Why Precision Is a Reporting Choice
A model can produce a number with many digits, but those digits do not automatically represent knowledge. A prediction of 127.846 claims, a probability of 3.746%, or an RMSE of $14.263 may be the output of software; the report still has to decide what precision is useful and justified. Reporting every available digit can make an uncertain result sound exact, while rounding too aggressively can obscure a distinction that matters for a decision.
Precision is about the detail shown in a number. Accuracy is about how close that number is to the quantity of interest. Neither a long decimal nor a narrow-looking interval guarantees accuracy. In a report, the estimate, its uncertainty, the unit, and the wording should work together: readers should be able to tell what was estimated and how much confidence to place in the displayed detail.
There is no universal rule that every result should have two decimal places or three significant figures. The appropriate rounding depends on the result's uncertainty, the scale of its units, and the decision being supported. A report should also preserve enough precision to reproduce calculations and should avoid rounding intermediate values before a final result is calculated.
A Practical Rule for Choosing Rounding
A useful starting point is to compare the rounding increment with the uncertainty in the estimate. Suppose an estimate is rounded to the nearest increment \(q\). Under ordinary rounding to the nearest value, the rounding error is at most \(q/2\). This does not measure model uncertainty; it shows how much extra numerical change the reporting choice itself can introduce.
For example, rounding to the nearest whole unit gives \(q=1\), so the rounding error is at most \(0.5\) unit. If the estimate's uncertainty spans many units, showing tenths may contribute little useful information. If a decision changes at a boundary less than one unit away, whole-unit rounding may conceal a relevant distinction; the report may need more digits, a clear threshold comparison, or both.
This rule is a reporting check, not a substitute for uncertainty analysis. Rounding should not be used to make an uncertain result appear decisive. State uncertainty using the uncertainty summary appropriate to the analysis, such as a range or interval, and make clear what that summary represents. Avoid describing an interval as a guarantee that the true value lies within it.
Worked Example: Reporting a Forecasted Claim Count
A hypothetical insurer's model produces a forecast of 127.846 claims for a defined future portfolio period. An uncertainty analysis reports a 90% interval from 111.2 to 145.9 claims. The report is for managers planning staffing in whole claims. Choose a readable presentation and assess whether the raw digits are informative.
Method. Match the point estimate and interval to the count scale, then state the uncertainty rather than letting the point estimate stand alone. The supplied interval is an analysis result; no new uncertainty calculation is being inferred from the point estimate.
Step 1: Choose the display increment. Staffing is planned in whole claims, so use the nearest whole claim for the forecast and interval endpoints. The increment is \(q=1\) claim, making the maximum rounding error \(q/2=0.5\) claim.
Step 2: Round consistently. The point estimate \(127.846\) rounds to 128 claims. The lower endpoint \(111.2\) rounds to 111, and the upper endpoint \(145.9\) rounds to 146. The reported interval therefore spans 111 to 146 claims.
Step 3: Compare rounding with uncertainty. The supplied interval's width is \(145.9-111.2=34.7\) claims. Rounding each displayed quantity to a whole claim changes it by no more than 0.5 claim, much less than the interval's width. Tenths or thousandths do not provide useful staffing detail here.
A suitable sentence is: “The forecast is about 128 claims; the 90% uncertainty interval is approximately 111 to 146 claims.” Calling 128 an exact number of future claims would be misleading: it is a model estimate, and actual claims are counted in whole numbers. Common mistake avoided: copying 127.846 into the report as though software precision made the forecast exact.
Probabilities: Keep the Scale and the Uncertainty Clear
Probabilities are often easier for decision-makers to interpret as percentages, but the conversion must be consistent. A probability \(p\) corresponds to \(100p\%\). Convert both the estimate and any interval endpoints using the same factor, then round to a level that does not suggest greater certainty than the uncertainty analysis supports. State the event and time period, not just a percentage.
Worked Example: Expressing an Estimated Event Probability
For a hypothetical group of commercial refrigeration units, a model estimates the probability of at least one temperature excursion during the next 12 months as \(0.03746\). The reported 90% uncertainty interval for this probability is \([0.0291, 0.0478]\). Prepare a sentence for an operational report, in percentage terms.
Method. Convert the estimate and both interval bounds to percentages, check their rounding, and identify the event and time horizon. The uncertainty interval is supplied; this example does not derive it.
Step 1: Convert the point estimate. Multiply by 100: \(0.03746 \times 100 = 3.746\%\).
Step 2: Convert the interval bounds. The lower bound is \(0.0291 \times 100=2.91\%\), and the upper bound is \(0.0478 \times 100=4.78\%\). The interval is therefore \([2.91\%,4.78\%]\) before rounding.
Step 3: Select useful precision. Reporting the point estimate to one decimal percentage point gives 3.7%. Reporting the bounds to one decimal percentage point gives 2.9% and 4.8%. The 0.1-percentage-point increment has a maximum rounding error of 0.05 percentage point, while the supplied interval is 1.87 percentage points wide. The chosen precision communicates the scale without displaying unsupported-looking thousandths of a percent.
A suitable sentence is: “The estimated probability of at least one temperature excursion over the next 12 months is about 3.7%; the 90% uncertainty interval is approximately 2.9% to 4.8%.” The interval describes uncertainty in the estimated probability; it is not a prediction that the event itself will occur between 2.9% and 4.8% of the time for a particular unit.
Common mistake and correction. Writing “the risk is exactly 3.746%” confuses a calculated decimal with certainty and omits the event horizon. Write “estimated probability,” retain the 12-month context, and give the uncertainty summary. Also avoid converting the estimate to percent while accidentally leaving interval bounds in probability units.
Performance Measures and Comparisons
Performance measures such as RMSE, accuracy, or AUC are estimates calculated from particular evaluation data. Their displayed digits do not by themselves establish that small differences between models are meaningful. A report should retain the metric's unit or definition, give a suitable precision, and avoid describing a small numerical difference as a proven improvement unless the analysis supports that interpretation.
For comparisons, calculate the difference from unrounded values and round the final difference only after the calculation. If the report displays the two model results at a chosen precision, verify that the displayed comparison does not imply more certainty than the comparison analysis supports. A difference between rounded entries can disagree with the rounded difference calculated from the original values.
Worked Example: Comparing Two RMSE Results
A hypothetical team evaluates two models on the same held-out data to predict repair cost, measured in hundreds of dollars. Model A has an RMSE of 14.263, and Model B has an RMSE of 14.118. A paired resampling analysis reports a 90% uncertainty interval for the difference, defined as RMSE for A minus RMSE for B, of \([-0.42, 0.67]\). Prepare a fair, concise comparison.
Method. Compare the unrounded estimates, round the metric values to a useful precision, and interpret the supplied interval using the stated difference order. Lower RMSE indicates smaller typical prediction error on this metric, but the interval is needed to qualify the comparison.
Step 1: Calculate the difference from the original values. The difference is \(14.263-14.118=0.145\) hundreds of dollars. In dollars, this is \(0.145 \times 100=\$14.50\).
Step 2: Round the displayed metric values. To one decimal place in the stated units, Model A is 14.3 and Model B is 14.1. Their displayed difference is 0.2 hundreds of dollars, or $20. This is not the exact difference calculated above: it results from subtracting rounded display values.
Step 3: Interpret the uncertainty interval. The supplied interval for A minus B runs from \(-0.42\) to \(0.67\), so it includes zero. Its dollar-scale endpoints are \(-0.42 \times 100=-\$42\) and \(0.67 \times 100=\$67\). The interval therefore does not support wording that treats Model B as definitively better on this comparison.
A suitable report is: “The RMSEs were 14.3 and 14.1 hundreds of dollars for Models A and B, respectively. The estimated difference (A minus B) was 0.145 hundreds of dollars, with a 90% uncertainty interval from −0.42 to 0.67; this comparison does not establish a clear difference.” The estimate difference is retained to three decimals here because its value and interval are reported together, with the units made explicit.
Common mistake and correction. Reporting “Model B reduces RMSE by exactly $20” subtracts the rounded values and presents the result as exact. The unrounded point difference is $14.50, and the interval includes zero. Report the calculation and uncertainty together rather than turning a rounded display difference into a decisive claim.
Wording That Does Not Overstate Results
Numerical precision and verbal certainty can reinforce each other. “The model will produce 128 claims” sounds more certain than “the model forecasts about 128 claims.” Likewise, “the probability is 3.7%” can sound like a known constant when the value is estimated. Choose verbs and qualifiers that identify an output as an estimate, forecast, or observed result, and specify the population or period it refers to.
Use qualifiers with care. Words such as “about” and “approximately” signal rounding; they do not replace an uncertainty interval when uncertainty matters to the decision. Conversely, adding many digits to an estimate does not communicate uncertainty. A reader should not have to infer from decimal places whether the number is a point estimate, a count, a rate, or an uncertainty bound.
Name the estimate or metric, its unit, and the population, event, or period to which it applies.
Use the uncertainty information available and ask what precision the decision actually needs. Do not treat rounding as evidence about uncertainty.
Round consistently, calculate differences from unrounded values, and verify that rounding does not hide a decision-relevant threshold or imply a stronger comparison.
Replace wording that implies certainty with language that accurately describes an estimate, and include a range or qualification when it affects interpretation.
The goal is not to make every number look cautious or imprecise. It is to preserve meaningful distinctions while avoiding detail that the evidence cannot support. This discipline is aligned with the SOA Exam PA (Predictive Analytics) syllabus.
Common Reporting Traps
- Keeping all software digits. A long decimal may be computationally reproducible but is not automatically informative. Choose reporting precision based on the unit, uncertainty, and decision.
- Rounding too early. Intermediate rounding can alter a final estimate or comparison. Carry adequate precision through the calculation and round the result that will be reported.
- Subtracting rounded model results. The difference between displayed values may not equal the rounded difference calculated from original values. State which comparison is being reported.
- Mixing units or scales. A probability of 0.04 and a probability of 4% express the same value, but a report must not combine one scale for the estimate and another for its interval without explanation.
- Using certainty language for an estimate. “Will be,” “exactly,” and “proves” can make model outputs sound guaranteed. Use wording that reflects what was estimated and what uncertainty remains.
Check Your Understanding
For each situation, select a defensible presentation and check that the wording does not imply more certainty than the supplied information supports.
- A hypothetical service model forecasts 63.482 requests for a week, with a supplied 90% uncertainty interval of 57.6 to 69.3. Staffing is planned in whole requests. Give a rounded forecast and interval. (Answer: About 63 requests; approximately 58 to 69 requests for the 90% uncertainty interval.)
- A model estimates a 12-month event probability of 0.00684, with a supplied interval of 0.0049 to 0.0093. Convert the estimate and interval to percentages, rounding to two decimal places. (Answer: 0.68%, with an interval of 0.49% to 0.93%.)
- Two hypothetical models have RMSEs of 8.746 and 8.691 units. The difference is defined as the first model's RMSE minus the second's, and its supplied 90% interval is \([-0.18, 0.29]\). Calculate the unrounded point difference and state whether the interval establishes a clear difference. (Answer: The difference is 0.055 units; no, the interval includes zero.)
- A report says, “The estimated probability is 4.26%,” based on a point estimate of 0.0426, with a supplied 90% interval of 0.031 to 0.056. Rewrite the result to one decimal place in percentage terms, including the interval. (Answer: The estimated probability is about 4.3%, with a 90% uncertainty interval of approximately 3.1% to 5.6%.)
- A model produces an estimated monthly count of 248.76. The supplied uncertainty interval is 230.4 to 268.1, and the operational decision changes at 250. Give one sentence that reports the estimate while making clear that the threshold comparison is not certain from the point estimate alone. (Answer: The estimated monthly count is about 249, with an uncertainty interval of approximately 230 to 268; because the interval spans the threshold of 250, the point estimate alone does not establish which side the actual count will fall on.)