Curve fitting is the process of finding a mathematical equation that best represents a set of data points. You start with scattered measurements, choose a type of equation (a straight line, a curve, an exponential), and then adjust that equation’s parameters until it matches your data as closely as possible. It’s one of the most widely used techniques in data analysis, showing up everywhere from predicting disease spread to tracking how a drug moves through the body.
How Curve Fitting Works
Imagine you’ve collected 50 temperature readings over the course of a day. Plotted on a graph, they form a rough pattern, but the points are scattered because of measurement error and natural variation. Curve fitting lets you draw a smooth line or curve through that scatter, capturing the overall trend while ignoring the noise in individual measurements.
The process has three basic steps. First, you pick a type of equation you think describes the relationship in your data. This might be a straight line for something simple, a polynomial for a wavy pattern, or an exponential curve for something that grows or decays rapidly. Second, you adjust the parameters of that equation (the slope and intercept of a line, for instance) to minimize the gap between the curve and your actual data points. Third, you check how well the resulting curve fits, using specific metrics to judge whether the model is good enough or whether you need a different equation.
Least Squares: The Standard Approach
The most common method for finding the best fit is called least squares fitting. For each data point, you calculate the “residual,” which is simply the difference between the actual measured value and what your curve predicts at that same point. Some residuals will be positive (the data point sits above the curve) and some negative (below it). If you just added them up, the positives and negatives would cancel out, hiding how far off the curve really is.
To avoid that, least squares fitting squares each residual, making every value positive, then adds them all together. The best-fit curve is the one where this total squared difference is as small as possible. This approach naturally penalizes large errors more heavily than small ones, which tends to produce curves that track the data well without being thrown off by a single outlier. It’s the default method in virtually every curve fitting tool you’ll encounter.
Curve Fitting vs. Interpolation
People sometimes confuse curve fitting with interpolation, but they solve different problems. Interpolation finds a function that passes exactly through every single data point. Curve fitting finds a function that captures the overall trend without necessarily hitting any individual point. When your data contains measurement noise or uncertainty (which real-world data almost always does), curve fitting is typically the better choice because it smooths out that noise rather than treating every tiny fluctuation as meaningful.
How to Tell if a Fit Is Good
Once you’ve fit a curve to your data, you need a way to measure how well it actually works. No single metric has become the universal standard, but a few are used far more than others.
R-squared (also written R²) is the most intuitive. It represents the proportion of variation in your data that the curve explains. A perfect fit gives an R² of 1.0, meaning the curve accounts for 100% of the variation. A value of 0.85 means the curve explains 85% of the pattern, with the remaining 15% unexplained. The closer to 1.0, the better.
Root mean square error (RMSE) takes a different angle. It tells you, on average, how far your curve’s predictions are from the actual data, expressed in the same units as your data. If you’re fitting temperature data and your RMSE is 0.3°C, your curve is off by about a third of a degree on average. A perfect fit gives an RMSE of zero. Unlike R², smaller is better here. Both metrics are useful: R² tells you what fraction of the pattern your curve captures, while RMSE tells you the typical size of the error in practical terms.
Overfitting and Underfitting
The biggest pitfall in curve fitting is choosing a model that’s too complex or too simple for your data.
Overfitting happens when your curve is so flexible that it starts matching the noise in your data rather than the underlying trend. Picture a wiggly line that hits every data point perfectly but makes wild, unrealistic predictions between those points. The fit looks great on the data you already have, but it performs poorly on new data from the same source. As model complexity increases, the error on your existing data keeps dropping, but the error on new data starts climbing. That gap is the signature of overfitting.
Underfitting is the opposite problem. Your model is too simple to capture the real pattern. A straight line forced through data that clearly follows a curve will miss the trend in the original data and also predict poorly on new data. An underfitted model performs badly everywhere, not just on new observations.
The goal is to find the sweet spot: a model complex enough to capture the real pattern but simple enough to avoid chasing noise. This tradeoff between accuracy and simplicity is central to good curve fitting. A model with fewer parameters that still fits the data well is almost always preferable to one with many parameters that fits slightly better.
Real-World Applications
Drug Development
In pharmacology, curve fitting is used to model how drug concentrations change in the body over time. Researchers collect blood samples at regular intervals after administering a drug, then fit curves to the concentration data to estimate how quickly the drug is absorbed, how it distributes through tissues, and how fast the body eliminates it. This works even in complex scenarios like repeated dosing, where the method can detect changes in absorption or elimination rates from one dose to the next, flagging phenomena like the body speeding up or slowing down its processing of the drug over time.
Epidemiology
During the COVID-19 pandemic, curve fitting became a frontline tool for public health forecasting. Researchers at institutions like Los Alamos National Laboratory and the Institute for Health Metrics and Evaluation (IHME) used curve fitting models to project confirmed cases and deaths across different regions. These models looked at the current trajectory of the outbreak, fit mathematical curves to the case data, and extrapolated the likely path forward. IHME’s curve fitting model was a primary forecasting tool during the critical early months of the pandemic in the U.S., from late March through the end of April 2020.
Engineering and Everyday Software
If you’ve ever added a trendline to a chart in Excel or Google Sheets, you’ve done curve fitting. That trendline is a mathematical equation whose parameters were calculated to minimize the distance between the line and your data points. The same principle scales up to sophisticated research tools. Python’s SciPy library includes dedicated curve fitting functions, MATLAB offers polynomial fitting and regression tools, and specialized statistical software provides even more options. The underlying math is the same whether you’re fitting a line in a spreadsheet or modeling satellite orbits.
Choosing the Right Model
The equation you choose matters as much as the fitting method. A linear model (straight line) has just two parameters: slope and intercept. It works well when the relationship between your variables is roughly proportional. A polynomial adds curves and bends, with each additional degree adding a new parameter and more flexibility. Exponential models suit data that grows or decays at a rate proportional to its current value, like population growth or radioactive decay. Logarithmic models work for relationships that rise quickly at first and then level off.
There’s no formula for picking the right model. Start by plotting your data and looking at the shape. Let domain knowledge guide you: if you know the underlying process is exponential (like compound interest), use an exponential model. If you’re unsure, try a few options and compare their R² and RMSE values, keeping in mind that adding more parameters will always improve the fit on your current data but won’t necessarily improve predictions. The simplest model that adequately describes the trend is usually the right choice.