How to fit and judge a model 01 Paste the data. Comment lines starting with # or // are skipped, and rows that do not parse as two numbers are reported by line number rather than silently dropped. 02 Choose a family, not a formula. The model list is organised by shape; a decay belongs to exponential growth/decay, a saturation curve to Michaelis–Menten or exponential plateau. 03 Read the parameters with their intervals. A parameter whose 95 % interval crosses zero is not evidence that it is needed; a tight interval is not evidence that the model is right. 04 Compare across families. The compare mode ranks fits by AIC, which penalises extra parameters, and shows adjusted R² next to plain R². 05 Check the prediction band before extrapolating. The band widens away from the data; if you are predicting outside the fitted range, look at the interval, not the curve.
Worked readouts Take the exponential y = 2.5·e^(−0.4x) sampled at x = 0, 0.5, …, 6, then add a small alternating noise of ±0.05:
Fit
Parameters
R²
Adjusted R²
RMSE
AIC
Exponential (true family)
a = 2.515414 , b = −0.402286
0.994984
0.994528
0.049678
−72.057
Quadratic
a = 0.070451 , b = −0.777154 , c = 2.436773
0.990802
0.988962
0.067272
−62.174
Exponential plateau
y₀ = 2.52585 , y<sub>m</sub> = 0.03714 , k = 0.41747
0.995116
0.994140
0.049019
−70.404
Double exponential
a = 0.081428 , b = 63.8975 , c = 2.46857 , d = 0.394163
0.995426
0.993901
0.047440
−69.256
The exponential parameters bracket the true values: the 95 % interval is 2.42511 … 2.605718 for a and −0.426244 … −0.378327 for b. At x = 7 the fitted value is 0.150535 with a band of 0.128419 … 0.172650 , against a true value of 0.152025 . Extrapolating to x = 20 gives 0.00080612 against a true 0.00083866 — a 4 % error in a region where there is no data to constrain the fit.
Why more parameters always fit better Least squares minimises the sum of squared residuals, and adding a parameter can only lower that sum. That is why plain R² is a poor referee: the double exponential's four parameters reach R² = 0.995426 , above the true model's 0.994984 , and the exponential plateau edges past it too. Adjusted R² and AIC subtract a penalty for the parameters used, and the ranking flips: the two-parameter exponential wins on AIC at −72.057 . "Best fit" is undefined until you say what a parameter costs.
The intervals come from the local curvature of the residual surface at the solution; they are local and symmetric, and become unreliable when parameters trade off against each other. The double exponential shows it — one amplitude (a) collapses toward zero against the other (c) while its decay rate (b) compensates. A noiseless data set is the limiting case: the fit recovers a = 2.5 , b = −0.4 with SSE 0 , and the band at x = 7 collapses to 0.152025 , because the residuals carry no uncertainty information. Zero-width intervals describe the data, not the model.
Limits and edges
Least squares needs a starting point and a shape. Levenberg–Marquardt is local; a model that starts far from the data can converge to a poor local minimum, and the panel reports convergence and iteration count so you can see when it did not settle.
The data must be numeric pairs. Missing values are not imputed and units are not tracked; a badly parsed row is listed by line number, and the compare tab declines data sets above 5,000 points instead of running blindly. What it does not do. It does not choose a model for you, test residuals for normality, or handle weighted and replicated measurements — it fits the model you name to the points you paste. For summary statistics and hypothesis tests use Statistics ; for a linear regression with inference on the coefficients, the Regression panel ; for controller identification from a step, Control Design .
The extrapolation trap The signature trap is trusting the curve outside the data. The fitted exponential predicts 0.00080612 at x = 20 while the truth is 0.00083866 ; the 4 % gap looks harmless until you notice that the confidence band at x = 7 is already ±0.022 , roughly 15 % of the value, and grows with distance. A quadratic fitted to the same decay would turn upward and cross zero somewhere past the data, a shape the physics forbids. Fit within the range you measured, quote the band, and treat any prediction beyond it as an assumption about the model, not a measurement.
The second trap is comparing R² across data sets. R² is relative to the variance of the particular y column; the same model can score 0.99 on a quiet signal and 0.85 on a noisy one without becoming worse. Compare models on the same data, and use RMSE when you need an error in the units of y — here 0.0497 for the exponential against 0.0673 for the quadratic.
Where it earns its place Calibrating a sensor response Fit a known exponential or power law to calibration points, then check that the parameters match the physics: a = 2.515 against a nominal 2.5 is a calibration statement, while the interval 2.425 … 2.606 says how much of that is sampling noise. The band at a working point becomes the uncertainty you carry into the next calculation.
Choosing between competing mechanisms When two mechanisms predict different shapes, the compare table settles it by information criterion rather than by eye: exponential AIC −72.057 against quadratic −62.174 is a gap of nearly ten, strong evidence for the decay form; a gap of one or two would call for more data instead of a conclusion.
Privacy Data, fits and bands are computed in your browser; nothing is uploaded.
References
Wikipedia, Least squares , en.wikipedia.org (访问日期:2026-10-07)— the objective every fit minimises.
Wikipedia, Coefficient of determination , en.wikipedia.org (访问日期:2026-10-07)— why R² never falls when parameters are added.
Wikipedia, Akaike information criterion , en.wikipedia.org (访问日期:2026-10-07)— the parameter penalty used in the compare table.
Wikipedia, Levenberg–Marquardt algorithm , en.wikipedia.org (访问日期:2026-10-07)— the damped Gauss–Newton iteration behind the fit.
NIST/SEMATECH e-Handbook of Statistical Methods, §4.1 Least Squares, itl.nist.gov (访问日期:2026-10-07)— residuals, degrees of freedom and uncertainty.
Calculators in this hub
Hand-picked tools, one click away. The mini versions compute live and carry your values into the full calculator.
Sources & review
Reviewed by CalcX Editorial Team
Updated 2026-10-07