Basically, I'm looking for the Bayesian equivalent of LinearModelFit. As of the moment of writing, Mathematica has no real (documented) built-in functionality for Bayesian fitting of data, but for linear regression there exist closed-form solutions for the posterior coefficient distributions and the posterior predictive distributions. Have these been implemented somewhere?

  • $\begingroup$ It might be of interest to look at the option FitRegularization in FindFit $\endgroup$
    – chris
    Jul 14, 2019 at 15:18
  • 1
    $\begingroup$ @chris I'm aware of the new fit regularization options. In particular, fitting with L2 regularization can be considered as an approximation to a MAP estimate with the appropriate priors. BayesianLinearRegression is meant to implement the full Bayesian treatment, not an approximation. $\endgroup$ Jul 14, 2019 at 15:26
  • $\begingroup$ OK I am not trying to lower the interest of what you did. I believe MAP falls within the Bayesian framework. If I use say a Gaussian prior, the MAP give an exact solution, not an approximation. I guess this is just semantics though. $\endgroup$
    – chris
    Jul 14, 2019 at 15:39
  • 3
    $\begingroup$ @chris It depends on what you call "the solution". I you ask me, the "solution" in a Bayesian setting is always a distribution, not a value or point estimate. A MAP estimate is a reduction of said distribution to a single value, which amounts to throwing away information. However, if you want to make a MAP estimate using my code, that's easy: all of the distributions are there. Just reduce them to numbers any way you want. Mean, median, mode: your choice. $\endgroup$ Jul 14, 2019 at 15:44
  • 1
    $\begingroup$ For Gaussian noise with a Gaussian prior, the MAP solution is a description of the Gaussian posterior, which is an exhaustive description of the corresponding Distribution. In any case I think people reading your post would also be interested in knowing that Mathematica now has FitRegularization now built in. The rest is linguistics IMHO. $\endgroup$
    – chris
    Jul 14, 2019 at 16:53

1 Answer 1


I submitted this question to answer it myself, since I recently updated my Bayesian inference repository on GitHub with a function called BayesianLinearRegression that does just this. I wrote a general introduction to its functionalities on the Wolfram Community and Wolfram Blog and the example notebook on GitHub shows some more advanced uses of the function. I also submitted the function to the Wolfram function repository and you can use this version simply by evaluating:

BayesianLinearRegression = ResourceFunction["BayesianLinearRegression", "Function"];

Please refer to the README.md file (shown on the front page of the repository link) for instructions on the installation of the BayesianInference package. If you don't want the whole package, you can also get the code for BayesianLinearRegression directly from the relevant package file.

Example of use

BayesianLinearRegression uses the same syntax as LinearModelFit. In addition, it also supports Rule-based definitions of input-output data as used by Predict (i.e., data of the form {x1 -> y1, ...} or {x1, x2, ...} -> {y1, y2, ...}). This format is particularly useful for multivariate regression (i.e., when the y values are vectors), which is also supported by BayesianLinearRegression.

The output of the function is an Association with all relevant information about the fit.

data = {
  {-1.5`,-1.375`},{-1.34375`,-2.375`},{1.5`,0.21875`}, {1.03125`,0.6875`},{-0.5`,-0.59375`},  {-1.875`,-2.59375`},{1.625`,1.1875`},
model = BayesianLinearRegression[data, {1, x}, x];

Out[21]= {"LogEvidence", "PriorParameters", "PosteriorParameters", "Posterior", "Prior", "Basis", "IndependentVariables"}

The posterior predictive distribution is specified as an x-dependent probability distribution:

model["Posterior", "PredictiveDistribution"]

Out[15]= StudentTDistribution[-0.248878 + 0.714688 x, 0.555877 Sqrt[1.05211 + 0.0164952 x + 0.031814 x^2], 2001/100]

Show the single prediction bands:

  predictiveDist = model["Posterior", "PredictiveDistribution"],
  bands = {95, 50, 5}
   Evaluate@InverseCDF[predictiveDist, bands/100],
   {x, -4, 4}, Filling -> {1 -> {2}, 3 -> {2}}, PlotLegends -> bands
  PlotRange -> All

Prediction bands of 1st degree fit

It looks like the data could also be fitted with a quadratic fit. To test this, compute the log-evidence for polynomial models up to degree 4 and rank them:

In[152]:= models = AssociationMap[
   BayesianLinearRegression[Rule @@@ data, x^Range[0, #], x] &,
   Range[0, 4]
ReverseSort @ models[[All, "LogEvidence"]]

Out[153]= <|1 -> -30.0072, 2 -> -30.1774, 3 -> -34.4292, 4 -> -38.7037, 0 -> -38.787|>

The evidences for the first and second degree models are almost equal. In this case, it may be more appropriate to define a mixture of the models with weights derived from their evidences:

 weights = Normalize[
   (* subtract the max to reduce rounding error *)
   Exp[models[[All, "LogEvidence"]] - Max[models[[All, "LogEvidence"]]]],
 mixDist = MixtureDistribution[
   Values @ models[[All, "Posterior", "PredictiveDistribution"]]
 (*regressionPlot1D is a utility function from the package*)
 regressionPlot1D[mixDist, {x, -3, 3}], 

Prediction bands of a mixture model of polynomials up to degree 4

As you can see, the mixture model shows features of both the first and second degree fits to the data.

Please refer to the example notebook for information about specification of priors (see the section "Options of BayesianLinearRegression" -> "PriorParameters") and multivariate regression.

Detailed explanation about the returned values

For purposes of illustration, consider a simple model like:

y == a + b x + eps

with eps distributed as NormalDistribution[0, sigma]. This model is fitted with BayesianLinearRegression[data, {1, x}, x]. Here is an explanation of the keys in the returned Association:

  • "LogEvidence": In a Bayesian setting, the evidence (also called marginal likelihood) measures how well the model fits the data (with a higher evidence indicating a better fit). The evidence has the virtue that it naturally penalizes models for their complexity and therefore does not suffer from over-fitting in the way that measures like the sum-of-squares or (log-)likelihood do.

  • "Basis", "IndependentVariables": Simply the basis functions and independent variable specified by the user.

  • "Posterior", "Prior": These two keys each hold an association with 4 distributions:

  1. "PredictiveDistribution": A distribution that depends on the independent variables (x in the example above). By filling in a value for x, you get a distribution that tells you where you could expect to find future y values. This distribution accounts for all relevant uncertainties in the model: model variance caused by the term eps; uncertainty in the values of a and b; and uncertainty in sigma.

  2. "UnderlyingValueDistribution": Similar to "PredictiveDistribution", but this distribution gives the possible values of a + b x without the eps error term.

  3. "RegressionCoefficientDistribution": The joint distribution over a and b.

  4. "ErrorDistribution": The distribution of the variance sigma^2.

  • "PriorParameters", "PosteriorParameters": These parameters are not immediately important most of the time, but they contain all of the relevant information about the fit.

People familiar with Bayesian analysis may note that one distribution is absent: the full joint distribution over a, b and sigma together (I only gave the marginals over a and b on one hand and sigma on the other). This is because Mathematica doesn't really offer a convenient framework for representing this distribution, unfortunately.

Sources of formulas used:

  • $\begingroup$ Awesome! Is there a way to set the prior? What prior did you use by default? $\endgroup$
    – Roman
    Jul 14, 2019 at 14:36
  • $\begingroup$ @Roman Yes you can: "Please refer to the example notebook for information about specification of priors" ;) (I updated that sentence with the section where you can find it) $\endgroup$ Jul 14, 2019 at 14:37
  • $\begingroup$ Nice. I believe version 12 provides penalty in all fitting routines of mathematica? This is equivalent to maximum a posteriori where exp(penalty) represents the prior? $\endgroup$
    – chris
    Jul 14, 2019 at 15:16
  • 4
    $\begingroup$ @chris Functions like LinearModelFit compute the Akaike Information Criterion and Bayesian Information Criterion. The serve a similar purpose, but are quite different from the Bayesian evidence. Furthermore, BayesianLinearRegression does not make point estimates (such as MAP) but retains the full distribution over the fit coefficients. Don't throw away information if you don't have to ;). $\endgroup$ Jul 14, 2019 at 15:23
  • 5
    $\begingroup$ Thank you for doing this work and sharing it with the community. I do find it frustrating that Wolfram has not made greater efforts to facilitate the use the Bayesian paradigm and statistical models for those users, such as myself, who are not capable of building/programming these models/implementations themselves; but nonetheless are able and desirous of following along with the competent and professional implementations of others. $\endgroup$
    – BeanSith
    Jul 17, 2019 at 3:12

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.