What this calculator does
This tool fits a chosen curve-regression model to your table of paired observations (x, y) using the method of least squares. It reports the fitted coefficients (A, B, and C for quadratic), the correlation coefficient r that measures how well the model describes the data, and an estimated value computed directly from the fitted equation. It is a universal mathematics and statistics tool with no region-specific rules.
The seven models
You can choose Linear (\(y = A + B\cdot x\)), Logarithmic (\(y = A + B\cdot \ln x\)), e-Exponential (\(y = A\cdot e^{Bx}\)), ab-Exponential (\(y = A\cdot B^{x}\)), Power (\(y = A\cdot x^{B}\)), Inverse (\(y = A + B/x\)), or Quadratic (\(y = A + B\cdot x + C\cdot x^{2}\)). All except the quadratic are fitted by linearizing into the form \(Y = a + b\cdot X\) and applying ordinary least squares; the quadratic is solved from its \(3\times 3\) normal equations.
How to use it
Enter your x values and y values as comma-separated lists of equal length, pick a regression type, choose whether you want to estimate y from x or x from y, and type the known input value. The calculator returns the coefficients, the correlation r, and the estimate.
The formula explained
For the linearized pairs (X, Y): \(b = S_{xy} / S_{xx}\) and \(a = \bar{Y} - b\cdot \bar{X}\), where \(S_{xx} = \sum X^2 - n\bar{X}^2\) and \(S_{xy} = \sum XY - n\bar{X}\bar{Y}\). The correlation is $$r = \frac{S_{xy}}{\sqrt{S_{xx}\cdot S_{yy}}}.$$ Coefficients are then back-substituted (for example \(A = e^a\) in exponential and power models).
Worked example
For x = [1,2,3,4,5,6,7] and y = [3,5,7,8,11,13,14] with a linear fit: \(S_{xx} = 28\), \(S_{xy} = 53\), so $$B = \frac{53}{28} = 1.892857$$ and $$A = 8.714286 - 1.892857\cdot 4 = 1.142857.$$ The correlation is $$r = \frac{53}{\sqrt{28\cdot 101.4286}} = 0.99453.$$ Estimating y at \(x = 64\) gives $$1.142857 + 1.892857\cdot 64 = 122.2857$$ (an extrapolation well beyond the data range).
Step-by-Step: Fitting a Curve by Hand
- Choose the model. Pick linear, logarithmic, exponential (e-form or ab-form), power, inverse or quadratic, based on the shape of a scatter plot and what you know about the process.
- Check domain restrictions. Logarithmic needs x > 0. Exponential needs y > 0. Power needs both x > 0 and y > 0. Inverse needs x ≠ 0. If any data point violates the rule, that model cannot be fitted to the data.
- Transform the data. Replace x and y with X and Y as the model requires: X = ln x, Y = ln y, or X = 1/x. Leave a variable unchanged if the model does not transform it.
- Compute the sums, means and sums of squares on the transformed data (n points):
$$S_{xx}=\sum X^2-\frac{(\sum X)^2}{n},\quad S_{xy}=\sum XY-\frac{\sum X\,\sum Y}{n},\quad S_{yy}=\sum Y^2-\frac{(\sum Y)^2}{n}$$ and the means \(\bar X=\sum X/n\), \(\bar Y=\sum Y/n\). - Get the slope and intercept of the straight line \(Y = a + bX\):
$$b=\frac{S_{xy}}{S_{xx}},\qquad a=\bar Y-b\,\bar X$$ - Back-substitute to the original parameters. For the ab-exponential model, \(A=e^{a}\) and \(B=e^{b}\). For the e-exponential and power models, \(A=e^{a}\) and \(B=b\). For the logarithmic and inverse models, \(A=a\) and \(B=b\).
- Compute the correlation coefficient:
$$r=\frac{S_{xy}}{\sqrt{S_{xx}\,S_{yy}}}$$ This is calculated on the transformed data. - Estimate. Put your x into the fitted equation (for example \(\hat y = A\,x^{B}\) for a power fit). To estimate x from a given y, rearrange the equation, and check that the result lies in the model's domain.
Quadratic note. The quadratic \(\hat y = A + Bx + Cx^2\) is not linearized. Instead, solve the 3×3 normal equations:
$$\begin{aligned}\sum y &= nA + B\sum x + C\sum x^2\\ \sum xy &= A\sum x + B\sum x^2 + C\sum x^3\\ \sum x^2y &= A\sum x^2 + B\sum x^3 + C\sum x^4\end{aligned}$$You need the sums Σx, Σx², Σx³, Σx⁴, Σy, Σxy and Σx²y. Solve the system by Cramer's rule or Gaussian elimination. At least three distinct x values are required.
Linearization Transforms by Model
| Model | Equation | Transformed X | Transformed Y | Linear form | Back-substitution (a, b from the line) | Domain restriction |
|---|---|---|---|---|---|---|
| Linear | \(y = A + Bx\) | x | y | \(y = a + bx\) | A = a, B = b | None |
| Logarithmic | \(y = A + B\ln x\) | ln x | y | \(y = a + bX\) | A = a, B = b | x > 0 |
| Exponential (e-form) | \(y = A e^{Bx}\) | x | ln y | \(\ln y = a + bx\) | A = ea, B = b | y > 0 |
| Exponential (ab-form) | \(y = A B^{x}\) | x | ln y | \(\ln y = a + bx\) | A = ea, B = eb | y > 0 |
| Power | \(y = A x^{B}\) | ln x | ln y | \(\ln y = a + bX\) | A = ea, B = b | x > 0 and y > 0 |
| Inverse | \(y = A + B/x\) | 1/x | y | \(y = a + bX\) | A = a, B = b | x ≠ 0 |
| Quadratic | \(y = A + Bx + Cx^2\) | No linearization | Solve the 3×3 normal equations | A, B, C come directly from the solution | None (needs at least 3 distinct x values) | |
The e-form and ab-form exponentials describe the same curve family: \(B_{ab}=e^{B_{e}}\). They give the same fit and the same r.
Interpreting r and Choosing Between Models
- Range and sign. r lies between −1 and 1. A positive r means Y tends to increase with X, and a negative r means it tends to decrease. |r| near 1 indicates a strong fit to the fitted form, and |r| near 0 indicates a weak fit.
- Transformed models. For logarithmic, exponential, power and inverse fits, r is computed on the transformed data (for example ln y against ln x). It is therefore not directly comparable with the r of a linear fit on the original scale. Comparing models by r is only a rough guide.
- r² (explained variance). r² is the proportion of variance in the (transformed) Y explained by the fitted line. For example, r = 0.90 gives r² = 0.81, so about 81% of the variance is explained.
- Correlation is not causation. A high r does not show that one variable causes the other, nor that the chosen model is the true relationship. Different curves can fit the same data about equally well.
- Check residuals. Plot the differences between observed and fitted values. They should scatter randomly around zero. A visible pattern, such as a curve or a funnel shape, means the model is poorly suited, even when r is high.
- Extrapolation risk. Estimates inside the range of your x data (interpolation) are usually far more reliable than estimates outside it. Exponential, power and quadratic curves can diverge sharply beyond the data.
- Quadratic versus linear. A quadratic always fits at least as well as a linear model on the same data, because it has more parameters and contains the linear model as a special case (C = 0). A better fit alone therefore does not justify the extra term. Prefer the simpler model unless the improvement is meaningful and the curvature makes sense for the process.
- Sample size. With very few points, even a high r can arise by chance. Use as much data as is practical.
FAQ
How is r interpreted? Roughly: \(0.7 < |r| \le 1\) strong, \(0.4 < |r| < 0.7\) moderate, \(0.2 < |r| < 0.4\) weak, below \(0.2\) negligible.
Why are some models rejected? Logarithmic and power models require \(x > 0\), exponential and power models require \(y > 0\), and inverse requires \(x \ne 0\), because of the log or reciprocal transforms.
Can I extrapolate? Yes, but estimates outside the observed data range are extrapolations and should be treated with caution.