Must interval width be constant?

No; it can depend on the input.

Can parameter dependencies matter?

Yes, even when individual parameter bounds stay unchanged.

AI Glossary / Statistics and probability / Advanced

Interval predictor model

Definition

An interval predictor model, or IPM, maps an input to an interval of predicted outputs rather than a single number. In a common parametric construction, a computational model is evaluated over an allowed set of parameter values. The resulting output range represents modeled uncertainty. The interval alone does not specify a probability distribution or a guaranteed future coverage rate.

What it means

A point model chooses one parameter vector and returns one output at each input. An IPM retains a set of parameter vectors and propagates that set through the model. Different inputs can therefore have intervals of different widths.
The distinction between a parameter set and an output interval matters. The way parameters combine with the input, including dependencies between parameters, determines the output range.

How it works

Choose a model structure, an input domain, and a representation of the allowed parameter set. Calculate the smallest and largest model outputs available within that set for each input of interest.
For models linear in their parameters and using a rectangular parameter set, the limits can often be calculated directly. More general parameter dependencies can require optimization over a constrained set. Optimization must respect the full parameter set.
Data-driven fitting can seek tight output envelopes while enclosing all or a specified portion of observations. Claims about future reliability require additional sampling assumptions and an applicable statistical result. Training-data enclosure is a different statement from a validated probability of covering future data.

Three practical examples

An engineering team represents uncertain coefficients in a computational response model and evaluates the resulting output envelope at several operating settings.
A simulation analyst compares a rectangular coefficient set with a set that preserves dependencies between coefficients. The comparison reveals how modeling those dependencies changes prediction width.
A forecasting team evaluates how often new observations fall outside a fitted interval model. It records those misses alongside interval widths rather than reporting width alone.

Think of it this way

Instead of selecting one route through a landscape, keep a defined family of routes. At each location, the predicted interval spans the outputs reachable by that family.

Worked example

Consider the invented model y = a + bx. First let a range from 1 to 2 and b range from 2 to 3, with every combination in this rectangle allowed. No probability is assigned to these combinations.
At x = 0, the output is simply a, so the predicted interval is [1, 2]. Its width is one unit.
At x = 2, both coefficients contribute positively. The smallest output is 1 plus 2 times 2 = 5. The largest is 2 plus 3 times 2 = 8. The interval is [5, 8], with width three.
At x = minus 2, the slope contribution reverses direction. The minimum is 1 minus 2 times 3 = minus 5, and the maximum is 2 minus 2 times 2 = minus 2. The interval is [minus 5, minus 2]. Substituting both lower coefficient bounds together would give minus 3, which is not the lower output limit.
The rectangle's center is a = 1.5 and b = 2.5, and each coefficient has half-width 0.5. Thus the output center is 1.5 plus 2.5x and its half-width is 0.5 plus 0.5 times the absolute value of x. At x = minus 2 this gives center minus 3.5 and half-width 1.5, reproducing the same limits.
Now change the parameter set. Keep a between 1 and 2, but require b = 4 minus a. The individual range of b is still [2, 3]. However, the pair a = 2, b = 3 is no longer allowed because its sum is five rather than four.
At x = 1, every allowed pair gives y = a + b = 4. The output interval collapses to [4, 4]. If the dependency were discarded and the rectangular ranges used instead, the interval would be [3, 5]. The same individual coefficient bounds can therefore produce substantially different output envelopes.
At x = 2, the constrained model gives y = a plus 2 times (4 minus a) = 8 minus a. Its interval is [6, 7], narrower than the rectangle's [5, 8].
Finally, suppose two recorded outputs at x = 0 are 1 and 2. The rectangular IPM encloses both because a can take either value. That does not mean one fixed parameter vector produces both observations at the same input. In this envelope interpretation, different observations may be accommodated by different members of the parameter set.
These are exact set-propagation calculations for stipulated sets. They are not fitted reliability bounds, confidence intervals, or evidence that every future outcome will remain inside the envelope.

Illustrative case study

Fictional company Cedar Systems builds an interval model for a simulated operating response. Its first version uses separate ranges for two coefficients. Engineers then identify a relationship linking those coefficients and compare the corresponding constrained envelope.
The revised interval is narrower at some settings, but the team does not label that narrowing an accuracy improvement without further checks. It tests additional settings and documents the evidence supporting the coefficient relationship.

Common misconceptions and mistakes

A parameter interval does not automatically make all values inside it equally probable.
Enclosing every training observation does not guarantee enclosure of every future observation.
Separate parameter bounds can lose important dependencies. Combining them as a rectangle may admit combinations excluded by the intended model.
A zero-width interval at one input can coexist with parameter uncertainty, as the constrained example shows.
Narrowing a set without justification can make predictions less reliable even though the resulting intervals look more precise.

Business use

Use an IPM when a decision benefits from a transparent output envelope and a defensible parameter-set model. Report interval widths, validation outcomes, the supported input domain, and the assumptions behind any reliability claim. Document why parameter sets change.

Check your understanding

Answer: What is the rectangular model's interval at x = 2? [5, 8].

Why does x = minus 2 require care?

Answer: Multiplying by a negative value reverses which slope bound minimizes the output.

What happens at x = 1 when a + b = 4?

Answer: Every allowed output equals four.

Related terms

Parameter uncertainty
set-valued prediction
interval analysis
uncertainty propagation
model calibration
scenario optimization
prediction coverage

Learn next

Study constrained parameter sets, interval-envelope optimization, and the assumptions required for finite-sample reliability bounds.

Sources

Crespo, Giesy and Kenny, Interval Predictor Models with a Formal Characterization of Uncertainty and Reliability: https://ntrs.nasa.gov/api/citations/20150002792/downloads/20150002792.pdf . Crespo and colleagues, Interval Predictor Models for Robust System Identification: https://ntrs.nasa.gov/citations/20210010650 . Numerical sets and company scenario are original illustrations.

Previous: Interval estimate Next: Interval scale