ML Performance Metrics

"The model has 95% accuracy" can mean brilliant or useless. Pick a type of machine learning, pick a metric, and play with a live simulation until the number makes sense. Written for people who do not build models but need to understand them.

100% in your browser. Every change re-runs the simulation instantly.

Play with the spam filter

Decision threshold 0.50 Emails scoring above this line get flagged as spam. This is the central trade-off of classification.

Model quality medium How well the model separates spam from real email.

How common is spam 30% Drop this to 5% and check accuracy again. Rare outcomes are where metrics fool people.

Metric0.00

Open-source acknowledgements

About the ML Performance Metrics Explorer

Machine learning results arrive as numbers: 95% accuracy, an AUC of 0.87, an RMSE of 12. If you sit in meetings where those numbers decide budgets and products, you need to know what they mean, where they mislead, and which question each one actually answers. This tool teaches exactly that, with live simulations instead of formulas.

How to use it

Pick the kind of problem. Classification sorts things into categories, regression predicts a number, and clustering finds groups nobody labeled. Each comes with a concrete running example: a spam filter, a house price model, and customer segments.
Pick a metric. The chips show every relevant metric with its live value from the simulation.
Play. Drag the sliders and watch the numbers respond. The card below the chart explains the selected metric in plain words: what it means, how to read it, and where it fools people.

Moments worth experiencing

The accuracy trap. Set spam to 5% of email and push the threshold high. The filter catches almost nothing, yet accuracy stays above 90%, because "it's probably fine" is nearly always right. This single demo explains more real-world ML confusion than any other.
Precision versus recall. Drag the threshold and watch the two numbers pull against each other. Every classification system in production sits somewhere on this trade, chosen by a human.
MAE versus RMSE. Flip on "one huge miss." MAE barely notices; RMSE jumps. Now you know which metric to ask for when rare disasters are what you fear.
The elbow. The clustering data secretly holds three groups. Watch inertia keep falling as k grows while the silhouette score quietly points at the honest answer.

Why the plain-English framing matters

Most metric explanations are written by practitioners for practitioners, so they start with formulas and end with Greek letters. The definitions here start with the question the metric answers, in the language of the example: how much spam slipped through, how far off the price guess was, whether the customer groups are really separate. The math is real underneath (every number is computed live from the simulated data), but you never need it to follow along.

A note on honesty

Every metric here comes with a "watch out" section, because every metric can be gamed or misread. Accuracy hides rare failures, precision can be inflated by timidity, R² rewards junk inputs, and the elbow is a judgment call. Knowing where a number lies to you is the difference between reading a dashboard and understanding one.

Common questions

What is the difference between precision and recall?

Precision asks: of everything the model flagged, how much was real? Recall asks: of everything real, how much did the model catch? Drag the decision threshold in the live simulation and watch them trade against each other, which is the relationship every model owner negotiates.

Why is accuracy a misleading metric?

With rare classes, a model can score 96% accuracy by catching almost nothing: predicting 'no fraud' always is very accurate and completely useless. The simulation makes the accuracy trap visible, which is why practitioners reach for precision, recall, and F1.

What does RMSE mean in regression?

Root mean squared error: the typical size of the model's mistakes, in the same units as the prediction, with big misses punished extra hard. Add an outlier in the simulation and watch RMSE panic while MAE shrugs; the comparison teaches when to use which.

What is a good silhouette score in clustering?

Silhouette measures whether points sit snugly in their own cluster versus near a neighboring one, from -1 to 1. Above 0.5 suggests real structure; near zero suggests the clusters are wishful thinking. Adjust the blobs and watch it respond.

Who is this tool for?

Anyone who reads or reports model results without wanting to take a course first: product managers, analysts, students, and engineers who want the vocabulary to actually mean something.