Choose an application
08 / Data & model testing
Fit one set.
Test on another.
Compare explicitly declared models using data in your browser. Inspect fitted coefficients, conditional uncertainty and residual structure before interpreting a difference.
Ordinary or weighted least squares · local CSV processingDeclare the comparison
The extension adds a specified basis term or harmonic. These are exploratory comparators; an extra fitted coefficient is not automatically a prediction of DTT.
Compare a baseline and an extension
The initial dataset is synthetic teaching data. Your CSV is read locally; neither its contents nor fitted values are sent to a server.
Required columns: x,y. Optional: sigma (a positive standard deviation for every row), split (train or test). Commas and decimal points; no missing values. Up to 5,000 rows / 2 MB. Column order may vary. Format and examples →
Enable JavaScript to load an example or analyze data.
Training fit and independent test error
Observed values and fitted models
Baseline · solid blueExtension · dashed orangeTraining · dark circles; test · orange squares
Vertical bars show supplied sigma when present. Up to 300 observations are drawn for readability; every row is used in the calculation and export.
Extended-model residuals
Residual = observed − predicted. Test residuals use coefficients fitted only to the training rows. Look for curvature, drift and changing spread.
Coefficients and conditional standard errors
Inspect observations and predictions
A repeatable sequence
Make the physical meaning of the columns and the model choices explicit before fitting.
- Specify the observable. Record units, preparation, acquisition order and uncertainty convention.
- Choose the comparator. State why the proposed term is expected, and identify relevant nuisance effects.
- Reserve test observations. Use a separate run or protocol when possible. The automatic row split does not guarantee independence.
- Fit the training subset. Keep coefficients fixed when evaluating the test subset.
- Inspect residuals and export. Save the input rows, allocation, model basis, covariance and results together.
Read the result at the right level
Lower training error is expected when adding parameters. A lower test error can be useful evidence for a better comparator, but does not establish the physical mechanism.
This workbench assumes independent errors in y and ignores uncertainty in x. It does not model correlated systematics, Poisson counts, parameter constraints or a full experimental likelihood. Coefficient errors are conditional on the selected model. Free fringe fits can produce values outside the probability range; such results require a constrained physical model.
NIST: weighted least squares · NIST: model validation and residuals