Machine learning models pose a unique set of challenges to model validators. While exponential increases in the availability of data, computational power, and algorithmic sophistication in recent years has enabled banks and other firms to increasingly derive actionable insights from machine learning methods, the significant complexity of these systems introduces new dimensions of risk. When appropriately implemented, machine learning models greatly improve the accuracy of predictions that are vital to the risk management decisions financial institutions make. The price of this accuracy, however, is complexity and, at times, a lack of transparency. Consequently, machine learning models must be particularly well maintained and their assumptions thoroughly understood and vetted in order to prevent wildly inaccurate predictions. While maintenance remains primarily the responsibility of the model owner and the first line of defense, second-line model validators increasingly must be able to understand machine learning principles well enough to devise effective challenge that includes: Analysis of model estimation data to determine the suitability of the machine learning algorithm Assessment of space and time complexity constraints that inform model training time and scalability Review of model training/testing procedure Determination of whether model hyperparameters are appropriate Calculation of metrics for determining model accuracy and robustness More than one way exists of organizing these considerations along the three pillars of model validation. Here is how we have come to think about it. Conceptual Soundness Many of the concepts of reviewing model theory that govern conventional model validations apply equally well to machine learning models. The question of “business fit” and whether the variables the model lands on are reasonable is just as valid when the variables are selected by a machine as it is when they are selected by a human analyst. Assessing the variable selection process “qualitatively” (does it make sense?) as well as quantitatively (measuring goodness of fit by calculating residual errors, among other tests) takes on particular importance when it comes to machine learning models. Machine learning does not relieve validators of their responsibility assess the statistical soundness of a model’s data. Machine learning models are not immune to data issues. Validators protect against these by running routine distribution, collinearity, and related tests on model datasets. They must also ensure that the population has been appropriately and reasonably divided into training and holdout/test datasets. Supplementing these statistical tests should be a thorough assessment of the modeler’s data preparation procedures. In addition to evaluating the ETL process—a common component of all model validations—effective validations of machine learning models take particular notice of variable “scaling” methods. Scaling is important to machine learning algorithms because they generally do not take units into account. Consequently, a machine learning model that relies on borrower income (generally ranging between tens of thousands and hundreds of thousands of dollars), borrower credit score (which generally falls within a range of a few hundred points) and loan-to-value ratio (expressed as a percentage), needs to apply scaling factors to normalize these ranges in order for the model to correctly process each variable’s relative importance. Validators should ensure that scaling and normalizations are reasonable. Model assumptions, when it comes to machine learning validation, are most frequently addressed by looking at the selection, optimization, and tuning of the model’s hyperparameters. Validators must determine whether the selection/identification process undertaken by the modeler (be it grid search, random search, Bayesian Optimization, or another method—see this blog post for a concise summary of these) is conceptually sound. Process Verification Machine learning models are no more immune to overfitting and underfitting (the bias-variance dilemma) than are conventionally developed predictive models. An overfitted model may perform well on the in-sample data, but predict poorly on the out-of-sample data. Complex nonparametric and nonlinear methods used in machine learning algorithms combined with high computing power are likely to contribute to an overfitted machine learning model. An underfitted model, on the other hand, performs poorly in general, mainly due to an overly simplified model algorithm that does a poor job at interpreting the information contained within data. Cross-validation is a popular technique for detecting and preventing the fitting or “generalization capability” issues in machine learning. In K-Fold cross-validation, the training data is partitioned into K subsets. The model is trained on all training data except the Kth subset, and the Kth subset is used to validate the performance. The model’s generalization capability is low if the accuracy ratios are consistently low (underfitted) or higher on the training set but lower on the validation set (overfitted). Conventional models, such as regression analysis, can be used to benchmark performance. Outcomes Analysis Outcomes analysis enables validators to verify the appropriateness of the model’s performance measure methods. Performance measures (or “scoring methods”) are typically specialized to the algorithm type, such as classification and clustering. Validators can try different scoring methods to test and understand the model’s performance. Sensitivity analyses can be performed on the algorithms, hyperparameters, and seed parameters. Since there is no right or wrong answer, validators should focus on the dispersion of the sensitivity results. Many statistical tactics commonly used to validate conventional models apply equally well to machine learning models. One notable omission is the ability to precisely replicate the model’s outputs. Unlike with an OLS or ARIMA model, for which a validator can reasonably expect to be able to match the model’s coefficients exactly if given the same data, machine learning models can be tested only indirectly—by testing the conceptual soundness of the selected features and assumptions (hyperparameters) and by evaluating the process and outputs. Applying model validation tactics specially tailored to machine learning models allows financial institutions to deploy these powerful tools with greater confidence by demonstrating that they are of sound conceptual design and perform as expected.