Model Limitations — I can analyze model limitations...

SECTION 1

Historical Context & Motivation

Humans have been building mathematical models—simplified representations of the real world using equations, graphs, and data—for thousands of years. Ancient astronomers modeled planetary motion as perfect circles, and for a while, those models worked well enough to predict eclipses and planting seasons. But every model eventually bumps up against situations where its predictions go wrong. Understanding when and why models fail has been just as important as building the models themselves.

~150 CE

Ptolemy's Geocentric Model

Ptolemy modeled the solar system with Earth at the center. It predicted planetary positions reasonably well for centuries but required increasingly complicated "epicycles" to patch its errors—a classic sign of a model pushed past its limits.

1687

Newton's Laws of Motion

Newton's equations modeled motion beautifully at everyday speeds. Scientists assumed these laws were universal until Einstein showed they break down near the speed of light—an example of hidden assumptions limiting a model.

1854

John Snow's Cholera Map

Snow used a spatial data model to trace cholera cases to a specific water pump in London. His model saved lives but could not explain the biological mechanism—it had clear limits in its explanatory power.

2008

Financial Model Failures

Banks relied on mathematical models that assumed housing prices would always rise. When that assumption broke, the models failed catastrophically, contributing to a global financial crisis and highlighting the danger of extrapolation.

2020

Pandemic Projection Models

COVID-19 models projected infection rates using early data. Wide uncertainty ranges communicated how limited the predictions were, teaching the public that models provide scenarios, not certainties.

The recurring lesson across these moments is clear: no model is perfect. The real skill isn't just building a model; it's knowing where the model stops being reliable and being honest about what it can't tell you. That's exactly the skill you'll develop in this lesson.

SECTION 2

Core Principles & Definitions

Before you can analyze a model's limitations, you need a shared vocabulary. Every mathematical model rests on simplifying choices, and those choices create predictable weak spots. The following core ideas will equip you to evaluate any model you encounter in math, science, or everyday life.

Interpolation vs. Extrapolation

Interpolation means making predictions within the range of your existing data—generally safe. Extrapolation means predicting beyond that range, where the model may no longer reflect reality.

Assumptions

Every model is built on assumptions—conditions that are taken to be true without proof. For example, a linear model assumes a constant rate of change. When assumptions don't match reality, the model's output becomes unreliable.

Domain of Applicability

The domain of applicability is the set of input values for which the model produces reasonable outputs. Outside this domain, predictions can become meaningless—like a model predicting negative population.

Residuals & Error

A residual is the difference between an observed value and the model's prediction. Large or patterned residuals indicate the model is missing something important in the data.

Uncertainty

Uncertainty describes how much confidence you should place in a model's predictions. Communicating uncertainty—through ranges, margins of error, or qualifying language—is an essential part of responsible modeling.

✦ KEY TAKEAWAY

Think of a model like a map of a city. The map is incredibly useful for navigating streets you can see on it (interpolation), but if you drive off the edge of the map (extrapolation), you have no idea what's out there. The map also assumes roads haven't changed since it was printed (assumptions). A good navigator acknowledges these limits instead of blindly trusting the map.

SECTION 3

Visualizing Interpolation vs. Extrapolation

The diagram below shows a scatter plot of data points with a best-fit curve drawn through them. The shaded green region marks the interpolation zone where the model is most trustworthy. Beyond the data, the curve extends into the extrapolation zones shown in red, where predictions become increasingly uncertain. Notice how the dashed uncertainty band widens dramatically as you move further from the data—this is a visual signature of growing uncertainty.

The cyan dots represent observed data. The solid violet curve is the model's best-fit line. Inside the green interpolation zone, the model closely matches the data. In the red extrapolation zones, the dashed uncertainty bounds spread apart, indicating that predictions there carry far more risk.

The key visual pattern to remember is this: the further your input travels from the observed data, the wider the uncertainty band becomes. Within the data range, residuals are small and the model is reliable. Beyond the data, the model is essentially guessing, and the uncertainty can grow without limit.

SECTION 4

Mathematical Framework

You don't need calculus to quantify how well a model performs or where it breaks down. A few straightforward formulas let you measure error, spot patterns in residuals, and put numbers on uncertainty.

RESIDUAL

eᵢ = yᵢ − ŷᵢ

where yᵢ is the observed value, ŷᵢ is the model's predicted value, and eᵢ is the residual (error) for the i-th data point. Positive residuals mean the model under-predicted; negative residuals mean it over-predicted.

PERCENT ERROR

Percent Error = |yᵢ − ŷᵢ| / |yᵢ| × 100%

This expresses the residual as a percentage of the observed value, making it easier to judge whether the error is large or small in context. A 5-unit error means very different things when the observed value is 10 versus 10,000.

COEFFICIENT OF DETERMINATION

R² = 1 − (Σeᵢ²) / (Σ(yᵢ − ȳ)²)

R² measures the proportion of total variation in y that the model explains. Values close to 1 indicate a strong fit; values near 0 indicate the model explains little. However, a high R² does not guarantee the model is appropriate for extrapolation—it only describes fit within the data range.

⚠️ Why R² Can Mislead

A polynomial of degree n − 1 can pass through all n data points, giving R² = 1. But such a model overfits: it captures noise rather than the true pattern, and its extrapolated predictions can be wildly inaccurate. Always check residual plots, not just R².

LINEAR MODEL WITH UNCERTAINTY

ŷ = mx + b ± margin of error

Reporting a prediction as a single number is incomplete. Including a margin of error (or confidence interval) communicates the range within which the true value is likely to fall. The margin typically grows as x moves further from the center of the data.

SECTION 5

Types of Model Limitations

Model limitations fall into several categories. Some are baked into the model's structure; others emerge from the data it was built on. The diagram below classifies the most common types of limitations you'll encounter in Grade 11 math and beyond.

Model limitations can be structural (built into the model's form), data-related (arising from the data used to build it), or contextual (due to changes or misinterpretations of what the model represents).

Let's look at a concrete example of each category. If you fit a straight line to data that actually follows an exponential curve, that's a structural limitation—the model's shape is fundamentally wrong. If you use that linear model to predict values ten years beyond your data, that's a data-related limitation because you're extrapolating. And if a pandemic changes consumer behavior after you built a sales model, that's a contextual limitation—the world shifted under your model.

SECTION 6

Worked Example — Coffee Cooling Model

A student measures the temperature of a cup of coffee every 5 minutes and fits a linear model to the first 20 minutes of data. The model is T = −3t + 85, where T is temperature in °C and t is time in minutes. The room temperature is 22 °C. Let's analyze this model's limitations step by step.

Analyzing the Coffee Cooling Model

Step 1 — Identify the Model's AssumptionsThe linear equation T = −3t + 85 assumes the temperature drops at a constant rate of 3 °C per minute. This is a simplifying assumption. In reality, Newton's Law of Cooling tells us the rate of cooling depends on the difference between the coffee and room temperature, so the rate slows as the coffee cools.

Assumption: constant cooling rate of −3 °C/min.

Step 2 — Determine the Domain of ApplicabilityThe data was collected from t = 0 to t = 20 minutes. Within this range, interpolation is reasonable. Beyond t = 20 minutes, the model is extrapolating. Let's check: at t = 20, T = −3(20) + 85 = 25 °C, which is plausible (just above room temp). But the model doesn't know to stop at room temperature.

Domain: 0 ≤ t ≤ 20 minutes (interpolation zone).

Step 3 — Test the Model with ExtrapolationAt t = 30: T = −3(30) + 85 = −5 °C. The model predicts the coffee is below freezing! Since the coffee can't go below room temperature (22 °C) without a freezer, this prediction is physically impossible. This is extrapolation failure in action.

Extrapolation at t = 30 gives T = −5 °C — impossible.

Step 4 — Compute Residuals for Known DataSuppose the actual measured temperatures at t = 5, 10, 15, 20 were 71, 60, 50, 42 °C. The model predicts 70, 55, 40, 25. Residuals: 71 − 70 = +1, 60 − 55 = +5, 50 − 40 = +10, 42 − 25 = +17. The residuals are growing systematically, which signals the linear model is increasingly wrong even within the data range.

Residuals: +1, +5, +10, +17 — a pattern indicating model misfit.

Step 5 — Communicate UncertaintyBased on our analysis, a responsible conclusion would be: "The linear model approximates the coffee's temperature during the first 10 minutes (residuals ≤ 5 °C) but increasingly underestimates the temperature after that. Predictions beyond 20 minutes are unreliable because the model does not account for the cooling rate slowing as the coffee approaches room temperature. An exponential decay model would likely perform better for longer time horizons."

Conclusion: reliable for t ≤ 10 min; unreliable beyond 20 min; recommend exponential model.

SECTION 7

How to Communicate Uncertainty

Identifying limitations is only half the job. The other half is communicating uncertainty clearly so that anyone using your model knows what to trust and what to question. Scientists, engineers, and analysts use specific strategies to do this, and you can apply the same techniques in your math class.

Five strategies for communicating uncertainty in model predictions

Strategy	What It Looks Like	Example
Qualifying Language	Words like "approximately," "within the data range," "assuming conditions remain constant"	"The model predicts approximately 150 users, assuming current growth continues."
Error Bounds	Reporting predictions with ± margins	"Predicted revenue: $42,000 ± $5,000"
Confidence Intervals	Specifying a range and a probability level	"We are 95% confident the true value lies between 38 and 46."
Residual Analysis	Plotting residuals to show patterns of misfit	"Residuals show a curved pattern, suggesting a nonlinear model may fit better."
Stating Assumptions Explicitly	Listing conditions the model relies on	"This model assumes no new competitors enter the market."

✦ KEY TAKEAWAY

Communicating uncertainty is like a weather forecast. A meteorologist doesn't just say "it will rain tomorrow"—they say "there's a 70% chance of rain, with accumulation between 0.5 and 1.2 inches." That extra information helps you decide whether to cancel an outdoor event. Similarly, reporting a model's prediction with uncertainty helps decision-makers understand the risk.

SECTION 8

Connection to Advanced Modeling

The principles you're learning now—checking assumptions, recognizing extrapolation, and communicating uncertainty—are the same ones used in advanced statistics, machine learning, and scientific research. The table below shows how Grade 11 concepts connect to what you'll encounter later.

How Grade 11 modeling concepts extend into advanced coursework

Grade 11 Concept	Advanced Version	What Changes
Residual = observed − predicted	Standardized residuals, Cook's distance	Residuals are scaled by their expected variability to identify influential points.
R² for goodness of fit	Adjusted R², AIC, BIC	Advanced metrics penalize models for having too many parameters (overfitting).
Stating assumptions	Formal hypothesis testing of assumptions	Statistical tests (e.g., Durbin-Watson) formally check if assumptions like independence hold.
± margin of error	Prediction intervals, Bayesian credible intervals	Intervals are computed rigorously from probability theory and account for sample size.
"Don't extrapolate too far"	Cross-validation, out-of-sample testing	Models are systematically tested on data they've never seen to gauge real-world performance.

The exciting takeaway is that you're already thinking like a data scientist. The specific tools get more powerful in AP Statistics, college-level regression, and machine learning, but the core question never changes: "How much should I trust this model, and where does it break?"

SECTION 9

Practice Problems

PROBLEM 1 — CONCEPTUAL

A student creates a linear model of plant height over 8 weeks of growth. They use the model to predict the plant's height at week 52. Is this interpolation or extrapolation? Why should the student be cautious about this prediction?

PROBLEM 2 — BASIC CALCULATION

A model predicts that a car's value after t years is V = −2500t + 30000. The actual values at t = 1, 2, 3 are $28,000, $24,500, and $22,000. Calculate the residual for each year. Do the residuals suggest the model is a good fit?

PROBLEM 3 — INTERMEDIATE

A quadratic model y = −0.5x² + 10x + 2 is fit to data for 0 ≤ x ≤ 12 with R² = 0.97. A classmate says, "R² is almost 1, so the model is perfect and we can use it for any value of x." Identify at least two flaws in this reasoning.

PROBLEM 4 — APPLIED

A town collects monthly electricity usage data from January to August and fits the model E = 120 + 15 sin(πm/6), where E is usage in kWh and m is the month number (1 = January). State two assumptions this model makes. Then predict usage for December (m = 12) and discuss whether this is interpolation or extrapolation, and how confident you should be.

PROBLEM 5 — CRITICAL THINKING

Two students each model the same data set of U.S. smartphone ownership from 2010 to 2018. Student A uses a linear model and gets R² = 0.95. Student B uses a logistic model (S-curve) and gets R² = 0.99. Both extrapolate to 2030. Student A predicts 120% ownership; Student B predicts 92%. Which student's extrapolation is more defensible, and why? What does this example teach us about the relationship between model type and extrapolation risk?

SUMMARY

Lesson Summary

Every mathematical model is a simplified version of reality, and that simplification introduces limitations. The three major categories are structural limitations (choosing the wrong model type or oversimplifying), data-related limitations (extrapolating beyond the data range, small samples, or outliers), and contextual limitations (changed conditions or confusing correlation with causation). Interpolation (predicting within the data range) is generally safe, while extrapolation (predicting beyond it) carries increasing risk the further you go.

To analyze a model responsibly, identify its assumptions, compute residuals to check for patterns of misfit, and determine the domain of applicability. When reporting predictions, always communicate uncertainty using qualifying language, error bounds, or confidence intervals. A model's value isn't diminished by having limitations—the value comes from understanding those limitations and using the model wisely within them.