Inference vs Prediction (2023)

Generative modeling or predictive modeling?



(Video) Inference Vs. Prediction

Inference vs Prediction (1)

The terms inference and prediction both describe tasks where we learn from data in a supervised manner in order to find a model that describes the relationship between the independent variables and the outcome. Inference and prediction, however, diverge when it comes to the use of the resulting model:

  • Inference: Use the model to learn about the data generation process.
  • Prediction: Use the model to predict the outcomes for new data points.

Since inference and prediction pursue contrasting goals, specific types of models are associated with the two tasks.

Model interpretability is a necessity for inference

In essence, the difference between models that are suitable for inference and those that are not boils down to model interpretability. What do I mean with model interpretability? I consider a model interpretable if a human, particularly a layman, could retrace how the model generates its estimates. Consider the following approaches for prediction:

  • Interpretable: Generalized linear models (e.g.linear regression, logistic regression), linear discriminant analysis, linear support vector machines (SVMs), decision trees
  • Less interpretable: neural networks, non-linear SVMs, random forests

Only a subset of interpretable methods is useful for inference. For example, linear SVMs are interpretable because they provide a coefficient for every feature such that it is possible to explain the impact of individual features on the prediction. However, SVMs do not allow for estimating the uncertainty associated with the model coefficients (e.g.the variance) and it is not possible to obtain an implicit measure of model confidence. Note that SVMs are capable of outputting probabilities but these probabilities are just a transformation of the decision values and are not basted on the confidence associated with the parameter estimates. This is why even interpretable methods such as linear SVMs and decision trees are unsuitable for inference.

In contrast, consider linear regression, which assumes that the data follow a Gaussian distribution. These models determine the standard error of the coefficient estimates and output confidence intervals. Since linear regression allows us to understand the probabilistic nature of the data generation process, it is a suitable method for inference. Bayesian methods are particularly popular for inference because these models can be adjusted to incorporate various assumptions about the data generation process.

(Video) Inference vs Prediction HD 720p

Examples for prediction and inference

Merely using a model that is suitable for inference does not mean that you are actually performing inference. What matter is how you are using the model. For example, although generalized linear models are suitable for inference, I recently used them solely for prediction purposes. Consider the following examples that make the distinction between prediction and inference clearer:

  • Prediction: You want to predict future ozone levels using historic data. Since you believe that there is a linear relationship between the ozone level and measurements of temperature, solar radiation, and wind, you fit several linear models (e.g.using different sets of features) on the training data and select the model minimizing the loss on the test set. Finally, you use the selected model for predicting the ozone level. Note that you do not care at all about the Gaussian assumption of the model or the additional information that is available on the model estimates as long as the model minimizes the test error.
  • Inference: You want to understand how ozone levels are influenced by temperature, solar radiation, and wind. Since you assume that the residuals are normally distributed, you use a linear regression model. In order to harness the information from the full data set and you do not care about predictive accuracy, you fit the model on the full data set. On the basis of the fitted model, you interpret the role of the features on the measured ozone level, for example, by considering the confidence bands of the estimates.

Workflows for inference and prediction

The basic workflows for inference and prediction are described in the following sections.


  1. Modeling: Reason about the data generation process and choose the stochastic model that approximates the data generation process best.
  2. Model validation: Evaluate the validity of the stochastic model using residual analysis or goodness-of-fit tests.
  3. Inference: Use the stochastic model to understand the data generation process .


  1. Modeling: Consider several different models and different parameter settings.
  2. Model selection: Identify the model with the greatest predictive performance using validation/test sets; select the model with the highest performance on the test set.
  3. Prediction: Apply the selected model on new data with the expectation that the selected model also generalizes to the unseen data.

The two modeling communities

Note that machine learning is often concerned with predictive modeling, while the statistical community often relies on stochastic models that perform inference. Due to the complexity of machine learning models, they are often treated as black boxes. For inference problems, on the other hand, the working principles of used models are well understood.


Three contradictions in statistical modeling

In his famous 2001 paper, Leo Breiman argued that there are three revolutions in the modeling community, which are represented by the following terms:

  1. Rashomon: There is often not a single model that fits a data set best but there usually is a multiplicity of models that are similarly appropriate. The term Rashomon refers to a classic 1950 Japanese feature film in which various characters depict different and even contradictory versions of the same incident.
  2. Occam: The simplest model is often the most interpretable. However, the simplest model is also often not the the most accurate model for a given task. Occam refers to William of Occam’s razor, which states that Entities should not be multiplied beyond necessity, underlining the importance of simplicity.
  3. Bellman: Large feature spaces are often seen as a disadvantage. However, modern machine learning methods heavily rely on expanding the feature space in order to improve predictive accuracy. So, are we subject to a curse or a blessing of dimensionality? The term curse of dimensionality was coined by Richard Bellman. The term is used to emphasize the challenges of statistical modeling in large feature spaces.

Blessing or curse of dimensionality?

Predictive modeling particularly embraces the idea that high dimensionality is a blessing. All of the recently developed popular machine learning models such as neural networks and SVMs rely on the idea of expanding the feature space in order to learn about the non-linear relationships between the independent variables. As a consequence, these models can also be thought to violate Occam’s razor because these models do not utilize compact representations of the variables but rely on partially redundant encodings.

Generative modeling, on the other hand, is often concerned with feature elimination in order to allow for improved interpretability and generalizability. Since these methods strive for simple models for explaining the data generation process with few features, they simultaneously fulfill Occam’s razor and circumvent the curse of dimensionality.

Feature selection and Rashomon

Models based on feature selection are, however, much more affected by Rashomon, the multiplicity of appropriate models. Imagine you are doing generative modeling and the original data set contains 10,000 features. Using a notion of variable importance, you reduce the number of features down to 100. You then interpret the resulting model according to the 100 features. Now, the question is: Is this model the only way to view the data generation process?

Probably not because, stochastically, it is likely that there exists another model with a different set of 100 features that explains the outcome similarly well. This is due to the existence of a huge number of subsets with exactly 1000 features, namely \({100\,000 \choose 100} = 10^{2430}\). Thus, when performing feature selection, small perturbations in the data may lead to the selection of a model using a different feature subset.

(Video) Inference vs. Prediction: An Overview


Prediction and inference can be differentiated according to the following criteria:

Model Selection- Evaluate a variety of models
- Select the best-performing model
- Reason about the data generation process
- Select model whose assumptions seem most reasonable
Validation- Empirically determine loss on test set- Use goodneess-of-fit tests
Application- Predict the outcome for new samples- Use the model to explain the data generation process
Ramnifcations- Model interpretability suffers
- Model validity shown for the test set
- Model may overfit if the test data are similar to the training data
- High model interpretability
- Model validity is uncertain since predictive accuracy was not considered
- Overfitting prevented through simplified assumptions


So what can the communities of predictive and generative modeling learn from each other? In my opinion, practitioners of machine learning should spend more time with the data in order to make well-informed decisions about the type of predictive model that is used. In many prediction scenarios, well-established methods such as SVMs or neural networks are used without thinking about the alternatives according to the saying if all you have is a hammer, everything looks like a nail. However, there may be other models that may be more appropriate to a given data set, for example because they perform better, allow for improved generalizability, or because they are more interpretable.

Additionally, there should be more skepticism about the dogmatic use of the test error as the sole measure of model validity. Particularly in settings where the test set is similar to the training data, a low test error may not guarantee a low generalization error. Finally, further effort should be made in order to peel away some of the layers of dark color that render predictive models intransparent.

In the inference setting, model performance should be included as a criterion for the evaluation of model validity since it is hazardous to draw conclusions from an inaccurate model. If the validity of a model has not been sufficiently established, all interpretations should be taken with a grain of salt and it may be worth to investigate the perspectives that other models offer. Additionally, model assumptions should be well argumented rather than assuming a certain distribution (e.g.normal) for mere convenience.

Finally, when pruning the feature space (in both inference and prediction), one should judiciously decide on the feature selection criterion and be aware that the selected model may not be the only reasonable model for the data at hand. To demonstrate the validity of the selected features, resampling methods such as the bootstrap can be used to demonstrate the robustness of the selected variables.


How is answering an inferential question different from answering a prediction question? ›

In general, if it's discussing a future event or something that can be explicitly verified within the 'natural course of things,' it's a prediction. If it's a theory formed around implicit analysis based on evidence and clues, it's an inference.

How is inferring different from prediction? ›

Predicting Definition: Good readers think about what is going to happen and make predictions based on what they know and what they've read. Inferring Definition: Good readers use what they know (schema) and what they read in the text to make a conclusion (reading between lines).

What is the relationship between inference and prediction? ›

Prediction is the process of using a model to make a prediction about something that is yet to happen. The inference is the process of evaluating the relationship between the predictor and response variables.

What is the difference between an inference and a Guess? ›

An inference is an idea or conclusion that's drawn from evidence and reasoning. An inference is an educated guess. We learn about some things by experiencing them first-hand, but we gain other knowledge by inference — the process of inferring things based on what is already known.

What is a simple definition of inference? ›

Definition of inference

1 : something that is inferred especially : a conclusion or opinion that is formed because of known facts or evidence. 2 : the act or process of inferring (see infer): such as.

Is assuming and inferring the same thing? ›

An assumption is something we usually take for granted or presuppose while inference is a conclusion you draw depending on your observations. The main difference between assumption and inference is that we make assumptions without any evidence or facts, while we make assumptions based on facts and evidence.

What is an example of a inference? ›

A. An inference is a conclusion that has been reached by way of evidence and reasoning. For example, if you notice someone making a disgusted face after they've taken a bite of their lunch, you can infer that they do not like it.

› Language ›

Difference Between Inference And Prediction What is the difference between inference and prediction? Both words refer to a conclusion based on some sort of fact...
by TeachThought Staff. Reading comprehension is a core tenet of schooling. Many K-12 academic standards emphasize–and require–strong literacy skills and some (i...
The main difference between inference and prediction is that prediction is foretelling a future event or an occurrence but, in inference, the future event.

What is the difference between an inference and a prediction quizlet? ›

Inferencing is often employed when a reader would like to determine a feeling or trait of a character or a particular circumstance at that moment using details in the text. Prediction is usually done when a reader would like to make a guess on what will happen in the text based on information or clues in the text.

What is the difference between descriptive questions and inferential questions? ›

Descriptive statistics summarize the characteristics of a data set. Inferential statistics allow you to test a hypothesis or assess whether your data is generalizable to the broader population.

What is the difference between a question hypothesis and prediction? ›

Scientific method

Observations lead to questions and these can be answered using experiments. Then an idea is thought of to help answer the question. This is known as a hypothesis . After this, a prediction is made.

Is inferential statistics better than descriptive? ›

A study using descriptive statistics is simpler to perform. However, if you need evidence that an effect or relationship between variables exists in an entire population rather than only your sample, you need to use inferential statistics.

How do you answer an inferential question? ›

QUICK TIP 1: Identify an inferential question. Look out for keywords such as 'why', 'how', 'give a reason' or 'explain'. QUICK TIP 2: Look for evidence in the passage. Gather relevant clues in the passage to support your answer.

What are the 3 types of inferential test are there? ›

There are three basic types of t-tests: one-sample t-test, independent-samples t-test, and dependent-samples (or paired-samples) t-test.

Is a hypothesis just a prediction? ›

A hypothesis is not a prediction. Rather, a prediction is derived from a hypothesis. A causal hypothesis and a law are two different types of scientific knowledge, and a causal hypothesis cannot become a law. A theory is not necessarily a well-supported explanation.

What 3 words should always be used in a prediction statement? ›

Predictions are often written in the form of “if, and, then” statements, as in, “if my hypothesis is true, and I were to do this test, then this is what I will observe.” Following our sparrow example, you could predict that, “If sparrows use grass because it is more abundant, and I compare areas that have more twigs ...

Does a hypothesis have to be a prediction? ›

Hypothesis and prediction are both a type of guess. That's why many people get the two confused. However, the hypothesis is an educated, testable guess in science. A prediction uses observable phenomena to make a future projection.


1. Inference and Predictions... What's the Difference???
(Ms c)
2. Prediction vs Inference
(Madina Kurmangaliyeva)
3. SCIENCE: Inference and Prediction
(samuel andanar)
4. Inference VS Prediction (Basic Science Process Skills) | RESEARCH I
(Khristine Min Jin)
5. Inference vs. Prediction: How do you distinguish between these two strategies?
(Dr. Vanessa Howard)
6. Prediction lecture 1: Prediction vs causal inference
Top Articles
Latest Posts
Article information

Author: Edmund Hettinger DC

Last Updated: 24/05/2023

Views: 6259

Rating: 4.8 / 5 (58 voted)

Reviews: 81% of readers found this page helpful

Author information

Name: Edmund Hettinger DC

Birthday: 1994-08-17

Address: 2033 Gerhold Pine, Port Jocelyn, VA 12101-5654

Phone: +8524399971620

Job: Central Manufacturing Supervisor

Hobby: Jogging, Metalworking, Tai chi, Shopping, Puzzles, Rock climbing, Crocheting

Introduction: My name is Edmund Hettinger DC, I am a adventurous, colorful, gifted, determined, precious, open, colorful person who loves writing and wants to share my knowledge and understanding with you.