GARP RAI Exam Math: What You Actually Need to Calculate
The GARP Risk and AI Exam is not a programming examination, but it is not entirely non-quantitative.
GARP states that no coding or programming is required. However, it describes the mathematical difficulty as comparable to an advanced undergraduate or introductory graduate course in finance, statistics, or economics. Candidates should therefore expect to interpret quantitative outputs and perform focused calculations—not build machine-learning models from scratch.
GARP does not publish a separate public formula sheet listing every calculation that may appear. The definitive scope is the 2026 RAI Study Guide and Learning Objectives, together with the registered curriculum, end-of-chapter questions, and official practice exam in GARP Learning. The guide identifies the tested learning objectives and number of questions associated with each chapter.
1. Classification Performance Metrics
This is one of the most important calculation areas because risk managers must determine whether a model is identifying the correct cases—and what types of mistakes it is making.
Start with the confusion matrix:
True positive
True negative
False positive
False negative
Be able to calculate and interpret:
Accuracy
Correct predictions ÷ total predictions
Precision
True positives ÷ all predicted positives
Recall or sensitivity
True positives ÷ all actual positives
Specificity
True negatives ÷ all actual negatives
False-positive rate
False positives ÷ all actual negatives
False-negative rate
False negatives ÷ all actual positives
F1 score
2 × (precision × recall) ÷ (precision + recall)
Do not study these as interchangeable formulas. Connect each metric to the cost of an error.
For a fraud-detection model, low recall may allow fraudulent transactions to escape detection. Low precision may create excessive false alerts and operational work. The exam may test which metric should receive priority in a particular risk scenario.
2. Regression Error Metrics
For models predicting a continuous value, such as loss severity or transaction value, understand how prediction errors are measured.
Prioritise:
Mean absolute error
Average of the absolute differences between predictions and actual values.
Mean squared error
Average of the squared differences between predictions and actual values.
Root mean squared error
Square root of the mean squared error.
You should understand why squared-error measures penalise large errors more heavily than absolute-error measures. Do not stop after calculating the number; be prepared to determine which metric is more appropriate when extreme prediction errors are particularly costly.
3. Basic Probability and Conditional Probability
You should be comfortable with:
Probabilities expressed as percentages or decimals
Joint and conditional probability
Expected values
Basic Bayes-style updating
The relationship between event frequency and model predictions
A common trap is confusing:
P(actual positive | predicted positive)
with:
P(predicted positive | actual positive)
The first is related to precision; the second is recall. The distinction matters greatly when the positive event is rare.
4. Descriptive Statistics and Relationships
Know how to calculate or interpret:
Mean and weighted mean
Variance and standard deviation
Covariance
Correlation
Percentiles and distributions
Standardisation or scaling
The exam may present model-development data and ask which statistic indicates dispersion, association, instability, or a shift in the underlying population.
Remember that correlation measures association—not causation. A high correlation also does not automatically mean that one variable should be included in a model.
5. Simple Model Outputs
You do not need to program a regression, decision tree, or neural network. GARP explicitly states that coding is not required. You may still need to work with a simplified model output.
Practise:
Substituting values into a simple linear equation
Interpreting the sign and size of a coefficient
Converting or interpreting a probability output
Following a basic decision-tree split
Calculating a weighted input or score
Comparing model results under different thresholds
The important skill is usually interpretation: what changes when an input, coefficient, classification threshold, or decision rule changes?
6. Fairness and Group Comparisons
Responsible AI questions may provide outcome rates for two groups and require you to compare them.
Be ready to calculate simple differences or ratios involving:
Selection or approval rates
False-positive rates
False-negative rates
True-positive rates
Error rates across protected groups
Do not conclude that a model is fair because its overall accuracy is high. Aggregate performance can conceal materially different outcomes across groups.
Use the Calculator You Will Receive
The RAI Exam contains 80 equally weighted multiple-choice questions and allows four hours. GARP provides an on-screen digital calculator; candidates do not bring a personal financial calculator.
Practise performing short calculations with a basic digital calculator. Avoid relying on spreadsheet functions, saved formulas, or programming tools that will not be available during the exam.
The Best Formula-Study Method GARP RAI Exam Math
For every quantitative measure, create a four-part record:
Formula
Meaning of the numerator and denominator
Situation in which the measure is useful
Risk of interpreting it incorrectly
Then complete the official end-of-chapter questions and practice exam. GARP includes both resources with registration and revises the RAI curriculum annually, so candidates should rely on the 2026 material rather than older formula lists. GARP RAI Exam Math
The RAI Exam does not require you to become a mathematician. It requires you to recognise the appropriate measure, calculate it correctly, and explain what the result means for AI risk, model performance, fairness, and governance.
Comments