Both are estimates made from the same fitted curve. Only one of them is backed by direct evidence on both sides.
Every model, curve fit, or trend line eventually gets used to estimate a value you didn't directly measure. Interpolation estimates a value that falls within the range of data you actually measured — bracketed by real observations on both sides. Extrapolation estimates a value that falls outsidethat range — extending the same fitted relationship into territory you have never observed at all. The math used to compute either one can look identical: plug an x-value into the fitted equation and read off y. But the confidence you're entitled to have in the answer is not the same, and treating them as equally trustworthy is a routine, well-documented source of engineering and scientific error.
Say you've tested a material, a process, or a trend at several conditions and fitted a curve through the results. Any x-value that falls between your lowest and highest tested x-values is an interpolation — you have direct measurements bracketing it on both sides, so the fitted curve is being asked to do something modest: describe the space between two things you already know, assuming the relationship behaves reasonably smoothly in between. Any x-value beyondthe range you tested is an extrapolation — the fitted curve is now being asked to describe territory where you have zero direct evidence, resting entirely on the assumption that whatever pattern held inside your tested range keeps holding outside it. That assumption is sometimes fine. It is also exactly the kind of assumption that silently fails: a material's stress-strain response can be beautifully linear across the tested load range and then yield or fracture entirely differently just beyond it; a growth trend can look linear for years and then plateau or reverse the moment it hits a real-world constraint that was never present in the historical data.
Engineers and scientists constantly need estimates beyond directly tested conditions — a material at a temperature never actually tested, a structure's response at a load level beyond what was directly measured, a trend extended a few years into the future. That's unavoidable and often necessary. What has to change is the confidence attached to the number. Inside the measured range, the fitted curve is doing something modest: filling a gap between two confirmed points, which is a much smaller leap than it sounds when the underlying relationship is reasonably well-behaved. Outside the measured range, the same equation is being asked to make a claim about territory where the relationship has never once been checked — and real physical systems routinely change behavior right at the edges of where they've been tested, precisely because that's where new physics, new failure modes, or new real-world constraints tend to show up. A good fit inside the data proves the model describes that data. It proves nothing about what lies beyond it.
False — a good fit within the observed data says nothing about whether the same relationship continues to hold beyond it. Real physical and real-world systems very often behave differently once outside a tested range: a material that responds linearly across its tested load range can yield or fracture entirely differently just past it; a chemical reaction that follows one rate law at moderate temperatures can switch mechanism entirely at a temperature never tested; a population or demand trend that looks linear for years can plateau or reverse once it runs into a real-world constraint that simply wasn't present in the historical data used to fit the curve. An excellent R² inside the tested range says nothing about the region outside it. Extrapolated estimates deserve substantially more caution — additional safety margin, independent verification, or a genuinely different physical model — than interpolated estimates drawn from data that actually brackets them.
Explains the difference between interpolation (estimating a value within the range of data you've actually measured, bracketed by direct evidence on both sides) and extrapolation (estimating a value outside that range, by extending a fitted trend into territory with no direct evidence) — and why the two deserve very different levels of confidence even when computed from the exact same fitted curve.
Interpolation estimates a value that falls between two data points you actually measured — the estimate is bracketed by real, direct evidence on both sides, so the fitted relationship only has to hold across a narrow, already-confirmed gap. Extrapolation estimates a value that falls outside the range of data you measured at all — beyond your highest or below your lowest tested x-value — resting entirely on the assumption that the pattern observed inside the data keeps holding once you leave it. The arithmetic used to compute either value from a fitted equation is identical; the evidentiary basis behind them is not.
A curve fit is only ever validated against the data used to build it. Inside that range, deviations from the fitted line would already have shown up as poor fit. Outside that range, there is no such check — the model could diverge from reality the moment conditions change, and nothing in the fit statistics computed from the observed data would warn you. This is exactly why a material's stress-strain curve can be reliably linear across a tested load range and then yield or fail in a completely different mode just beyond it, or why a population, demand, or growth trend that looks linear for years can plateau or reverse the instant it meets a real-world constraint that was never present while the historical data was collected.
Materials and structural engineering: predicting behavior at a temperature, pressure, or load never directly tested requires either a validated physical model (not just a curve fit) or an explicit safety margin sized for the uncertainty of extrapolating. Forecasting and trend analysis: projecting a trend a few years beyond the historical data it was fitted to is common and often necessary, but the further beyond the observed range the projection reaches, the less the historical fit statistics say about its reliability. Calibration and instrumentation: a sensor or calibration curve is only trustworthy across the range it was actually calibrated over — readings taken outside that calibrated range are extrapolations and should be flagged as such, not reported with the same confidence as in-range readings.
No — it is often necessary and can be perfectly reasonable, especially just past the edge of the tested range or when there is a strong physical reason to expect the same relationship to continue. The point isn't that extrapolation should never be done; it's that it should be flagged, treated with more caution, and backed with additional safety margin or independent justification rather than reported with the same confidence as an interpolated value.
There's no fixed rule, but risk generally grows the further you go — and grows faster near known physical limits (yield points, saturation, phase changes, resource constraints) where real systems are known to change behavior. A rough rule of thumb used in engineering practice is to treat extrapolation much more conservatively even a small distance past the tested range, and to require independent verification (physical modeling, targeted additional testing) before extrapolating any significant distance.
No. R² and other goodness-of-fit statistics describe how well the curve matches the data it was fitted to — they say nothing about the region outside that data. A near-perfect fit inside the tested range is fully compatible with the true relationship diverging sharply just beyond it; the fit statistics simply have no information about that region at all.
Yes, though less commonly. If the true relationship is not smooth between two measured points — for example if there's a sharp transition, discontinuity, or missed feature hiding in a large gap between data points — a simple interpolation can still be wrong. This is why interpolation is described as lower risk, not risk-free: it depends on the relationship behaving reasonably between the bracketing points, and widely spaced data points weaken that assumption.
Flag it explicitly as an extrapolation rather than presenting it identically to interpolated values, apply a larger safety margin or uncertainty band, and where possible validate it against an independent source — additional targeted testing, a first-principles physical model, or known limiting behavior — rather than relying solely on extending the same empirical curve fit.
Try our STEM Learning Studio
More calculators, simulators, and guides for this discipline.