4. Sources of uncertainty

A prediction interval is only as trustworthy as the uncertainty budget behind it. When a pass-fail decision depends on whether a predicted value stays inside a specification band, the width and placement of that interval determine the decision margin. Two models can report the same point prediction yet produce very different intervals because they account for different uncertainty sources. This section separates those sources, shows how each one enters the interval, and demonstrates how to reason about their propagation with a concrete engineering example.

Two fundamentally different kinds of uncertainty

Aleatoric uncertainty (from the Latin alea, "dice") is the irreducible randomness inherent in the data-generating process. Even with a perfect model and infinite training data, repeated measurements of the same input would still scatter. Sources include sensor noise, manufacturing variation, turbulent flow, and stochastic material behavior. Aleatoric uncertainty is a property of the world, not of the model, and it cannot be reduced by collecting more data — only characterized more precisely.

Epistemic uncertainty (from the Greek episteme, "knowledge") is uncertainty due to limited knowledge: insufficient training data, model misspecification, or parameters that are poorly constrained. It is reducible in principle — more representative data, a better model structure, or additional calibration shrinks it. In regions of the input space with sparse data, epistemic uncertainty grows; near dense training data, it shrinks.

This distinction matters operationally. If a prediction interval is dominated by aleatoric uncertainty, no amount of extra data will narrow it, and the pass-fail margin must be accepted as a physical limit or the process itself must be changed. If epistemic uncertainty dominates, the correct response is to gather more data or improve the model before trusting the decision.

Model error and measurement noise

Two further sources sit alongside the aleatoric/epistemic split and are often conflated with it:

• Model error (bias) — systematic discrepancy between the model's expected output and the true value, caused by an inadequate functional form, missing features, or distribution shift. Unlike variance, bias does not average out over repeated samples; it shifts the entire interval away from the truth.
• Measurement noise — error introduced by the instrumentation and acquisition chain: sensor resolution, calibration drift, sampling jitter, and quantization. It contributes to aleatoric uncertainty in the observed targets but is distinct from the process's own randomness.

A useful mental model is the decomposition of expected prediction error into bias², variance, and irreducible noise. The prediction interval must cover all three, but each is estimated and reduced differently.

How the sources propagate into the interval

Consider a regression model that predicts a quality characteristic y from process settings x, used to decide pass/fail against a specification limit. A well-constructed prediction interval at input x combines:

PI(x) = ŷ(x) ± t · sqrt( σ²_aleatoric(x) + σ²_epistemic(x) + σ²_model_bias )

where ŷ(x) is the point prediction, σ²_aleatoric captures irreducible noise (including measurement noise folded into the target), σ²_epistemic reflects parameter and data uncertainty, and σ²_model_bias is an allowance for systematic model error. The multiplier t sets the coverage level (e.g., 1.96 for a 95% interval under normality).

Propagation is not always additive. When uncertainty enters through an input variable rather than the output, it must be pushed through the model's sensitivity. If a measured input x has noise σ_x, its contribution to output uncertainty is approximately |∂ŷ/∂x| · σ_x — the local gradient times the input noise. This first-order propagation is exact only for linear models; for nonlinear models, the gradient must be evaluated at the operating point, and near steep regions a small input noise can produce a large output interval.

Worked example: propagation in a pass-fail decision

Suppose a model predicts coating thickness y (µm) from line speed v (m/min). At v = 40 m/min the model gives ŷ = 52.0 µm with a 95% prediction interval of ±3.2 µm, so the interval spans [48.8, 55.2]. The lower specification limit is 50 µm. The point prediction passes, but the lower interval bound falls below the limit — the decision is marginal.

Decomposing the ±3.2 µm: aleatoric noise contributes ±1.8 µm (process and gauge variation), epistemic uncertainty ±1.4 µm (sparse data near this speed), and a bias allowance ±2.0 µm from a known systematic offset at high speed. These combine in quadrature to roughly ±3.1 µm, consistent with the reported interval. Because epistemic and bias terms together account for most of the width, the correct action is to collect more data near 40 m/min and recalibrate the high-speed offset — not to accept the marginal decision as final. After adding 30 runs and correcting the offset, the interval might narrow to ±2.1 µm, moving the lower bound to 49.9 µm and still leaving a thin margin that warrants a process adjustment.

Practical implications for pass-fail margins

The decomposition changes how you set a decision rule. A naive rule compares the point prediction to the specification limit. A margin-aware rule compares the interval bound to the limit, or equivalently requires the point prediction to clear the limit by at least the interval half-width. The latter is more conservative and correctly reflects the risk of a false pass.

When diagnosing a wide interval, ask which source dominates. If aleatoric noise dominates, the interval is near its physical floor; the options are to tighten the process, improve the gauge, or accept a wider margin. If epistemic uncertainty dominates, the interval will shrink with more data — prioritize sampling in the sparse region. If model bias dominates, the interval is systematically misplaced; recalibrate or re-specify the model before using it for decisions. Distinguishing these cases prevents wasted effort: collecting more data will not fix a bias problem, and recalibrating will not fix irreducible noise.

In summary, aleatoric and epistemic uncertainty answer different questions — "how noisy is the world?" versus "how much do we not yet know?" — while model bias and measurement noise add systematic and instrumental components. Each propagates into the prediction interval through a distinct mechanism: additive variance for output-side noise, gradient-weighted propagation for input-side noise, and a non-averaging offset for bias. Reading the interval as a sum of these contributions turns a single number into an actionable diagnosis for the pass-fail decision.