Algorithmic System Accuracy Reviews – Choosing the Right Approach

Even small errors in algorithmic systems can lead to customer harm, financial losses, failure to meet contractual obligations or regulatory issues. 

Accuracy reviews help us find and resolve errors that may exist, providing assurance about the overall level of reliability of the system.

When reviewing the accuracy of algorithmic systems, the method used can significantly influence how robust the review is and what it can reveal about reliability and potential improvements.

There are several approaches that we can use. A common one focuses on processes. A relatively newer method focuses on outcomes.

Opting for an approach that focuses on the outcomes, rather than the processes, can lead to more insightful and actionable results. 

This outcome-focused strategy offers reassurance by verifying whether your systems meet their intended goals - such as upholding customer promises or meeting contractual obligations.

It contrasts with traditional methods that primarily examine processes, which might miss critical discrepancies.

A major drawback with the outcome focused method is that it typically takes more time and effort. It is usually more difficult, and more expensive. It often involves full population testing. It won’t be appropriate for all situations.

This article explains why verifying outcomes is preferred over tracing through processes, and how it works.

Why depart from traditional methods?

Focusing on outcomes means evaluating whether your algorithmic systems are producing the intended results, such as accurate fees or precise broker commission payments.

This contrasts with traditional methods, which primarily examine the design and implementation of processes.

While a process focus can be useful in certain situations—such as identifying inefficiencies or ensuring regulatory compliance—it may not be the best method for accuracy reviews.

Here’s why focusing on outcomes can be more beneficial:

• Direct Verification of Results: By concentrating on outcomes, you directly verify whether the system delivers what it promises. This helps identify discrepancies between expected and actual results, allowing for timely corrections.
• Overcoming Translation Errors: Translation errors often occur when business requirements are converted into technical specifications, leading to discrepancies. By focusing on outcomes, you can bypass these translation layers and verify that results align with expectations, catching any misalignments early.
• Addressing Quick Fixes and Process Changes: Processes often evolve through quick fixes and adjustments, which can introduce inconsistencies. An outcome-focused approach can identify errors that are masked by outdated rules or undocumented changes.
• Detecting Unintended Consequences: Process-focused methods may miss critical discrepancies that arise from complex interactions within the system. For example, a seemingly minor change in one part of the process could have unforeseen effects on the final output that may not be evident when examining the process alone.
• Revealing Edge Cases: Process-focused methods often concentrate on typical scenarios, potentially missing rare but critical edge cases. Outcome focused analysis can highlight these outlier situations where the system may produce unexpected or incorrect results.
• Identifying Data Quality Issues: Process reviews may overlook problems with input data quality. An outcome-focused approach can reveal discrepancies caused by incorrect or outdated data that may not be apparent when simply examining the process flow.
• Slowly changing dimensions: The outcome focused approach is designed to deal with changes in master data during the period under review. These can be handled by a process focused review, but only if it involves significant data testing.

So, there are several good reasons for adopting an outcome focused approach.

A step-by-step explanation about how it works may help in understanding how these benefits manifest.

How the outcome focused approach works

This method involves seven key steps, one of which is iterative:

1. Scope definition: define the review objective and scope.
2. Base understanding: examine relevant documents that outline expectations (e.g., product disclosure statements, third party contracts), overall process flows and system architecture/data flows.
3. Data collection: gather relevant data: inputs (as close as possible to the source) and targets (the final outputs).
4. Reperformance: reperform the calculations, based on agreed contractual terms, to produce modelled (expected) results; then compare those with the system's outputs (the actual results).
5. Iterative refinement: if there are differences between the expected and actual, check whether this is a problem with the modelled result or the actual result. For example: a valid exception applies; the modelled result has missed a key step or made an incorrect assumption; or a data fix was applied (in the actual process) because of a bug or issue in the source system.
6. Root cause identification: once the specific inaccuracies have been confirmed, identify the source(s) of those.
7. Reporting: document the result, how the review was performed, the issues and how they will be resolved.

Other approaches also start with defining scope and end in a report. The other five steps will look quite different.

Choose wisely

While this approach can take significantly more time and effort, it is certainly possible. Newer analytics software and techniques enable testing across large datasets, reducing manual effort.

It is not the norm; most reviews follow a traditional approach.

But don’t let anyone tell you that it is impossible. It can be done, even if you have an exceptionally large dataset (we do this routinely, across hundreds of millions of records). 

By choosing an outcome-focused approach, you can improve the chances that the review will pick up errors.

This method is more robust for reviewing algorithmic accuracy than traditional assessment methods, addressing common challenges and giving you better assurance.

Disclaimer: The information in this article does not constitute legal advice. It may not be relevant to your circumstances. It was written for specific algorithmic contexts within banks and insurance companies, may not apply to other contexts, and may not be relevant to other types of organisations.