Understanding

For each problem, identify what the interval says about the relevant future observation and retain the stated assumptions when interpreting coverage.

A hypothetical forecast is 50 units, with forecast-error standard deviation 3 units. Find an approximate central 95% prediction interval using \(1.96\). (Answer: \([44.12,\;55.88]\) units.)
For a positive index, the one-step forecast is 80 with standard deviation 5; the six-step forecast is 100 with standard deviation 12. Use \(z=1.645\) for an approximate central 90% interval. Give both intervals and relative margins, rounding endpoints to two decimals and percentages to two decimals. (Answer: One step: \([71.78,\;88.23]\), relative margin \(8.225/80=10.28\%\); six steps: \([80.26,\;119.74]\), relative margin \(19.74/100=19.74\%\).)
A point forecast is 18.0 units and its forecast-error standard deviation is 2.5 units. Find the approximate central 95% interval using \(1.96\), its half-width, and its relative margin. Round the endpoints and half-width to two decimals and the relative margin to four decimals. (Answer: \([13.10,\;22.90]\); half-width \(4.90\); relative margin \(4.90/18.0=0.2722\), or 27.22%.)
A hypothetical model gives a forecast of 125 units with forecast-error standard deviation 10 units. Using \(1.96\), find the approximate central 95% interval. If the observation is 146, find its standardized forecast error and explain whether the observation is impossible under the model. (Answer: Interval \([105.40,\;144.60]\); standardized error \((146-125)/10=2.10\); the observation is outside the interval but not impossible.)
At two horizons, a hypothetical positive series has forecasts 30 and 60 units, with forecast-error standard deviations 4 and 8 units, respectively. Use \(1.96\) to find both approximate central 95% intervals and compare absolute and relative margins. (Answer: First interval \([22.16,\;37.84]\), half-width 7.84 and relative margin \(7.84/30=26.13\%\); second interval \([44.32,\;75.68]\), half-width 15.68 and relative margin \(15.68/60=26.13\%\). The second interval is twice as wide in absolute units, while the relative margins are equal.)

What You'll Learn