Quick Answer: There Is No Universal Cutoff
- Universal good R² cutoff: None — context always governs
- R² = 0: The model accounts for none of the response variation relative to predicting the mean
- R² = 0.50: Roughly half the observed variation in Y is accounted for by the fitted model
- R² = 1: Perfect in-sample fit under standard OLS — investigate if unexpected
- Most important question: Is the model useful for its intended purpose?
- Also inspect: Residuals, RMSE, assumptions, adjusted R² for multiple regression, and out-of-sample performance
What R-Squared Actually Tells You
For standard least-squares regression with an intercept, R² is defined as:
SSE = Sum of Squared Errors (residuals from model)
SST = Total Sum of Squares (variation around the mean)
The interpretation: R² tells you the proportion of observed variation in the response variable that the fitted model accounts for, relative to the baseline of simply predicting the response mean. An R² of 0.60 means 60% of the observed variation in Y is accounted for by the fitted model — the remaining 40% is unexplained by the predictors in this model on this data.
Two things this does not mean. First, it is not a percentage accuracy score — the model is not "60% accurate" in any classification or precision sense. Second, the word "accounts for" is statistical, not causal. R² says nothing about whether X causes Y. For a deeper look at the formula and its derivation, the dedicated R-squared page covers the mechanics. For the coefficient-of-determination framing, see the coefficient of determination page.
In a model with one predictor and an intercept, R² = r², where r is the Pearson correlation coefficient. So r = 0.8 and r = −0.8 both produce R² = 0.64. R² loses direction — check the slope or correlation to know whether the relationship is positive or negative.
Why There Is No Single Good R-Squared Value
The reason there is no universal threshold comes down to one fact: different outcomes have different amounts of inherent, irreducible variation. Some things are easier to predict than others, and that difference has nothing to do with model quality.
A calibration experiment measuring instrument output against a known physical quantity might routinely produce R² above 0.99 — not because the analyst is especially skilled, but because the physical relationship is deterministic and measurement error is minimal. Running the same model on, say, student exam scores would be a completely different situation. Academic performance is shaped by preparation, sleep, prior knowledge, anxiety, test design, and dozens of other factors that no single predictor fully captures. Getting R² = 0.35 in that setting might reflect excellent work.
Comparing those two R² values — 0.99 versus 0.35 — and concluding the second model is worse is like comparing a clear photograph of a building to a slightly blurry photograph of fog, and concluding the second photographer needs a better camera. The subject explains the outcome, not the equipment.
You will sometimes see tables claiming: R² < 0.3 = poor, 0.3–0.7 = moderate, > 0.7 = good. No major statistical body endorses these cutoffs as universal standards. They may describe what is common in a particular field, but applying them across different domains produces systematically wrong conclusions.
How to Decide Whether Your R-Squared Is Good
Instead of comparing your R² to an arbitrary threshold, work through these six questions. They address the things that actually determine whether a model is fit for purpose.
What Is the Model Supposed to Do?
A model built to explain a relationship — to estimate a coefficient, understand a direction, or test a hypothesis — can be scientifically valuable even when R² is low, provided the coefficient estimate is precise and meaningful. A model built to predict new observations faces a stricter test: prediction error must be small enough to be useful. The same R² means different things in these two settings.
How Inherently Predictable Is the Outcome?
Outcomes driven by many unmeasured or random factors — human behavior, biological processes, social phenomena — have a practical upper ceiling on predictability. If the response variable itself contains large irreducible noise, even a well-specified model will have moderate R². That is not a flaw; it reflects the nature of the subject. Outcomes in tightly controlled physical or engineering contexts have much higher achievable R².
How Does R² Compare with Similar Models?
The most informative benchmark is not a threshold table — it is what other credible models achieve predicting the same outcome on similar data. If published work modeling weekly retail sales with comparable data typically shows R² in the 0.55–0.70 range, and your model is at 0.62, that is a reasonable result. If your model is at 0.30, that gap is worth investigating. Compare like with like: the same response variable, similar data quality, similar populations.
Do the Residuals Show Serious Problems?
A high R² does not validate a model. If the residuals show a clear curved pattern, the linear form is likely wrong regardless of R². If they fan out with increasing fitted values, the constant-variance assumption is violated. Residual diagnostics tell you things about model adequacy that R² cannot. Always inspect the residual plot before praising or dismissing an R² value.
Is the Prediction Error Acceptable in the Units of Y?
R² is unitless and relative — it does not tell you whether a typical prediction error of $200 or $20,000 is acceptable. RMSE (Root Mean Squared Error) expresses prediction error in the same units as the response. Two models can share the same R² = 0.80 while one has RMSE of $500 and another has RMSE of $50,000 — which is fine depends entirely on the decision context.
Does the Model Perform Well on New Data?
In-sample R² is computed on the data used to fit the model, so it can look strong simply because the model has been tuned to that particular dataset. When prediction is the goal, test-set or cross-validated R² is the honest measure. A model with training R² = 0.88 and test R² = 0.51 is almost certainly overfit and will not perform as hoped on real new observations.
What Specific R-Squared Values Mean
Is an R-Squared of 0.2 Good?
R² = 0.20 means the model accounts for about 20% of the observed variation in the response — 80% remains unexplained by the predictors in this model.
Imagine a study trying to predict employee productivity from weekly hours of training. With n = 400 employees, the regression produces R² = 0.19 and a slope of 2.3 additional units of output per training hour (p = 0.001, 95% CI: 1.1 to 3.5).
The low R² reflects something obvious in retrospect: productivity depends on many factors — motivation, prior experience, team dynamics, task complexity — that a single training variable cannot capture. But the coefficient is precisely estimated, statistically clear, and practically meaningful. The finding that each training hour associates with roughly 2.3 extra productivity units is a useful result for planning. The model answered the question it was built to answer.
✅ Verdict: Not bad — it serves its explanatory purpose. The R² tells you variation remains unexplained, not that the finding is wrong.
Is an R-Squared of 0.5 Good?
R² = 0.50 means the model accounts for roughly half the observed variation in Y. Whether that is impressive or underwhelming is entirely context-dependent. In macroeconomic forecasting, 0.50 might represent strong performance for a quarterly GDP growth model. In a calibration study for a precision instrument, 0.50 would indicate a serious problem with the measurement approach.
The number itself carries no quality verdict. Ask: what does credible work in this domain typically achieve? Is the prediction error within an acceptable range for the decision at hand? Are the residuals well-behaved?
Is an R-Squared of 0.7 Good?
R² = 0.70 is sometimes cited in informal guides as the threshold for a "good" model. This rule has no statistical basis. In a study predicting health outcomes from a handful of demographic variables, 0.70 might be genuinely strong. In a study predicting electricity demand from temperature and time-of-day data, 0.70 might suggest important predictors are missing. In a study of daily stock price changes, 0.70 would be extraordinarily suspicious — and likely evidence of data leakage rather than a genuinely predictive model.
No standard statistical reference endorses 0.7 as a universal threshold. Using it as one imposes an arbitrary standard that is simultaneously too demanding for noisy outcomes and too lenient for precise ones.
Is an R-Squared of 0.9 Good?
R² = 0.90 sounds impressive. In many engineering and physical science contexts, it is routine. In behavioral and social science, it would be unusually high and warrant scrutiny about model specification and data quality. An R² near 0.90 achieved with many predictors in a multiple regression warrants checking whether adjusted R² is similarly high — ordinary R² cannot decrease when predictors are added, so a high value may simply reflect a large model rather than genuine predictive ability.
A model fits a straight line to data generated by a quadratic function. With n = 60 observations in a moderate range, the linear fit achieves R² = 0.91 — but the residual plot shows a clear U-shaped curve.
The linear model misses the curve in the data. It has captured a large fraction of the variation simply because the overall trend is positive and the observations span a moderate range. Someone relying solely on R² = 0.91 would miss the problem entirely. The residual plot reveals it immediately. High R² can coexist with a misspecified model; the residuals are what diagnose it.
⚠️ Verdict: Impressive R² number, fundamentally flawed model. Residual diagnostics are not optional.
When a Low R-Squared Can Still Be Good
Several situations produce low R² values in models that are doing exactly what they should.
The most common: when the outcome is shaped by many unmeasured factors. A regression predicting individual test scores from study hours will have low R² because test performance also depends on baseline ability, prior coursework, sleep quality, test anxiety, and the specific questions asked. None of that makes study hours an unimportant predictor. A clear positive slope with a well-estimated confidence interval is a meaningful finding independent of R².
Another situation: measurement error in the response variable. If Y itself is measured with substantial noise — for example, self-reported dietary intake, where recall bias is large — there is a practical ceiling on achievable R². The noise floor in the data limits how much variation any model can explain. Low R² here reflects the measurement situation, not the underlying relationship.
A third: restricted range. If the sample was selected to be homogeneous on Y — say, only high-achieving students — R² will be artificially low even if the predictor has a real relationship with Y in the broader population. The restricted variation in Y means SST is small, which mechanically reduces R².
A predictor can have a clearly estimated, statistically significant, and practically important coefficient while R² remains low. Many real outcomes are inherently noisy. The coefficient and its confidence interval answer the explanatory question; R² describes how much variation those predictors capture in that sample.
When a High R-Squared Can Still Be Bad
Several situations produce high R² in models that should not be trusted.
Overfitting is the most common. When a model has many parameters relative to the sample size, it can trace the noise in the training data rather than the underlying signal. The result is high training R² and poor performance on new observations. If you add enough predictors to a multiple regression, R² approaches 1.0 regardless of whether those predictors are meaningful — ordinary R² cannot fall when predictors are added.
Data leakage is more subtle. If information about the response variable is inadvertently included in the predictors — a variable derived from Y, a future observation used to predict the past, or test data accidentally mixed into training — R² will be artificially inflated. The model appears much more predictive than it actually is.
Spurious correlation is a related problem, especially with time-series data where two unrelated variables both trend upward over time. The linear fit may explain 85% of the variation in Y, but that variation is being driven by common time trends, not a meaningful causal relationship. In these cases, R² is high but the model has no real predictive content beyond "things generally go up."
Finally, a high R² says nothing about whether the form of the model is correct. As the worked scenario above shows, a misspecified model can have high R² while systematically misfitting the data in ways the residuals reveal immediately. For a discussion of how influential observations can pull a line toward high R² in misleading ways, see the influential points page.
High R² does not prove linearity, constant variance, independent errors, correct model specification, or causal relationships. It also does not guarantee good predictions on new data. Always inspect residuals, check RMSE, and validate on held-out data when prediction matters.
R-Squared vs RMSE
R² and RMSE answer different questions. R² is a relative, unitless measure of fit. RMSE expresses the typical scale of prediction errors in the same units as the response variable.
| Feature | R-Squared | RMSE |
|---|---|---|
| What it measures | Variation accounted for (relative) | Typical prediction error magnitude |
| Units | Unitless (0 to 1 in standard OLS) | Same units as Y |
| Reference point | Compared against the mean baseline | Compared against zero error |
| Higher means | More variation accounted for | Larger typical errors |
| Tells you error size? | No | Yes, directly |
| Useful for | Comparing fit vs baseline; model comparison | Practical error assessment; decision relevance |
Consider two models, both with R² = 0.78. Model A predicts house prices with RMSE of $8,000. Model B predicts house prices in a different city with RMSE of $95,000. The same R² represents very different practical situations. R² alone does not tell you which model is more practically useful — RMSE does.
R-Squared vs Adjusted R-Squared
Ordinary R² has a problem in multiple linear regression: adding any predictor to the model — including noise variables that have no real relationship with Y — can only keep R² the same or increase it. This means you can inflate R² simply by adding more predictors. A model with 20 predictors will almost always show higher R² than a model with 5 predictors on the same data, regardless of whether those extra 15 predictors contribute anything meaningful.
Adjusted R² corrects for this by penalizing model complexity:
n = number of observations
k = number of predictors (excluding intercept)
When a new predictor genuinely improves the model, adjusted R² increases. When the predictor adds little explanatory power, adjusted R² stays flat or decreases. This makes adjusted R² more useful than ordinary R² for comparing models with different numbers of predictors — for example, deciding whether adding two more variables to a regression is worth the added complexity.
| Feature | Ordinary R² | Adjusted R² |
|---|---|---|
| When predictors are added | Can only stay same or increase | Can decrease if predictor adds little |
| Accounts for model size | No | Yes |
| Best use | Describing in-sample fit | Comparing models with different numbers of predictors |
| Can reward over-modeling? | Yes | Partially corrects for this |
Adjusted R² penalizes unnecessary predictors but does not replace cross-validation, held-out evaluation, residual diagnostics, or theoretical justification for including predictors. It is one useful tool in model selection, not the final word. For worked examples, see multiple regression examples.
Training R-Squared vs Test R-Squared
In-sample (training) R² measures how well the model fits the data it was built on. This is what most software reports by default. But if prediction is the goal, the question that actually matters is: how well does the model perform on new observations it has never seen?
Test-set R² — computed by applying the model to held-out data and calculating the same formula — answers that question. A model can have high training R² and low test R², and this gap is the signature of overfitting. The model has learned the noise in the training data rather than the underlying signal.
Model A uses 3 meaningful predictors: training R² = 0.72, test R² = 0.68. Model B adds 15 more variables: training R² = 0.89, test R² = 0.44.
Model B looks much better by the training metric. But its test performance is worse than Model A — it has fitted the training noise. Selecting Model B on the basis of higher training R² would be a mistake. The 17-point gap between training R² and test R² (0.89 vs 0.44) is a clear overfitting signal.
⚠️ Verdict: Never select a predictive model on training R² alone. Validate on held-out data.
Can R-Squared Be Negative?
Standard in-sample OLS regression with an intercept always produces R² between 0 and 1. This is guaranteed by the mathematics: with an intercept, the model always fits at least as well as predicting the mean of Y.
However, when R² is computed on a held-out test set, it can be negative. This happens when the model's predictions are worse than simply predicting the mean of the test data for every observation. A negative test-set R² is a clear indication that the model has failed to generalize — it may be overfit, or the test data may come from a different distribution than the training data.
Negative R² can also arise in no-intercept (forced-through-origin) models, where some software uses an alternative "uncentered" formulation that does not use the same SST reference. Be careful when comparing R² values from models with and without intercepts.
This is only true for standard in-sample OLS with an intercept. Test-set R² and some alternative formulations can take negative values. Understand which R² your software is reporting before drawing conclusions.
R-Squared Does Not Prove Causation
A high R² says the predictors account for a large fraction of the observed variation in Y. It says nothing about whether that relationship is causal. Confounding variables, common trends, reverse causality, and selection effects can all produce high R² in the absence of any direct causal connection.
Two unrelated time series both trending upward will produce high R² in a simple regression. A predictor that shares a common cause with Y will show a strong association without directly affecting Y. R² measures the correlation structure in the data; establishing causality requires appropriate study design, theory, and in most cases techniques that go well beyond fitting a regression line. The correlation vs causation page covers this distinction in depth.
R-Squared Calculator
Calculate R² from Your Data
Enter paired X and Y values (comma-separated) to compute R², the slope, and RMSE.
For more advanced regression calculations, use the Simple Linear Regression Calculator. To explore interactively how changing data spread affects R², the R-Squared Visualizer demonstrates this clearly.
Model Evaluation Checklist
Before deciding an R² value is "good enough," work through this checklist. R² is one input, not the final verdict.
- Is the model form appropriate for this data? (Check residual plot for curves or patterns)
- Are the residuals randomly scattered, or do they show systematic problems?
- Are there influential points pulling R² upward artificially?
- Is RMSE / prediction error within an acceptable range for the real decision?
- For prediction tasks: does the model perform well on held-out data?
- Is R² competitive with what credible models achieve for this outcome?
- For multiple regression: does adjusted R² support the same conclusion as ordinary R²?
- Are coefficient estimates meaningful and within plausible ranges?
- Does the model answer the actual scientific or business question?
- Are assumptions (independence, constant variance, linearity) adequately met?
Practical Judgment at Common R² Values
| Situation | Appropriate Judgment |
|---|---|
| Low R², meaningful slope, noisy human outcome | May still support a useful explanatory finding |
| High R², curved residual pattern | Model form is likely wrong — check for nonlinearity |
| High training R², much lower test R² | Likely overfitting — simplify the model or get more data |
| Moderate R², small prediction error relative to decision needs | Can be practically strong for the task |
| High R², large unacceptable prediction error | Numerically impressive but practically inadequate |
| Slight R² gain from many additional predictors | Check adjusted R² and validation — gain may be spurious |
| R² near 1.0 unexpectedly | Investigate for data leakage, duplicated variables, or overfit |
| Negative test-set R² | Model worse than the mean baseline on new data — poor generalization |
Common R-Squared Mistakes
| The Mistake | Why It's Wrong | The Correction |
|---|---|---|
| "R² above 0.7 is always good" | No universal threshold exists in statistical literature | Compare against what is typical for your outcome and domain |
| "R² = 0.80 means the model is 80% accurate" | R² measures variation accounted for, not prediction accuracy | Use RMSE to understand error magnitude in the units of Y |
| "Low R² means no relationship exists" | A meaningful slope can coexist with large unexplained variation | Look at the coefficient estimate and its confidence interval |
| "High R² means the model is correctly specified" | Misspecified models can have high R² | Always inspect residual plots for patterns or curvature |
| "High R² proves X causes Y" | R² measures fit — it contains no causal information | Causality requires appropriate study design and theory |
| "Choose the model with the highest R²" | Ordinary R² can only increase with more predictors | Use adjusted R², validation error, and parsimony together |
| "Training R² measures prediction quality" | In-sample fit can be inflated by overfitting | Evaluate performance on held-out test data |
| "R² can never be negative" | Only true for in-sample OLS with an intercept | Test-set R² and no-intercept models can produce negative values |
| "Remove outliers to improve R²" | Removing legitimate data to raise R² corrupts the analysis | Investigate influential points on their own merits |
| "Add variables to improve R²" | Ordinary R² mechanically increases with any new predictor | Add predictors only when theoretically justified; use adjusted R² to evaluate |
Frequently Asked Questions
There is no universally good R² value. Whether R² is sufficient depends on what is being predicted, how inherently variable that outcome is, what comparable models achieve in the same domain, whether the residuals show problems, and whether prediction error is acceptable in real units. A value of 0.30 may be strong in behavioral science; the same value may be inadequate in a tightly controlled physical experiment.
R² = 0.50 means the model accounts for roughly half the observed variation in the response. Whether that is adequate depends on the context: 0.50 might be competitive in macroeconomic modeling or behavioral research, while it would be insufficient in a precision engineering application. Check the residuals, RMSE, and what comparable models achieve before drawing a conclusion.
Same principle applies. R² = 0.60 means 60% of the variation in the response is accounted for by the model. In many social science and business contexts, this would be considered a reasonable result. In other contexts, it could suggest the model is missing important predictors. The number alone does not decide the question — what matters is whether the model is fit for its intended purpose.
Informally, 0.7 is often cited as a "good" threshold, but this has no statistical basis. In some fields it would be strong; in others, underwhelming or suspicious. An R² of 0.7 with well-behaved residuals and acceptable prediction error is more meaningful than one with curved residuals or poor test-set performance. Always examine what the number means in your specific context.
R² = 0.90 is impressive in many settings and routine in others. In physical calibration or engineering it may be expected; in behavioral science it would be unusually high and worth investigating. A high R² achieved through many predictors in multiple regression warrants checking adjusted R² and validation performance. Near-perfect R² should prompt a search for data leakage or model misspecification.
No. R² is a relative measure of fit — it compares the model to a simple mean baseline. A model with R² = 0.85 can still produce large prediction errors in absolute terms if the response variable has high variance. RMSE tells you the actual scale of errors in the units of Y, which is what matters for most practical decisions. The two metrics answer different questions and both are needed.
Standard in-sample OLS with an intercept always produces R² between 0 and 1. Test-set R² — computed on held-out data — can be negative when the model performs worse than predicting the mean for every observation. This signals overfitting or distribution shift. No-intercept models can also produce alternative R² formulations that take negative values. Always understand which formulation your software is using.
R² = 1 means all observed response values fall exactly on the fitted model. In a deterministic physical relationship (Hooke's Law, for example), this can be legitimate. In most empirical work, it warrants investigation: possible causes include data leakage (the target variable appears in the predictors), a constructed relationship, duplicated columns, a very small sample with as many parameters as observations, or a coding error. Do not celebrate R² = 1 automatically.
Adjusted R² penalizes model complexity. Ordinary R² can only increase when predictors are added; adjusted R² can decrease if a new predictor adds little explanatory value. Use adjusted R² when comparing multiple regression models with different numbers of predictors fitted to the same response on the same data. It is a more honest comparison than ordinary R² in that context. It does not replace validation, residual analysis, or theoretical justification for predictors.
No. Maximizing R² is not a scientific objective. It can lead to overfitting, inclusion of irrelevant predictors, and models that look good in-sample but generalize poorly. Improve the model when changes are justified by theory, better measurement, appropriate functional form, or validated predictive performance — not because they raise R².
Key Takeaways
There is no universal R-squared value that makes a regression model "good." Judge R² in the context of the outcome, the purpose of the model, comparable work in your field, residual diagnostics, practical prediction error, model complexity, and performance on new data. A low R² can still support a useful finding; a high R² can still belong to a poorly specified or overfit model.
- R² = proportion of observed variation in Y accounted for by the fitted model — not percentage accuracy
- No universal threshold (0.7, 0.8, etc.) applies across all domains or model types
- Low R² does not mean no relationship; a coefficient can be meaningful with substantial unexplained variation
- High R² does not validate model form, linearity, or assumptions — residual plots do
- High R² does not imply causation — fit and causality are separate questions
- For multiple regression, use adjusted R² when comparing models with different numbers of predictors
- For prediction, validate on held-out data; test-set R² is the honest measure of generalization
- Use RMSE alongside R² to understand the practical scale of prediction errors