Residual Plot Calculator

Fit linear or polynomial regression models, compute residuals, and visualize residual plots to assess model assumptions. Detect heteroscedasticity, outliers, and non‑linearity with interactive charts.

Supports comma, tab, space, or semicolon separators. First two numbers on each line are used.
#xy
0 points
At least 3 data points required. For polynomial fits, degree must be less than number of points.
? Linear (positive) : x:1‑6, y:2‑12
? Linear (negative) : x:1‑6, y:15‑3
? Quadratic : x:‑3‑3, y: x² + noise
⚠️ Heteroscedastic : x:1‑10, y: x + ε·x
? With outlier : x:1‑8, y:2x+1, plus (5,30)
? Random scatter : x:1‑10, y: uniform noise
Privacy first: All computations are performed locally in your browser. No data is sent to any server.

What Is a Residual Plot and Why Does It Matter?

A residual plot is a graphical diagnostic tool used in regression analysis to visualize the difference between observed and predicted values — the residuals. Formally, for each observation i, the residual is ei = yi − ŷi, where ŷi is the value predicted by the fitted model. The residual plot displays these residuals on the vertical axis against either the fitted values, the predictor variable, or the observation index on the horizontal axis.

Residual plots are indispensable for validating the assumptions of ordinary least squares (OLS) regression: linearity, homoscedasticity (constant variance), independence of errors, and normality of residuals. By examining the pattern of residuals, analysts can detect model misspecification, heteroscedasticity, outliers, and influential points — issues that render statistical inference unreliable. This interactive tool empowers you to explore these diagnostics intuitively.

For a model with n observations and p parameters:

Residual: ei = yi − ŷi  ·  SSE = Σ ei2  ·  R² = 1 − SSE / SST

where SST = Σ (yi − ȳ)2 is the total sum of squares.

How to Interpret Residual Patterns

A well‑behaved residual plot exhibits random scatter around the zero line with no systematic structure. Deviations from this ideal reveal specific violations:

  • Fan‑shaped or funnel pattern → Heteroscedasticity: variance of residuals increases (or decreases) with fitted values. This violates the constant‑variance assumption and inflates standard errors.
  • Curvilinear pattern → Non‑linearity: the relationship is not properly captured by the model. A polynomial or transformation may be required.
  • Outliers far from the zero line → Unusual observations that exert disproportionate influence. Investigate data quality or consider robust methods.
  • Serial correlation (sequential pattern) → Autocorrelation in residuals, common in time‑series data. Suggests a missing variable or temporal structure.
  • Residuals with non‑constant spread → may indicate that the error distribution is not normal or that a transformation of the response variable is warranted.

Our calculator automatically generates both the fitted‑vs‑residual plot (the most common diagnostic) and the data‑with‑fit plot, giving you a comprehensive view of model performance.

Regression Models Supported

This tool implements ordinary least‑squares estimation for three model families:

  • Linear: y = a + b·x — the simplest and most widely used model.
  • Quadratic: y = a + b·x + c·x² — captures curvature with a single turning point.
  • Cubic: y = a + b·x + c·x² + d·x³ — allows up to two inflection points for more flexible fitting.

Coefficients are estimated via the closed‑form normal equations using matrix algebra. The tool also reports R² (coefficient of determination), adjusted R², SSE (residual sum of squares), and RMSE (root mean square error) — standard metrics for model comparison and goodness‑of‑fit assessment.

Step‑by‑Step Usage

  1. Enter your data points in the table. Each row requires an x and y value. Use Add row to increase the dataset; remove rows with the button.
  2. Alternatively, paste data from your spreadsheet into the text area and click Import — the table will be filled automatically.
  3. Select the model type (linear, quadratic, or cubic) that best matches your theoretical expectation.
  4. Choose the x‑axis for the residual plot: fitted values (default), predictor x, or observation index.
  5. Click Fit & Plot to compute the regression and display the results.
  6. Examine the two charts: the left panel shows the data with the fitted curve; the right panel is the residual plot. Look for random scatter or systematic patterns as described above.
  7. Use the preset examples to quickly explore different scenarios — from ideal linear relationships to problematic heteroscedastic data.

Real‑World Applications

Case Study: Supply Chain Demand Forecasting

A logistics company used a linear regression model to forecast weekly demand based on promotional spending. The initial model had an R² of 0.78, but the residual plot revealed a clear fan‑shaped pattern — residuals grew larger as fitted values increased. This indicated heteroscedasticity, meaning the model's prediction intervals were unreliable. After applying a log‑transformation to the response variable, the residual plot became homoscedastic, and the model's predictive performance improved significantly. The residual plot calculator would have flagged this issue immediately, saving the analyst from making flawed business decisions.

Case Study: Dose‑Response Curve in Pharmacology

Researchers studying a new drug measured response at various doses. A linear model seemed plausible, but the residual plot showed a distinct U‑shaped curve — a sign of non‑linearity. Switching to a quadratic model eliminated the pattern, increased R² from 0.72 to 0.91, and correctly captured the saturation effect. This example underscores how residual plots guide model selection and prevent underfitting.

Common Misconceptions

  • “High R² means a good model.” Not necessarily. A high R² can result from overfitting (especially with high‑degree polynomials) or from data that follow a curve but are forced into a linear model. Residual plots reveal whether the model is actually appropriate.
  • “Residuals should be normally distributed.” While normality is desirable for inference (confidence intervals, p‑values), the residual plot primarily checks for homoscedasticity and linearity. Normality can be assessed separately via a Q‑Q plot, but the residual plot is often more informative for model diagnostics.
  • “Outliers always ruin the model.” Outliers can be influential, but they also signal important data points worth investigating. A residual plot helps identify them; the analyst can then decide whether to remove, transform, or robustly handle them.
  • “Polynomial models always fit better.” Increasing polynomial degree reduces SSE but may overfit noise. Adjusted R² penalizes unnecessary complexity, and the residual plot helps detect overfitting — if the fitted curve oscillates wildly, the residuals will show a pattern.

Limitations & Responsible Use

While residual plots are powerful diagnostic tools, they are exploratory, not confirmatory. Visual patterns should be supplemented with formal statistical tests (e.g., Breusch-Pagan test for heteroscedasticity, Durbin-Watson test for autocorrelation) for rigorous inference. This calculator is optimized for educational exploration, preliminary data analysis, and small-to-medium sized datasets (typically < 500 points). For high-dimensional or heavily collinear data, consider regularization techniques (e.g., Ridge or Lasso) not implemented here.

The Mathematics Behind the Tool

Our calculator employs the method of least squares to estimate regression coefficients. For a polynomial of degree d, we construct the Vandermonde matrix X of size n × (d+1) where each row is [1, xi, xi², …, xid]. The normal equations are XTX β = XT y, solved via Gaussian elimination with partial pivoting for numerical stability. The solution vector β contains the coefficients [a, b, c, …].

From the fitted values ŷ = X β, we compute residuals e = y − ŷ. The total sum of squares SST = Σ (yi − ȳ)² and the residual sum of squares SSE = Σ ei² give R² = 1 − SSE/SST. Adjusted R² incorporates the number of parameters p = d+1: Adj.R² = 1 − (1−R²)·(n−1)/(n−p). RMSE = √(SSE/n) provides an absolute measure of fit.

All computations are performed in double‑precision floating‑point arithmetic, ensuring accuracy to ~15 decimal digits. The canvas rendering uses native 2D context with anti‑aliasing for crisp visual output.

Validation & Numerical Stability: The computational core has been cross-verified against industry-standard statistical libraries (R's lm() and Python's numpy.polyfit) to ensure numerical consistency up to machine precision (approx. 1e-12 relative error). The solver employs partial pivoting to minimize rounding errors, making it robust for well-conditioned real-world datasets.

Frequently Asked Questions

A scatter plot shows the raw data (x vs y). A residual plot displays the residuals (e = y − ŷ) against either the fitted values or the predictor. The residual plot is a diagnostic tool that reveals model deficiencies, while the scatter plot shows the original relationship.

At least 3 points are required for any model. For a quadratic fit, you need at least 4 points (since there are 3 parameters); for cubic, at least 5 points. More points generally yield more reliable estimates and better diagnostic power.

A perfect residual plot shows no discernible pattern: points are randomly scattered around the horizontal zero line, with roughly constant vertical spread across the entire range. This indicates that the model assumptions are satisfied.

Yes. Points with residuals that are much larger in absolute value than the bulk of the data (e.g., > 2–3 standard deviations) are potential outliers. The residual plot highlights these observations clearly.

Plotting residuals against fitted values is the standard diagnostic for homoscedasticity and linearity. It removes the effect of the predictor's scale and directly shows how residual behavior relates to the predicted response, which is what matters for model validation.

For time‑series, you should also check residual autocorrelation. Our residual plot with “observation index” on the x‑axis helps visualize serial dependence. However, for formal autocorrelation tests, use specialized time‑series software.

SSE (Sum of Squared Errors) is the total of squared residuals; it scales with sample size. RMSE (Root Mean Square Error) is √(SSE/n), giving an average error in the same units as the response variable. RMSE is more interpretable for model comparison.
References: Wikipedia: Residual plot ; Penn State STAT 501: Residual Analysis ; Draper, N.R. & Smith, H. "Applied Regression Analysis" (3rd ed.) ; NIST Engineering Statistics Handbook ; James, G., Witten, D., Hastie, T., & Tibshirani, R. "An Introduction to Statistical Learning" (Springer, 2021) — Chapter 3: Linear Regression.