Whether a model's complexity is locked in before it sees a single row of data, or grows right along with the dataset.
A parametric model commits to a fixed functional form before it ever sees data — linear regression, for instance, is always "intercept plus slope times x," whether you fit it on 10 rows or 10 million. Its parameter count is fixed by that choice, not by the dataset, which makes it fast to fit and low-variance, but wrong wherever the true pattern doesn't actually look like the shape it assumed. A non-parametric model — k-nearest neighbors, decision trees, kernel methods — makes no such upfront commitment, and its effective complexity grows as more training data arrives. That flexibility lets it fit patterns a fixed-form model never could, at the cost of needing more data to do it safely and a much higher risk of fitting noise instead of signal.
Parametric means the model's functional form — and therefore its number of parameters — is chosen in advance, independent of how much data you have. Linear regression on one input has exactly two parameters, an intercept and a slope, whether the training set has 10 rows or 10 million. Once fit, prediction only ever needs those two numbers; the training data itself can be discarded. That fixed dimensionality is what makes parametric models fast, cheap to store, and low-variance — but also means they carry a hard ceiling on the patterns they can represent. If the true relationship curves and a line can't, no amount of additional data will fix that: the error from a wrong assumed shape is bias, and more data doesn't reduce it.
Non-parametric means no fixed functional form is assumed up front — and, critically, the model's effective complexity is allowed to grow as the training set grows. K-nearest neighbors makes a prediction by consulting the nearby training points directly, so it effectively keeps all of them as live "parameters" — more training rows means more stored reference points shaping every prediction. A decision tree can keep splitting into more leaves as more data gives it more room to split safely. A kernel method places one kernel per training point. In every case, the more data you feed the model, the more effective parameters it ends up using — which is exactly what lets it approximate arbitrarily complex true patterns, given enough data, at the cost of needing that data and being more exposed to variance (for the full mechanics of that tradeoff, see the companion Bias-Variance Tradeoff explainer).
For a parametric model, the number of parameters is a property of the chosen model class, fixed before training starts — linear regression on one input is always 2 parameters; a degree-3 polynomial is always 4. Training data only refines the values of those fixed slots; it never adds more of them.
For a non-parametric model, the effective parameter count is a property of the dataset. K-nearest neighbors literally stores every one of the N training points and consults them at prediction time — its effective parameter count isN. A decision tree's leaves can multiply as more data provides safe room to split further. A kernel method centers one kernel on every training point, so N points means N kernels. In every non-parametric case, more data structurally means a more complex model — which is precisely the mechanism behind Diagram 2 below.
Now swap the regression line for a classification boundary and use a non-parametric method — nearest-neighbors-style — to draw it. With few points, the simplest boundary that separates the two classes is smooth and simple. Add more points, including a couple of noisy ones sitting on the "wrong" side, and the boundary doesn't just get more detailed in a good way — it visibly sprouts extra loops chasing individual outliers, which is the overfitting risk that comes bundled with this flexibility.
False — and it's almost the opposite of what actually happens. "Non-parametric" is a statistical term of art meaning the number of effective parameters is not fixed in advance and is allowed to grow as more training data arrives — not that the model has zero parameters. K-nearest neighbors doesn't discard the training set the way linear regression does; it keeps every single training point as live information it consults for every future prediction, which means a non-parametric model can end up with far moreeffective parameters than a parametric one, not fewer. A deep, fully-grown decision tree can have a number of leaf regions approaching the number of training rows itself. The dividing line was never "has parameters" versus "has none" — it's whether the parameter count is nailed down by the model class before you see data (parametric), or determined by the dataset itself and free to grow with it (non-parametric).
Explains the core distinction in how model complexity relates to training data. A parametric model (linear regression, logistic regression) commits to a fixed functional form before seeing data, so its parameter count stays constant no matter how much data you add — fast and low-variance, but biased if the assumed shape is wrong. A non-parametric model (k-nearest neighbors, decision trees, kernel methods) makes no fixed assumption about the functional form, and its effective complexity grows as the training set grows — flexible enough to fit complex real-world patterns, but more data-hungry and more prone to overfitting.
"Non-parametric" is one of the most misleading names in statistics, because it sounds like it should mean "has no parameters." It means the opposite in an important sense: the number of effective parameters is not pinned down by the model class in advance, and is instead determined by — and grows with — the training data. K-nearest neighbors keeps every training point as a live reference; a fully grown decision tree can have a leaf for nearly every distinct region in the data. Both can end up with vastly more effective parameters than a simple parametric model ever has, not fewer.
A parametric model's parameter count is a property of the chosen model class: linear regression on one input is always 2 parameters (intercept, slope); a degree-k polynomial is always k+1 parameters. Training data changes the fitted values of those parameters, never their count. Predictions can be made from the fitted parameters alone — the original training data can be thrown away.
A non-parametric model's effective parameter count is a property of the dataset. K-nearest neighbors needs the entire training set at prediction time — more rows means more stored reference points shaping every future prediction. A decision tree can keep splitting into more leaves as more data provides statistically safe room to split further. A kernel method (like kernel density estimation or a kernel-based regression) centers one kernel on every training point, so N points structurally means N kernels. The common thread: as N grows, so does the model's working complexity.
Parametric models are fast to fit, cheap to store and deploy, and low-variance — retraining on a slightly different sample from the same distribution barely changes the fitted line. Their weakness is bias: if the true relationship doesn't look like the assumed functional form, no amount of additional data fixes that, because there is no extra capacity for the model to absorb the missed structure.
Non-parametric models can, given enough data, approximate essentially any underlying pattern, because their capacity grows to match it. Their weakness is variance and data requirements: with too little data relative to the pattern's true complexity, a non-parametric model's flexibility gets spent fitting noise instead of signal — exactly what the two dashed loops in the decision-boundary diagram illustrate. Choosing between the two families is really a bet about how much you trust an assumed functional form versus how much clean data you have to let a flexible model find its own shape safely.
Parametric: linear regression, logistic regression, Naive Bayes (given its distributional assumption), a fixed-architecture neural network (its layer sizes are chosen before training, so its parameter count is fixed once architecture is set — though in practice it often has enough parameters to behave close to non-parametric).
Non-parametric: k-nearest neighbors, decision trees and random forests, kernel density estimation, kernel regression / support vector machines with an RBF kernel, non-parametric statistical tests (like the Mann-Whitney U test, which makes no assumption about the underlying distribution's shape).
No — this is the single most common misreading of the term. It means the number of effective parameters isn't fixed by the model class in advance and can grow with the training data, not that there are literally none. K-nearest neighbors, for instance, effectively treats every stored training point as a parameter, so more data means more parameters, not fewer.
Technically parametric, in the strict sense that a network's architecture (number of layers and units) is fixed before training and doesn't change size as more data arrives — the parameter count is set by the architecture choice, not the dataset. In practice, though, large modern networks are often deliberately over-parameterized relative to the data, so they can behave much more like flexible non-parametric methods than classic small parametric models like linear regression.
Because their effective complexity scales with the data, a non-parametric model fit on a small dataset either has too little capacity to be useful yet, or — more commonly — has more than enough capacity relative to the data available, and spends that extra flexibility fitting sampling noise instead of genuine structure. Parametric models avoid this specific failure mode because their capacity is capped regardless of dataset size, though they carry their own risk of being too capped to represent the truth at all.
Only if the assumed functional form happens to match the true relationship — in which case the parametric model is not just competitive but strictly better, since it needs far less data to estimate a handful of fixed parameters accurately. If the assumed form is wrong, no amount of data closes that gap, because the parametric model has no mechanism to represent structure outside its assumed shape. That bias-versus-flexibility bet is the whole reason both families exist.
Directly — parametric versus non-parametric is one of the main structural choices that determines where a model sits on the bias-variance curve. A parametric model is biased toward the high-bias, low-variance end because its capacity is capped. A non-parametric model is biased toward the low-bias, high-variance end because its capacity grows with data, letting it approach the true pattern but making it more sensitive to whatever particular sample it was trained on. See the companion Bias-Variance Tradeoff explainer for the full mechanics.
Try our AI and Data Science Studio
More calculators, simulators, and guides for this discipline.