How and When to Use a Calibrated Classification Model with scikit-learn

Last Updated on September 25, 2019

Instead of predicting class values directly for a classification problem, it can be convenient to predict the probability of an observation belonging to each possible class.

Predicting probabilities allows some flexibility including deciding how to interpret the probabilities, presenting predictions with uncertainty, and providing more nuanced ways to evaluate the skill of the model.

Predicted probabilities that match the expected distribution of probabilities for each class are referred to as calibrated. The problem is, not all machine learning models are capable of predicting calibrated probabilities.

There are methods to both diagnose how calibrated predicted probabilities are and to better calibrate the predicted probabilities with the observed distribution of each class. Often, this can lead to better quality predictions, depending on how the skill of the model is evaluated.

In this tutorial, you will discover the importance of calibrating predicted probabilities and how to diagnose and improve the calibration of models used for probabilistic classification.

After completing this tutorial, you will know:

  • Nonlinear machine learning algorithms often predict uncalibrated class probabilities.
  • Reliability diagrams can be used to diagnose the calibration of a model, and methods can be used to better calibrate predictions for a problem.
  • How to develop reliability diagrams and calibrate classification models in Python with scikit-learn.

Kick-start your project with my new book Probability for Machine Learning, including step-by-step tutorials and the Python source code files for all examples.

Let’s get started.

How and When to Use a Calibrated Classification Model with scikit-learn

How and When to Use a Calibrated Classification Model with scikit-learn
Photo by Nigel Howe, some rights reserved.

Tutorial Overview

This tutorial is divided into four parts; they are:

  1. Predicting Probabilities
  2. Calibration of Predictions
  3. How to Calibrate Probabilities in Python
  4. Worked Example of Calibrating SVM Probabilities

Predicting Probabilities

A classification predictive modeling problem requires predicting or forecasting a label for a given observation.

An alternative to predicting the label directly, a model may predict the probability of an observation belonging to each possible class label.

This provides some flexibility both in the way predictions are interpreted and presented (choice of threshold and prediction uncertainty) and in the way the model is evaluated.

Although a model may be able to predict probabilities, the distribution and behavior of the probabilities may not match the expected distribution of observed probabilities in the training data.

This is especially common with complex nonlinear machine learning algorithms that do not directly make probabilistic predictions and instead use approximations.

The distribution of the probabilities can be adjusted to better match the expected distribution observed in the data. This adjustment is referred to as calibration, as in the calibration of the model or the calibration of the distribution of class probabilities.

[…] we desire that the estimated class probabilities are reflective of the true underlying probability of the sample. That is, the predicted class probability (or probability-like value) needs to be well-calibrated. To be well-calibrated, the probabilities must effectively reflect the true likelihood of the event of interest.

— Page 249, Applied Predictive Modeling, 2013.

Calibration of Predictions

There are two concerns in calibrating probabilities; they are diagnosing the calibration of predicted probabilities and the calibration process itself.

Reliability Diagrams (Calibration Curves)

A reliability diagram is a line plot of the relative frequency of what was observed (y-axis) versus the predicted probability frequency  (x-axis).

Reliability diagrams are common aids for illustrating the properties of probabilistic forecast systems. They consist of a plot of the observed relative frequency against the predicted probability, providing a quick visual intercomparison when tuning probabilistic forecast systems, as well as documenting the performance of the final product

Increasing the Reliability of Reliability Diagrams, 2007.

Specifically, the predicted probabilities are divided up into a fixed number of buckets along the x-axis. The number of events (class=1) are then counted for each bin (e.g. the relative observed frequency). Finally, the counts are normalized. The results are then plotted as a line plot.

These plots are commonly referred to as ‘reliability‘ diagrams in forecast literature, although may also be called ‘calibration‘ plots or curves as they summarize how well the forecast probabilities are calibrated.

The better calibrated or more reliable a forecast, the closer the points will appear along the main diagonal from the bottom left to the top right of the plot.

The position of the points or the curve relative to the diagonal can help to interpret the probabilities; for example:

  • Below the diagonal: The model has over-forecast; the probabilities are too large.
  • Above the diagonal: The model has under-forecast; the probabilities are too small.

Probabilities, by definition, are continuous, so we expect some separation from the line, often shown as an S-shaped curve showing pessimistic tendencies over-forecasting low probabilities and under-forecasting high probabilities.

Reliability diagrams provide a diagnostic to check whether the forecast value Xi is reliable. Roughly speaking, a probability forecast is reliable if the event actually happens with an observed relative frequency consistent with the forecast value.

Increasing the Reliability of Reliability Diagrams, 2007.

The reliability diagram can help to understand the relative calibration of the forecasts from different predictive models.

Want to Learn Probability for Machine Learning

Take my free 7-day email crash course now (with sample code).

Click to sign-up and also get a free PDF Ebook version of the course.

Download Your FREE Mini-Course

Probability Calibration

The predictions made by a predictive model can be calibrated.

Calibrated predictions may (or may not) result in an improved calibration on a reliability diagram.

Some algorithms are fit in such a way that their predicted probabilities are already calibrated. Without going into details why, logistic regression is one such example.

Other algorithms do not directly produce predictions of probabilities, and instead a prediction of probabilities must be approximated. Some examples include neural networks, support vector machines, and decision trees.

The predicted probabilities from these methods will likely be uncalibrated and may benefit from being modified via calibration.

Calibration of prediction probabilities is a rescaling operation that is applied after the predictions have been made by a predictive model.

There are two popular approaches to calibrating probabilities; they are the Platt Scaling and Isotonic Regression.

Platt Scaling is simpler and is suitable for reliability diagrams with the S-shape. Isotonic Regression is more complex, requires a lot more data (otherwise it may overfit), but can support reliability diagrams with different shapes (is nonparametric).

Platt Scaling is most effective when the distortion in the predicted probabilities is sigmoid-shaped. Isotonic Regression is a more powerful calibration method that can correct any monotonic distortion. Unfortunately, this extra power comes at a price. A learning curve analysis shows that Isotonic Regression is more prone to overfitting, and thus performs worse than Platt Scaling, when data is scarce.

Predicting Good Probabilities With Supervised Learning, 2005.

Note, and this is really important: better calibrated probabilities may or may not lead to better class-based or probability-based predictions. It really depends on the specific metric used to evaluate predictions.

In fact, some empirical results suggest that the algorithms that can benefit the more from calibrating predicted probabilities include SVMs, bagged decision trees, and random forests.

[…] after calibration the best methods are boosted trees, random forests and SVMs.

Predicting Good Probabilities With Supervised Learning, 2005.

How to Calibrate Probabilities in Python

The scikit-learn machine learning library allows you to both diagnose the probability calibration of a classifier and calibrate a classifier that can predict probabilities.

Diagnose Calibration

You can diagnose the calibration of a classifier by creating a reliability diagram of the actual probabilities versus the predicted probabilities on a test set.

In scikit-learn, this is called a calibration curve.

This can be implemented by first calculating the calibration_curve() function. This function takes the true class values for a dataset and the predicted probabilities for the main class (class=1). The function returns the true probabilities for each bin and the predicted probabilities for each bin. The number of bins can be specified via the n_bins argument and default to 5.

For example, below is a code snippet showing the API usage:

Calibrate Classifier

A classifier can be calibrated in scikit-learn using the CalibratedClassifierCV class.

There are two ways to use this class: prefit and cross-validation.

You can fit a model on a training dataset and calibrate this prefit model using a hold out validation dataset.

For example, below is a code snippet showing the API usage:

Alternately, the CalibratedClassifierCV can fit multiple copies of the model using k-fold cross-validation and calibrate the probabilities predicted by these models using the hold out set. Predictions are made using each of the trained models.

For example, below is a code snippet showing the API usage:

The CalibratedClassifierCV class supports two types of probability calibration; specifically, the parametric ‘sigmoid‘ method (Platt’s method) and the nonparametric ‘isotonic‘ method which can be specified via the ‘method‘ argument.

Worked Example of Calibrating SVM Probabilities

We can make the discussion of calibration concrete with some worked examples.

In these examples, we will fit a support vector machine (SVM) to a noisy binary classification problem and use the model to predict probabilities, then review the calibration using a reliability diagram and calibrate the classifier and review the result.

SVM is a good candidate model to calibrate because it does not natively predict probabilities, meaning the probabilities are often uncalibrated.

A note on SVM: probabilities can be predicted by calling the decision_function() function on the fit model instead of the usual predict_proba() function. The probabilities are not normalized, but can be normalized when calling the calibration_curve() function by setting the ‘normalize‘ argument to ‘True‘.

The example below fits an SVM model on the test problem, predicted probabilities, and plots the calibration of the probabilities as a reliability diagram,

Running the example creates a reliability diagram showing the calibration of the SVMs predicted probabilities (solid line) compared to a perfectly calibrated model along the diagonal of the plot (dashed line.)

We can see the expected S-shaped curve of a conservative forecast.

Uncalibrated SVM Reliability Diagram

Uncalibrated SVM Reliability Diagram

We can update the example to fit the SVM via the CalibratedClassifierCV class using 5-fold cross-validation, using the holdout sets to calibrate the predicted probabilities.

The complete example is listed below.

Running the example creates a reliability diagram for the calibrated probabilities.

The shape of the calibrated probabilities is different, hugging the diagonal line much better, although still under-forecasting in the upper quadrant.

Visually, the plot suggests a better calibrated model.

Calibrated SVM Reliability Diagram

Calibrated SVM Reliability Diagram

We can make the contrast between the two models more obvious by including both reliability diagrams on the same plot.

The complete example is listed below.

Running the example creates a single reliability diagram showing both the calibrated (orange) and uncalibrated (blue) probabilities.

It is not really an apples-to-apples comparison as the predictions made by the calibrated model are in fact a combination of five submodels.

Nevertheless, we do see a marked difference in the reliability of the calibrated probabilities (very likely caused by the calibration process).

Calibrated and Uncalibrated SVM Reliability Diagram

Calibrated and Uncalibrated SVM Reliability Diagram

Further Reading

This section provides more resources on the topic if you are looking to go deeper.

Books and Papers




In this tutorial, you discovered the importance of calibrating predicted probabilities and how to diagnose and improve the calibration of models used for probabilistic classification.

Specifically, you learned:

  • Nonlinear machine learning algorithms often predict uncalibrated class probabilities.
  • Reliability diagrams can be used to diagnose the calibration of a model, and methods can be used to better calibrate predictions for a problem.
  • How to develop reliability diagrams and calibrate classification models in Python with scikit-learn.

Do you have any questions?
Ask your questions in the comments below and I will do my best to answer.

Get a Handle on Probability for Machine Learning!

Probability for Machine Learning

Develop Your Understanding of Probability

...with just a few lines of python code

Discover how in my new Ebook:
Probability for Machine Learning

It provides self-study tutorials and end-to-end projects on:
Bayes Theorem, Bayesian Optimization, Distributions, Maximum Likelihood, Cross-Entropy, Calibrating Models
and much more...

Finally Harness Uncertainty in Your Projects

Skip the Academics. Just Results.

See What's Inside

76 Responses to How and When to Use a Calibrated Classification Model with scikit-learn

  1. Elie Kawerk September 3, 2018 at 11:34 pm #

    I wish you include the last 2 tutorials in “Machine learning mastery with Python”.

    • Jason Brownlee September 4, 2018 at 6:06 am #

      Thanks. They may be a little advanced for that beginner book.

  2. Mark Littlewood September 4, 2018 at 1:30 am #

    Thanks for this, very interesting but on my Gradient Descent Boosting UK horse racing ratings it did not improve performance sadly

    • Jason Brownlee September 4, 2018 at 6:10 am #

      Thanks, nice one for trying!

      • Mark Littlewood March 18, 2020 at 10:15 pm #

        Correction actually I was taking the code for creating the calibration plot from another example and I was setting normalize to True. For horse racing you really want to take the probabilities as is and when I did this the probabilities were pretty well calibrated without adjustment

      • Mark Littlewood March 18, 2020 at 10:17 pm #

        Is there a Python Sklearn function the measures the fit of the calibration curve ie a kind of MSE on the curve

  3. Ratiba September 7, 2018 at 6:47 am #

    Thanks, very useful.

  4. Raj September 10, 2018 at 2:40 pm #

    In reliability plot, how to plot observed values on y-axis bins. For eg: what can be possible value for a 0.2 bin on y-axis. May be a detailed illustration of the reliability plot will be helpful. Thank you.

    • Jason Brownlee September 11, 2018 at 6:25 am #

      Good question, I would expect that the calculated bins will be accessible some how, I don’t have an example.

  5. Joe F October 31, 2018 at 8:53 pm #

    This is a great article – could you explain a but more about how to calibrate with an oversampled dataset (which will have produced incorrect probabilities) – should you adjust the probabilities to account for the oversampling yourself, before calibration, or after, or should you just use the weights parameter within the fit() function ? Any thoughts?

    • Jason Brownlee November 1, 2018 at 6:06 am #

      You could fit with an oversampled dataset, but calibrate with an un-touched validation dataset.

  6. M K November 9, 2018 at 6:32 am #

    Yet another cool contribution of yours to ML community, as usual!

    Could you please indicate how this approach may be used with a random forest as a classifier — as RF doesn’t output a decision function as SVC.

    Many thanks.

    • Jason Brownlee November 9, 2018 at 1:58 pm #

      You can use the calibration method directly.

      Perhaps I don’t understand the problem you’re having?

  7. M K November 9, 2018 at 6:48 am #

    Hi, again.

    Please disregard my previous question — it seems that returning predict_proba(testX)[:, 1] in the function uncalibrated (when using a random forest) resolves my issue.



  8. Jen K. March 22, 2019 at 6:07 am #

    Hi Jason,

    Would it be possible to use CalibratedClassifierCV to calibrate a non scikit learn model? For example, how would I go about calibrating a classifier written in tensorflow?


    • Jason Brownlee March 22, 2019 at 8:43 am #

      Perhaps, but you may need to wrap the model to look like a sklearn classifier (like we can do in Keras).

  9. Andres M. March 29, 2019 at 6:45 am #

    Hi Jason , thanks for making such detailed explanation! I’ve been reading about these issues as I am trying to look for the best way of comparing different classifiers performances, and it is very common to use the ROC curve and the AUC. However, as you posted here, “not all machine learning models are capable of predicting calibrated probabilities”, so i guess it only make sense to speak about probabilities and thresholds for some classifiers only. But most of the models in sklearn either have the .predict_prob() method or the .decision_function() one, so i would like to know which specific models are we talking about here. I don’t see what it means to choose the threshole for a model that doesn’t return a decision function (like in a decision tree for example), but it does have a predict_prob() method nevertheless.

    In simple words, which classifiers can I compare using the ROC and callibrate the probabilities?

    • Jason Brownlee March 29, 2019 at 8:49 am #

      Simple models are built on a probabilistic model, such as logistic regression, whereas many nonlinear methods don’t operate that way.

      One approach might be to treat models separately with and without calibration, throw them all in the mix and see what performs best in your comparison. The calibration won’t hurt probabilistic models like logistic regression and will help non-probabilistic models.

  10. cs student April 21, 2019 at 5:19 pm #

    Could you please elaborate more on how can this be extended to multi-class problems?

    • Jason Brownlee April 22, 2019 at 6:17 am #

      Is there a specific aspect that you’re having trouble with?

      • Ahmad December 12, 2019 at 3:34 am #

        Hi Janson, when I try to run on multi-class, the calibration_curve is throwing error as “bad input shape.

        “fop_uncalibrated, mpv_uncalibrated = calibration_curve(testy, yhat_uncalibrated, n_bins=10, normalize=True)”

        testy size is [59, 8], np_utils categorical data.
        yhat_uncalibrated size is [59, 8], which is “model.decision_function(testX)” output.

        • Jason Brownlee December 12, 2019 at 6:29 am #

          Yes, the implementation is for binary classification only as far as I recall.

          • Ahmad December 12, 2019 at 4:58 pm #

            Okay, thank you for the information. If I want to implement it for multi-class how can I do that. Can you suggest a way

          • Jason Brownlee December 13, 2019 at 5:54 am #

            Sorry, I don’t have an example. I cannot give you off the cuff advice.

  11. Blanca September 4, 2019 at 12:55 am #

    Thanks so much for this post, it’s been extremely useful!
    I’ve been running a random forest and applying the calibration as you suggest here, but now the size of the dataset is increasing and I might need to shift to Spark. Do you know if there is a way of doing the calibration on Spark?

  12. Nemo October 3, 2019 at 4:40 am #


    Can calibration change relative position of predictions for some objects? I mean, if predicted probability for object1 is greater than that for object2 before calibration, but it is contrariwise after.

    If not, does it mean that calibration doesn’t affect ROC-AUC score?

    Thank you!

    • Jason Brownlee October 3, 2019 at 6:54 am #

      It often does not, but it may, e.g. for probabilities around the threshold.

      I would expect the ROC to change after calibration!

  13. Samarth Dhawan October 6, 2019 at 5:33 pm #

    Hi, How can I do a gridseach along with CaliberatedClassifierCV to tune my parameters?

    • Jason Brownlee October 7, 2019 at 8:28 am #

      I’m not sure, perhaps try using the CalibrateClassifier without the CV as part of a pipeline inside a gridsearchcv?

  14. Mohammad October 22, 2019 at 12:39 am #

    Awesome 🙂

  15. Marina Drus November 2, 2019 at 5:58 am #

    Very well explained! Thank you!

  16. Elias December 11, 2019 at 12:31 am #

    Is it really necessary to use CalibratedClassifierCV within a SVC?

    I think you get a 5 fold CV calibration using sv = SVC(probabilities=True). Right?

    • Jason Brownlee December 11, 2019 at 7:00 am #

      It will use cv to estimate the corrected probabilities, not to evaluate the model.

      • Elias December 11, 2019 at 9:22 pm #

        Thank you!

        So would you recommend using CalibratedClassifierCV instead or does it depend on the use casw?

  17. Ahmad December 12, 2019 at 3:30 am #

    Hi Jason,
    Can you please explain me about generating calibration_curve on multi class classifier probabilities.

    • Jason Brownlee December 12, 2019 at 6:29 am #

      Sorry, I don’t have an example. It might not make sense, I have only seen it for binary classes (I think).

  18. Abhilash Menon January 8, 2020 at 2:15 am #

    Hi Jason, Thanks for all your help! I am working on a pet project of mine and I have been ranking instances in the positive class in the decreasing order of their probabilities generated by XGBoost. From your article, I can see that calibration is basically rescaling. This was my understanding as well. However, I find that the order of candidates is changing as well once I choose to sort based on calibrated class probabilities. For example:: 0.94 became 0.91 and 0.85 became 0.95 after calibration. This changes the order of instances (ranking). In some cases, the probabilities dropped by 0.2 which is a margin.
    If the underlying mechanism for calibration is Logistic regression on probs generated by the model to predict the actual outcome why does it not scale and cause this change in order? Please correct me if I have made any wrong assumptions here. Thanks!

    Note: the calibration method I used was sigmoid.

    • Jason Brownlee January 8, 2020 at 8:28 am #

      Yes, think of it as a non-linear mapping from uncalibrated probs to calibrated probs.

      E.g. one probability distribution function to another.

    • Pepe September 5, 2020 at 5:51 am #

      Hey man, same doubt.

      Take a look here:

      CalibratedClassifierCV uses a cross-validation approach to fit both the classifier and the regressor. For each of the k (trainset, testset) couple, a classifier is trained on the train set, and its predictions on the test set are used to fit a regressor. We end up with k (classifier, regressor) couples where each regressor maps the output of its corresponding classifier into [0, 1]. Each couple is exposed in the calibrated_classifiers_ attribute, where each entry is a calibrated classifier with a predict_proba method that outputs calibrated probabilities.

      ->>>The output of predict_proba for the main CalibratedClassifierCV instance corresponds to the AVERAGE of the predicted probabilities of the k estimators in the calibrated_classifiers_ list.

      The output of predict is the class that has the highest probability.

      • Jason Brownlee September 5, 2020 at 6:54 am #

        Thanks for sharing.

      • John September 5, 2020 at 7:55 am #

        mmmm it should not be the case as you are averaging monotonically functions. It may occur because some observations, which were close (e.g. 0.0021 and 0.002), may get “squished” together, being rounded. The ranking, the AUC, should be almost the same

  19. Vera February 19, 2020 at 8:55 pm #

    Hi Jason!

    I’m a bit confused and maybe you can help me.
    When I use a different strategy for calibration, the predictions of the model change. I was trying different methods (platt, isotonic, prefit, cv) just to get an idea of what is going on. But I found my predictions change everytime. I am working with text data, classifying 25 classes.


    • Jason Brownlee February 20, 2020 at 6:11 am #

      Yes. Perhaps pick a metric and optimize it via CV or a hold out validation set.

  20. Gledson February 20, 2020 at 1:32 am #

    Hello, I would like to know if it is possible to perform the calibration with more than two classes? If so, how to do it?


    • Jason Brownlee February 20, 2020 at 6:16 am #

      It may be possible. I don’t have a tutorial on the topic, sorry.

  21. Riz March 13, 2020 at 7:35 am #

    How do we finalize the model? Usually, after we are happy with the a model, we fit it on all of the data. Is that the same with calibration? Thanks!

    • Jason Brownlee March 13, 2020 at 8:23 am #

      Yes, but you will need a validation set or a CV procedure for the final model to calibrate.

  22. Riz March 13, 2020 at 9:02 am #

    So if I understand correctly, using the example below, the model is finalized in step 3 and then can be used to predict new data (step 4)?

    1) model=RandomForestClassifier()
    2) calibrated=CalibratedClassifierCV(model, method=’sigmoid’, cv=5)
    4)calibrated_probs = calibrated.predict_proba(X)[:,1]

  23. riz April 25, 2020 at 1:59 pm #

    Is saving/loading the calibratedclassifiercv the same as any other model? I have tried saving and loading the calibrated and running predictions, but they end up being way off from when the model is originally created, thanks!

    • Jason Brownlee April 26, 2020 at 6:03 am #

      Yes, I believe so.

      Sorry to hear that you are having trouble. Perhaps try posting your code and error to stackoverflow?

  24. Som Dubey May 27, 2020 at 4:57 pm #

    This is great for reading and can be useful for predictions in classifier. I believe i will now always would like to do this exercise to see how my predictions are compared to calibrated ones. Thanks a lot sir.

  25. mathswoman July 3, 2020 at 7:46 pm #

    sorry i don’t get what you mean by : Below the diagonal: The model has over-forecast; the probabilities are too large.
    Above the diagonal: The model has under-forecast; the probabilities are too small. surely if the curve is below the line then we are under foreceasting?? e..g given probability of 0.5, a perfect classifier will predict 0.5 of population will be posotives. however if our model is saying given probabliluty 0.5 the proportion is actually 0.4 then we are underestimating the positives?

  26. mathswoman July 10, 2020 at 4:40 pm #

    ok, say i fit a classifier using my training data (from grid search e.g. xgboost). i then run calibration calling ‘prefit’ variable, and this fit this calibrated classifier on training data – why can i not do this? — apparently you cannot

    • Jason Brownlee July 11, 2020 at 6:02 am #

      Why can’t you do it? I have not tried, I’m curious.

      Perhaps just fit the model again with the chosen hyperparameters and calibration?

  27. mathswoman July 13, 2020 at 8:50 pm #

    because accoridng to the documentation the data must be disjoint from the one used for training.

    • Jason Brownlee July 14, 2020 at 6:19 am #

      Yes, perhaps prepare a hold out dataset, or split some data off from train – e.g. manually.

  28. Maryam August 22, 2020 at 2:42 am #

    Dear Jason,
    Thank you for the useful tutorial.
    I would like to know what the difference is between “confidence calibration” and “joint calibration”?

    would you mind please editing this tutorial to add these definitions?

    Thank you in advance to consider our requirements.


    • Jason Brownlee August 22, 2020 at 6:19 am #

      Thanks for the suggestion.

      Where did you hear these terms?

  29. Maryam August 29, 2020 at 6:36 am #

    Dear Jason,
    I am grateful for your quick response.
    I would like to search for bayesian deep learning and uncertainty and in these papers I have faced some definition about calibration that I could not understand:

    1- Pitfalls of In-Domain Uncertainty Estimation and Ensembling in Deep Learning
    2-On Calibration of Modern Neural Networks

    would you mind explaining how to apply calibration in detail?

    Thank you in advance


    • Jason Brownlee August 29, 2020 at 8:07 am #

      Thanks for the suggestion, perhaps in the future.

  30. Bruno Barre September 12, 2020 at 9:22 pm #

    Hi, I would like to have your opinion to a question/idea I just asked on the “Cross Validated” site, about a global measure for calibration. Here is the link .

    It would gladly help if you have already some papers about it!

    Thanks for your great work and advice!

Leave a Reply