Extrapolation is unreliable because you’re making predictions in a region where you have no data to confirm whether your assumptions still hold. Within the range of your observations, you can check your model against reality. Outside that range, you’re trusting that the patterns you’ve seen will continue unchanged, and that trust is frequently misplaced. The further you project beyond your data, the wider your margin of error grows, and the more likely you are to encounter conditions your model never accounted for.
Uncertainty Grows With Distance From Your Data
The most fundamental problem with extrapolation is mathematical. When you fit a model to data, your predictions are most reliable near the center of your observations and become less reliable as you move toward the edges. Once you move beyond the edges entirely, uncertainty expands rapidly.
This happens because the margin of error in any prediction depends partly on how far your new data point sits from the average of your existing data. As that distance increases, the confidence interval around your prediction widens. Near the middle of your data, you have plenty of neighboring observations keeping your estimate anchored. At the extremes, you’re leaning on fewer and fewer data points, and any small error in your model’s slope or curvature gets magnified over the longer distance. A prediction that’s off by a small amount at the boundary of your data could be off by an enormous amount two or three times beyond it.
Real Systems Rarely Stay Linear
Most extrapolation assumes that whatever relationship you’ve observed will continue in the same way. If sales grew 5% per year for a decade, you might assume they’ll keep growing at 5%. If doubling a drug dose doubled its effect in early trials, you might assume that pattern continues. But real systems are full of thresholds, feedback loops, and saturation points that make this assumption dangerous.
Engineering research on vibrating structures illustrates this clearly. A beam that behaves predictably under small vibrations can behave in completely different ways under larger ones, because the physics changes at higher amplitudes. A model built from low-amplitude data will fail when asked to predict high-amplitude behavior. The structure’s response doesn’t just scale up proportionally; it shifts into a different regime entirely. This is true of countless systems: population growth hits resource limits, chemical reactions saturate, materials reach breaking points, and markets hit ceilings or floors that didn’t exist in the training data.
When something behaves linearly in a narrow range, it’s tempting to assume it’s linear everywhere. But linearity is often just a local approximation. Zoom out far enough and nearly every real-world relationship curves, flattens, or breaks.
More Complex Models Make It Worse, Not Better
A natural reaction to failed extrapolation is to build a more complex model, one that captures every wiggle in your data. This backfires spectacularly. Research using COVID-19 data across multiple countries showed that the most complex models (those using high-dimensional mathematical functions) achieved near-perfect accuracy on the data they were trained on, yet produced the worst predictions on new, unseen data. The simplest models, by contrast, had mediocre performance on the training data but gave the best real-world forecasts.
This tradeoff is called overfitting. A complex model doesn’t just learn the true underlying pattern; it also memorizes the noise and random fluctuations in the training data. Those fluctuations won’t repeat in the future, so a model that’s learned them will confidently predict things that never happen. A phenomenon called Runge’s effect demonstrates this vividly: as you increase the degree of a polynomial fitted to data points, the curve passes perfectly through every point but swings wildly between them and especially beyond them. The in-sample fit looks flawless while the out-of-sample predictions become absurd.
The lesson is counterintuitive. When it comes to predicting outside your data, a simpler, slightly “wrong” model will typically outperform a complex, perfectly fitted one.
Long-Range Forecasts Lose Predictive Power
Economic forecasting provides some of the starkest evidence for extrapolation’s limits. Analysis by the Federal Reserve found that prediction accuracy for economic variables like growth, inflation, and interest rates deteriorates steadily as the forecast horizon lengthens. At short horizons (a quarter or two ahead), professional forecasts are meaningfully better than simply guessing the historical average. At longer horizons, they aren’t. The forecasts explain little to none of the actual variation in what eventually happens.
This finding isn’t unique to one institution. It has been documented across the Survey of Professional Forecasters, the Federal Reserve’s own internal forecasts, and forecasts for other large economies. The pattern is consistent: the further into the future you try to project current trends, the less value that projection has. At some point, your elaborate forecast is no more useful than the long-run average.
Part of the reason is that economic conditions shift in ways that historical data can’t anticipate. Recessions, policy changes, technological disruptions, and financial crises create breaks in the underlying structure of the economy. While researchers have studied whether modeling these breaks improves forecasts, the results are surprisingly mixed. Small shifts are often better ignored because trying to account for them introduces its own errors: imprecisely estimated timing and poorly calibrated post-break parameters can make forecasts worse, not better. This creates a frustrating situation where you know the world has changed but can’t reliably incorporate that knowledge into your predictions.
Biological Extrapolation Can Be Dangerous
Some of the most consequential extrapolation failures happen in medicine, where scaling results from one species to another is routine but fraught with risk. Simple body-weight scaling of drug doses from animals to humans is unreliable because species differ in how quickly they process drugs, how much of a drug binds to proteins in the blood, and how their immune systems respond.
A well-known example involves resveratrol, a compound found in red wine. When researchers gave mice a dose of 22.4 mg per kilogram of body weight, media outlets simply multiplied by human body weight and reported the equivalent dose as about 1,344 mg per day for a person. That calculation ignored the differences in metabolism between mice and humans, producing a number so high it undermined public trust in the research.
More dangerous cases exist. A dose of LSD that produced only mild effects in cats and primates proved toxic to an elephant, because elephants metabolize the drug more slowly, leading to higher and longer-lasting blood concentrations than simple weight-based scaling would predict. The most tragic example may be TGN1412, a drug targeting immune cells that appeared safe in animal testing. When six healthy human volunteers received it in a 2006 clinical trial, all six developed multi-organ failure. The drug interacted with the human immune system in ways that animal models simply couldn’t predict.
These cases highlight a core problem: extrapolation assumes that the relationship you’ve observed in one context transfers to another. When the underlying biology, chemistry, or physics differs between contexts, even slightly, that assumption can fail catastrophically.
Why People Extrapolate Anyway
Given all these risks, extrapolation persists because it’s often the only option available. You can’t always collect data from the exact conditions you want to predict. You can’t run a clinical trial on every species, forecast the economy by living through it first, or test a bridge at every possible load before building it. Extrapolation fills the gap between what you know and what you need to decide.
The key is recognizing what you’re doing when you extrapolate. You’re not extending knowledge; you’re extending assumptions. The further you go, the more you’re betting that nothing important changes. Keeping projections close to your data range, using simpler models, acknowledging widening uncertainty, and looking for independent evidence that your assumptions hold in the new range are all ways to extrapolate more responsibly. But the fundamental limitation remains: outside your data, you’re guessing with math, and the math can’t warn you about things it’s never seen.