Communication of Model Uncertainty | Malone Math

The Central Idea

Suppose a model produces an uncertain quantity of interest

\[ Y=g(X,\theta,M), \]

where $X$ represents uncertain inputs, $\theta$ represents uncertain parameters, and $M$ represents modeling assumptions or model structure. A complete communication does not report only one estimate $\hat Y$. It also explains the uncertainty surrounding that estimate and the conditions under which the conclusion should be trusted.

In practical terms, uncertainty communication should answer four questions:

What is the best current estimate?
How uncertain is it?
Why is it uncertain?
Does that uncertainty change the decision?

Point Estimates Are Not Enough

A point estimate such as

\[ \hat Y=42 \]

may look precise, but it does not show whether plausible values are tightly concentrated near $42$ or spread across a much wider range.

Compare the statements

\[ \hat Y=42 \]

and

\[ Y\approx42,\qquad 95\%\text{ interval }[39,45]. \]

The second statement gives substantially more decision-relevant information.

Uncertainty Is Not the Same as Error

Error is the difference between a reported value and a true or reference value. Uncertainty describes the range, distribution, or degree of doubt surrounding a quantity when the true value is not known exactly.

A forecast can have substantial uncertainty even before its eventual error is observable. Likewise, a model can communicate uncertainty honestly even when the final realized outcome later falls close to the point estimate.

Name the Source of the Uncertainty

A useful uncertainty statement identifies what produces the uncertainty rather than reporting a range with no explanation.

Measurement uncertainty: inputs are observed imperfectly.
Parameter uncertainty: model coefficients are estimated rather than known.
Process variability: the system itself contains randomness.
Sampling uncertainty: only a finite sample was observed.
Structural uncertainty: the mathematical form of the model may be imperfect.
Scenario uncertainty: future conditions may follow different plausible paths.
Numerical uncertainty: approximation and computation introduce error.

Different sources call for different summaries and different management responses.

Aleatory and Epistemic Uncertainty

A common distinction separates uncertainty into two broad forms.

Aleatory uncertainty represents inherent variability or randomness in the system. Epistemic uncertainty represents lack of knowledge that may be reduced through better data, measurement, or modeling.

Communicating this distinction matters because it tells the audience whether uncertainty can potentially be reduced or whether it is an irreducible feature of the process being modeled.

Absolute and Relative Uncertainty

An uncertainty amount can be reported in the same units as the estimate. If

\[ x=50\pm2, \]

then the absolute uncertainty is $2$.

Relative uncertainty compares the uncertainty with the magnitude of the quantity:

\[ \frac{2}{50}=0.04=4\%. \]

Relative uncertainty is often easier to compare across quantities with different scales.

Communicating with Intervals

An interval summarizes a range of values supported by the modeling procedure. Examples include confidence intervals, credible intervals, prediction intervals, tolerance intervals, and engineering uncertainty bounds.

The interval should always be labeled by type. Writing only

\[ [18,26] \]

is incomplete because different intervals have different meanings.

Communicating a Confidence Interval

A frequentist confidence interval is tied to the long-run behavior of the interval-producing procedure. If a valid 95% confidence procedure were repeated across many independent samples, approximately 95% of the resulting intervals would contain the true parameter under the model assumptions.

For communication, it is usually clearest to report the estimate, interval, confidence level, and assumptions together rather than giving the interval alone.

Communicating a Bayesian Credible Interval

A Bayesian credible interval is interpreted through the posterior distribution. For example, a 95% posterior credible interval $[a,b]$ may be described as a region containing 95% of the posterior probability for the parameter, conditional on the model, prior, and observed data.

Confidence and credible intervals may look similar numerically, but their interpretations are not interchangeable.

Communicating a Prediction Interval

A prediction interval describes uncertainty about a future observation rather than uncertainty about a population parameter or mean response.

For example, a model might report

\[ \hat y=120, \qquad \text{95% prediction interval }[90,150]. \]

The interval is wider than a corresponding interval for the mean when it must also include individual outcome variability.

Communicating Probabilities

Probability is most useful when attached to a clearly defined event. Instead of saying that an outcome is "unlikely," a model may estimate

\[ P(Y>y_c)=0.08. \]

This states that the model assigns an 8% probability to exceeding the critical threshold $y_c$, subject to the assumptions used to construct the probability model.

Natural Frequencies

Some audiences understand probabilities more easily when they are translated into natural frequencies. A probability of

\[ p=0.02 \]

can be expressed as approximately 2 cases out of 100 under comparable conditions.

Natural frequencies can improve intuition, especially when probabilities are small, but the denominator and reference population must remain explicit.

Probability and Odds

Probability and odds are related but are not the same quantity. If an event has probability $p$, its odds in favor are

\[ \frac{p}{1-p}. \]

For $p=0.75$, the odds are

\[ \frac{0.75}{0.25}=3, \]

or 3 to 1 in favor. Reports should avoid switching casually between odds and probability because the numerical values differ.

Communicating Scenarios

Scenarios describe internally consistent alternative conditions such as a baseline case, downside case, upside case, or stress case. Unless probabilities have been explicitly assigned and justified, scenarios should not be presented as though they are probabilistic forecasts.

A scenario table should normally identify the assumptions that change, the resulting model outputs, and the operational or strategic implication of each case.

A Range Is Not Automatically a Probability Statement

Suppose a stress analysis produces outputs from $70$ to $130$. The statement

\[ 70\le Y\le130 \]

does not imply that all values in the range are equally likely, nor does it imply a particular probability that the true outcome lies inside the range.

The communication should state whether the range comes from probabilistic simulation, deterministic scenario bounds, expert judgment, sensitivity analysis, or another procedure.

Report Center and Spread Together

A probability distribution is rarely described adequately by a measure of center alone. A useful summary often pairs a mean or median with a measure of spread.

\[ \mu=E[Y], \qquad \sigma=\sqrt{\operatorname{Var}(Y)}. \]

For skewed or heavy-tailed distributions, the median and quantiles may communicate the uncertainty more faithfully than the mean and standard deviation alone.

Quantiles and Percentiles

The $q$-quantile $y_q$ satisfies, approximately,

\[ P(Y\le y_q)=q. \]

Quantiles are especially useful for uncertainty communication because they do not require a symmetric distribution. Reporting the 5th, 50th, and 95th percentiles can show downside, central, and upside outcomes in one compact summary.

Do Not Hide Skewness

A symmetric summary such as

\[ \hat y\pm\delta \]

can be misleading when the uncertainty distribution is strongly asymmetric. For example, a cost model may have limited downside but a long upper tail.

In such cases, asymmetric intervals or separate quantiles communicate the shape more accurately.

Communicating Tail Risk

Some decisions depend more on rare extreme outcomes than on average behavior. Useful tail summaries include exceedance probabilities,

\[ P(Y>y_c), \]

high quantiles such as $y_{0.99}$, and conditional tail measures such as expected shortfall.

Tail estimates should be accompanied by a warning when the available data contain little information about extreme events.

Communicating Risk as Expected Loss

When outcomes have different consequences, probability alone may not be enough. If outcome $i$ has probability $p_i$ and loss $L_i$, expected loss is

\[ E[L]=\sum_i p_iL_i. \]

A low-probability event can still be decision-critical when its consequence is large. Good communication therefore separates probability from consequence before combining them into a risk measure.

Connect Uncertainty to Decision Thresholds

Uncertainty becomes operationally important when it crosses a decision threshold. Suppose a project is acceptable only if

\[ Y\ge100. \]

A forecast of $\hat Y=110$ may sound favorable, but if

\[ P(Y<100)=0.35, \]

the decision is materially exposed to uncertainty. Reporting this threshold probability is often more useful than reporting the point forecast alone.

Does Uncertainty Reverse the Decision?

One of the most important communication questions is whether uncertainty can reverse a recommendation.

If alternative $A$ is preferred across nearly all plausible assumptions, the decision is robust. If small changes in assumptions make alternative $B$ preferable, the decision is fragile.

This distinction should be stated directly because decision robustness is often more important than the exact width of an uncertainty interval.

Report What Drives the Uncertainty

Sensitivity analysis identifies which uncertain inputs contribute most strongly to variation in the output. A useful uncertainty report should highlight the dominant drivers rather than presenting all uncertain inputs as equally important.

For a differentiable model $Y=f(x_1,\ldots,x_n)$, local sensitivity to input $x_i$ is represented by

\[ \frac{\partial Y}{\partial x_i}. \]

Global methods can extend this idea across a wider parameter region and include interactions among inputs.

Uncertainty Budgets

An uncertainty budget allocates portions of total uncertainty to identifiable sources. In a simple variance decomposition,

\[ \operatorname{Var}(Y)\approx V_1+V_2+\cdots+V_k, \]

where each component represents a source or group of sources.

The purpose is not merely accounting. It helps determine where additional data, improved measurement, or better modeling would produce the greatest reduction in uncertainty.

State the Assumptions That Matter

Every uncertainty statement is conditional on assumptions. Effective reporting distinguishes assumptions that are routine from assumptions that materially drive the conclusion.

Important assumptions may concern distribution shape, independence, stationarity, linearity, parameter ranges, future policy, operating conditions, missing data, or the structure of the underlying model.

The objective is not to list every technical detail in the headline result. It is to disclose the assumptions whose failure could change the interpretation or decision.

Use Conditional Language

Model-based uncertainty is conditional. A statement such as "the probability is 12%" is stronger than the analysis usually supports unless the modeling assumptions are understood.

More precise language is:

Under the stated model, data, and assumptions, the estimated probability is 12%.

This does not weaken the result. It describes the scope of the claim accurately.

Avoid False Precision

Reporting more decimal places than the model can support creates a false impression of certainty. If the uncertainty is approximately $\pm8$, a forecast such as

\[ 103.7426 \]

is usually less informative than

\[ 104\pm8. \]

Numerical precision should be consistent with the quality of the data and the scale of the uncertainty.

Match Significant Digits to Uncertainty

If an uncertainty is reported as $\pm0.3$, reporting the central estimate as $12.34782$ suggests accuracy that the uncertainty does not justify.

A more coherent presentation is

\[ 12.3\pm0.3. \]

The exact rounding convention depends on the field, but the central principle is universal: displayed precision should not exceed meaningful precision.

Use Verbal Probability Labels Carefully

Words such as "likely," "unlikely," "possible," and "almost certain" can mean different things to different readers. If verbal labels are used, they should be tied to explicit numerical ranges whenever practical.

For example, an organization might define "likely" as a probability between 70% and 90%. The mapping should be stated so that the language has a stable quantitative meaning.

Visualizing Uncertainty

Visual displays can make uncertainty easier to understand when they reveal both the central result and the range of plausible outcomes. Common displays include:

error bars;
confidence or credible bands;
prediction bands;
histograms and density plots;
box plots;
fan charts;
scenario tables;
tornado diagrams;
probability-of-exceedance curves.

The visual form should match the question being asked rather than merely display all available output.

Error Bars Need Labels

Error bars are ambiguous unless the report states what they represent. They may show standard deviations, standard errors, confidence intervals, credible intervals, prediction intervals, or other uncertainty measures.

Two charts with visually identical error bars can therefore have very different statistical meanings.

Fan Charts

A fan chart displays how predictive uncertainty changes over time. Narrower bands close to the present and wider bands farther into the future can communicate the accumulation of uncertainty in forecasting models.

Different shaded regions may represent central probability intervals such as 50%, 80%, and 95%. Each band should be identified explicitly.

Tornado Diagrams

A tornado diagram summarizes how much a model output changes as individual inputs move through specified ranges. It is useful for communicating sensitivity and prioritizing uncertainty drivers.

The diagram should not be interpreted automatically as a probability ranking. A wide bar shows a strong output response over the chosen input range, not necessarily that the input is the most uncertain or most likely to vary.

Show the Distribution When Shape Matters

If the uncertainty distribution is multimodal, skewed, or heavy-tailed, a single interval can hide important structure. Displaying the simulated or posterior distribution can reveal features that summary statistics suppress.

This is especially important when different modes correspond to different mechanisms, regimes, or strategic scenarios.

Calibration

An uncertainty statement should be evaluated not only for sharpness but also for calibration. Roughly speaking, a procedure is well calibrated when events assigned a probability near $p$ occur approximately a fraction $p$ of the time across comparable cases.

For interval forecasts, a nominal 90% interval should contain the realized outcome about 90% of the time under the relevant repeated-use conditions.

Calibration and Sharpness

Extremely wide intervals can achieve high coverage while being operationally unhelpful. Good probabilistic forecasts therefore seek both calibration and sharpness.

Calibration asks whether stated uncertainty is statistically reliable. Sharpness asks whether the uncertainty ranges are as concentrated as reasonably possible while retaining that reliability.

Report Validation Evidence

Uncertainty claims are more credible when they are accompanied by evidence that the model has been tested against relevant data or conditions.

Depending on the model, useful validation evidence can include held-out prediction error, backtesting, residual analysis, calibration checks, stress tests, sensitivity analysis, or comparison with alternative model structures.

State Important Limitations

A limitation is not merely a disclaimer. It identifies a condition under which the model's uncertainty estimates may fail or become less reliable.

Examples include extrapolation beyond observed data, sparse observations in the tails, omitted variables, unresolved dependence, regime changes, poor measurement, and uncertain model structure.

Quantified and Unquantified Uncertainty

Not all uncertainty can be reduced to a probability distribution. A model may quantify measurement error and parameter uncertainty while leaving structural uncertainty only partially characterized.

Reports should distinguish uncertainty that has been explicitly quantified from uncertainty that remains outside the formal analysis.

A narrow numerical interval is not evidence that unmodeled uncertainty is small.

Communicate Model Risk

Model risk is the possibility that conclusions are wrong because the model itself is misspecified, miscalibrated, used outside its intended domain, or implemented incorrectly.

Model risk is often different from the stochastic uncertainty generated inside the model. A simulation may be extremely precise about the behavior of an incorrectly specified model.

Flag Extrapolation Clearly

Predictions made outside the range of observed data usually rely more heavily on model structure than predictions made within the data-supported region.

If $x_0$ lies far beyond the observed input range, an interval produced by the fitted model may not fully capture the risk that the assumed functional form is wrong there.

Extrapolated results should therefore be identified explicitly rather than being visually or verbally blended with interpolated results.

Communicate When Estimates Will Be Updated

Some model results are snapshots rather than permanent conclusions. Forecasts, parameter estimates, and risk measures may change as new information arrives.

A useful report can state what future evidence would trigger an update, such as a new observation period, a threshold crossing, a parameter drift signal, or a change in operating conditions.

Match the Communication to the Audience

The mathematical analysis may be identical while the presentation changes by audience.

A technical audience may need model equations, diagnostics, and distributional assumptions.
An operating team may need thresholds, ranges, triggers, and contingency actions.
An executive audience may need the central forecast, downside exposure, confidence level, and decision implication.
A public audience may need plain-language explanation of what the uncertainty does and does not mean.

Simplifying the presentation should not change the meaning of the uncertainty.

Layer the Information

A practical reporting structure places the most decision-relevant information first and technical detail later.

State the decision-relevant result.
Give the uncertainty range or probability.
Identify the main uncertainty drivers.
State the assumptions that could change the conclusion.
Provide technical methods and diagnostics for readers who need them.

This layered approach preserves rigor without forcing every reader through the entire technical derivation before understanding the conclusion.

Baseline, Downside, and Upside Summaries

For many planning models, a compact three-case summary is useful:

Baseline: the central planning case;
Downside: a plausible adverse case;
Upside: a plausible favorable case.

These cases should be tied to explicit assumptions. They should not be labeled "90% case" or "10% case" unless those probabilities or quantiles have actually been estimated.

Separate Stress Cases from Expected Cases

A stress case is designed to test resilience under adverse conditions. It does not need to be the most probable outcome.

Mixing a stress case into the same presentation as a probabilistic forecast without labeling the distinction can cause readers to interpret severity as likelihood.

Relative and Absolute Risk

Relative changes can sound dramatic when the baseline probability is small. Suppose risk increases from 1% to 2%.

The relative increase is

\[ \frac{0.02-0.01}{0.01}=1=100\%, \]

while the absolute increase is only

\[ 0.02-0.01=0.01=1\text{ percentage point}. \]

Reporting both measures often gives a more balanced interpretation.

Percent and Percentage Points

If a probability moves from 20% to 25%, the increase is 5 percentage points but the relative increase is

\[ \frac{25-20}{20}=0.25=25\%. \]

These descriptions are both correct but communicate different comparisons. Reports should distinguish them explicitly.

Communicate Inconclusive Results Clearly

Sometimes uncertainty is wide enough that the model does not support a clear conclusion. That is a substantive result.

If two alternatives have strongly overlapping performance distributions and the ordering changes across plausible assumptions, the appropriate communication may be that the available evidence does not support a robust preference.

A forced ranking can communicate more certainty than the model actually contains.

Explain What Would Reduce Uncertainty

A useful uncertainty report can identify the next information that would be most valuable. If parameter $\theta_2$ dominates output uncertainty, improving the estimate of $\theta_2$ may have more value than collecting additional data on already well-known inputs.

This connects uncertainty quantification to measurement strategy, research design, and resource allocation.

Value of Information

Additional information is valuable when it can change a decision or reduce expected loss. Conceptually, the value of information compares the expected performance of decisions made with current information to decisions made after obtaining better information.

If no plausible data update would alter the decision, reducing uncertainty may have little operational value even if it improves the model scientifically.

Common Communication Failure Modes

reporting only the point estimate;
showing intervals without identifying their meaning;
using too many significant digits;
treating scenarios as probabilities;
hiding assumptions that materially drive the result;
reporting precision while ignoring structural model risk;
using vague probability words without quantitative definitions;
presenting relative risk without absolute risk;
failing to distinguish uncertainty from known error;
omitting the decision threshold that makes the uncertainty relevant.

Uncertainty Does Not Mean Ignorance

A quantified uncertainty range does not mean the model has no information. In fact, identifying a constrained set of plausible outcomes can be much more useful than claiming an unjustified exact answer.

The purpose of uncertainty analysis is to replace false certainty with structured knowledge about what remains possible.

Uncertainty Does Not Make All Outcomes Equally Plausible

A wide range of possible outcomes does not imply that every value in the range is equally supported. Probability distributions, likelihoods, posterior densities, and scenario plausibility may differ greatly across the range.

Communication should preserve this structure whenever it is decision-relevant.

Probabilities Are Conditional on the Model

A model-generated probability is not detached from the assumptions used to produce it. In symbolic form,

\[ P(A\mid M,D), \]

emphasizes that the probability of event $A$ is being evaluated conditional on model $M$ and information $D$.

If the model or data change, the reported probability can change as well.

Communicating Across Multiple Models

When several plausible models produce materially different outputs, reporting only one preferred model can understate structural uncertainty.

Analysts may report the range across models, model-specific estimates, weighted ensembles, or the effect of model choice on the final decision. The key is to distinguish variation within a model from variation across models.

Include a Robustness Statement

A concise robustness statement summarizes whether the conclusion survives reasonable changes in assumptions.

For example:

The preferred option remains preferred across the tested parameter ranges, alternative model structures, and downside scenarios; the recommendation changes only when demand falls below the identified break-even threshold.

This communicates more about decision stability than a point estimate alone.

Use Technical Appendices Strategically

Complex uncertainty analyses may require distributional assumptions, simulation settings, convergence diagnostics, priors, covariance structures, sensitivity methods, and validation procedures.

These details should remain available for audit and replication, but they do not all need to appear in the headline communication. A technical appendix allows rigor and accessibility to coexist.

Reproducibility Supports Credibility

When uncertainty is produced by simulation or computational analysis, the report should preserve enough information to reproduce the result. Depending on the method, this may include data versions, code versions, random seeds, parameter distributions, scenario definitions, and the number of simulation runs.

Reproducibility does not guarantee correctness, but it makes the uncertainty analysis inspectable and testable.

A Reliable Communication Workflow

Identify the decision or question the model supports.
State the central estimate or principal conclusion.
Choose the uncertainty measure that matches the question.
Label the uncertainty measure correctly.
Identify the dominant sources of uncertainty.
State assumptions that could materially change the result.
Connect uncertainty to thresholds, consequences, or decision reversals.
Show robustness or fragility under plausible alternatives.
Use visuals that preserve the statistical meaning.
State important limitations and unquantified uncertainty.
Explain what new information would reduce uncertainty.
Preserve technical details for audit and replication.

Example: Communicating a Forecast

Suppose a model forecasts annual demand of

\[ \hat D=24{,}000 \]

units, with a 90% prediction interval of

\[ [19{,}000,31{,}000]. \]

Capacity becomes strained above $28{,}000$ units, and simulation estimates

\[ P(D>28{,}000)=0.18. \]

A decision-oriented summary is therefore not merely "forecast demand is 24,000." It is that central demand is about 24,000 units, substantial uncertainty remains, and the model assigns roughly an 18% probability to demand exceeding the current capacity threshold under the modeled conditions.

Example: Communicating an Investment Model

Suppose a project has modeled net present value

\[ E[\mathrm{NPV}]=\$12\text{ million} \]

but also

\[ P(\mathrm{NPV}<0)=0.22. \]

Reporting only the positive expected NPV hides an important feature of the decision. A stronger communication states both the expected value and the modeled probability of loss, together with the assumptions that dominate the downside cases.

Example: Structural Uncertainty

Suppose three plausible model structures estimate the same outcome as

\[ 84,\qquad101,\qquad117. \]

If each model individually reports a narrow interval, presenting only the preferred model can create a misleading appearance of precision. The dispersion across model structures is itself decision-relevant information and should be disclosed.

Practice 1: Absolute and Relative Uncertainty

A measurement is reported as

\[ 80\pm4. \]

Find the relative uncertainty as a percentage.

Practice 2: Probability as a Natural Frequency

A model assigns probability $0.03$ to a failure event. Express this as an approximate natural frequency per 1,000 comparable cases.

Practice 3: Percent or Percentage Points?

A modeled probability rises from 8% to 12%. Find both the increase in percentage points and the relative percentage increase.

Practice 4: Expected Loss

A failure has probability $0.04$ and would create a loss of $\$500{,}000$. Find the expected loss contribution from this event.

Practice 5: Threshold Probability

A production model has mean output $105$, but the minimum acceptable output is $100$. Which quantity is more directly relevant to the decision than the mean alone?

Practice 6: Identify the Interval

A report says only that the uncertainty range is $[45,60]$. Explain why this is not enough information for correct interpretation.

Practice 7: Scenario or Probability?

An analyst constructs a downside case with demand 20% below baseline but has not assigned probabilities to future demand. Is the downside case a 20th-percentile forecast? Explain.

Practice 8: False Precision

A model reports $37.42819$ with uncertainty approximately $\pm2.1$. Suggest a more appropriate numerical presentation.

Practice 9: Decision Fragility

Model A recommends option 1, while two equally plausible structural variants recommend option 2. What uncertainty issue should be communicated?

Practice 10: Relative and Absolute Risk

A modeled adverse-event probability increases from 0.5% to 1.0%. Describe the change using both relative and absolute terms.

Practice Answers

Practice 1

\[ \frac{4}{80}=0.05=\boxed{5\%}. \]

Practice 2

\[ 0.03\times1000=\boxed{30}. \]

The model corresponds to approximately 30 failures per 1,000 comparable cases, conditional on the modeled conditions.

Practice 3

The absolute change is

\[ 12\%-8\%=\boxed{4\text{ percentage points}}. \]

The relative increase is

\[ \frac{12-8}{8}=0.5=\boxed{50\%}. \]

Practice 4

\[ E[L]=0.04(500{,}000)=\boxed{\$20{,}000}. \]

Practice 5

The more decision-relevant quantity is the modeled probability of falling below the threshold:

\[ \boxed{P(Y<100)}. \]

Practice 6

The range is ambiguous because it does not identify whether it is a confidence interval, credible interval, prediction interval, scenario range, tolerance interval, or another uncertainty measure. The procedure and interpretation must be labeled.

Practice 7

No. The downside case is a constructed scenario. Without a probability model or empirical quantile calculation, it cannot be interpreted as the 20th percentile or any other probabilistic statement.

Practice 8

A presentation such as

\[ \boxed{37.4\pm2.1} \]

is more consistent with the scale of the stated uncertainty than reporting five decimal places.

Practice 9

The recommendation is structurally fragile. The report should communicate structural model uncertainty and state that the preferred action changes across plausible model forms.

Practice 10

The probability rises by

\[ 1.0\%-0.5\%=\boxed{0.5\text{ percentage point}}. \]

Relative to the original probability, it has doubled:

\[ \frac{1.0-0.5}{0.5}=1=\boxed{100\%\text{ increase}}. \]

What to Remember

A point estimate should rarely be reported without information about uncertainty.
Uncertainty describes lack of exact knowledge; error is a realized difference from truth or reference.
Name the source of uncertainty whenever it affects interpretation.
Distinguish aleatory variability from epistemic lack of knowledge when useful.
Label confidence, credible, prediction, tolerance, and scenario intervals correctly.
A range does not automatically imply a probability distribution.
Use probabilities for clearly defined events and thresholds.
Natural frequencies can make small probabilities easier to interpret.
Report center and spread together.
Use quantiles or asymmetric intervals when the distribution is skewed.
Tail risk may matter more than the mean for high-consequence decisions.
Separate probability from consequence before combining them into a risk measure.
Connect uncertainty to the threshold that drives the actual decision.
State whether uncertainty can reverse the recommendation.
Report the inputs and assumptions that dominate uncertainty.
Avoid false precision and unnecessary decimal places.
Do not use verbal probability labels without a shared quantitative meaning.
Error bars and shaded bands must state what they represent.
Calibration and sharpness both matter for probabilistic forecasts.
Disclose important model limitations, extrapolation, and unquantified uncertainty.
Structural model risk can remain even when numerical uncertainty inside one model is small.
Relative and absolute risk should often be reported together.
An inconclusive result is preferable to an unjustified precise conclusion.
Explain what information would most efficiently reduce uncertainty.
Layer communication so decision-relevant conclusions come first and technical detail remains available.

The essential principle is: communicate uncertainty in the form that preserves its mathematical meaning while making its decision consequence visible. A useful model report should tell the reader not only what the model predicts, but how strongly the prediction is supported, what could move it, and whether those movements matter.

Section Completion

This lesson completes 5.17 Uncertainty Within Models. The section began with measurement and parameter uncertainty, developed probability-based and simulation-based treatments, examined reliability, scenarios, sensitivity, Bayesian updating, and robustness, and concludes here with the communication of uncertainty to the people who must interpret and act on model results.

The unifying idea is that uncertainty is not an afterthought added to a finished model. It is part of the model itself: it must be represented, propagated, tested, interpreted, and communicated alongside the central prediction or decision.