By Sonja Kelly, Director of Research and Advocacy, and Mehrdad Mirpourian, Senior Data Analyst
In 2020, we began a journey to understand algorithmic bias as it relates to women’s financial inclusion. What is it? Why does it matter especially now? Where does it emerge? How might it be mitigated? This topic is especially important as we speed into a digital finance future. Women are less likely to own a phone, less likely to own a smartphone, and less likely to access the internet. Under these conditions, it is not a guarantee that digital credit underwriting will keep women’s digital constraints in mind. We focused our inquiry on the risks of algorithm-based underwriting to women customers. Today, we’re sharing what we’ve learned and where this research is taking Women’s World Banking in the future.
In Algorithmic Bias, Financial Inclusion, and Gender: A primer on opening up new credit to women in emerging economies, we emphasize that finding bias is not as simple as finding a decision to be “unfair.” In fact, there are dozens of definitions of gender fairness, from keeping gendered data out of credit decisions to ensuring equal likelihood of granting credit to men and women. We started with defining fairness because financial services providers need to start with an articulation of what they mean when they say they pursue it.
Pursuing fairness starts with a recognition of where biases emerge. One source of bias is the inputs used to create the algorithms—the data itself. Even if an institution does not use gender as an input, the data might be biased. Looking at the data that app-based digital credit providers collect gives us a picture of what biased data might include. Our analysis shows that the top digital credit companies in the world collect data on GPS location, phone hardware and software specifications, contact information, storage capacity, and network connections. All of these data sources might contain gender bias. As mentioned, a woman has more unpaid care responsibilities and is less likely to have a smartphone or be connected to the internet. Other biases might include the model specifications themselves, based on parameters set by data scientists or developers. We heard from practitioners in our interview sample about mistakes that coders make—either through inexperience or through subconscious biases—that all but guarantee bias in the model outputs. Finally, the model itself might introduce or amplify biases over time as the model continues to learn from itself.
For institutions wanting to better approximate and understand their own biases in decision-making, Women’s World Banking put together a simple tool that estimates bias in credit models. The tool is free and anonymous (we’re literally not collecting any data), and lives here. It simply asks a series of quick questions about a company’s applicant pool and decisions about who to extend credit to, and makes some judgements about whether the algorithm might be biased. We hope this is useful to financial services providers wanting to understand what this topic means for their own work (we certainly learned a lot through creating and testing it with synthetic data).
There are many easily implementable bias mitigation strategies relevant to financial institutions. These strategies are relevant for algorithm developers and institutional management alike. For developers, mitigating algorithmic bias may mean de-biasing the data, creating audits or checks to sit alongside the algorithm, or running post-processing calculations to consider whether outputs are fair. For institutional management, mitigating algorithmic bias may mean asking for regular reports in plain language, working to be able to explain and justify gender-based discrepancies in the data, or setting up an internal committee to systematically review algorithmic decision-making. Mitigating bias requires intentionality at all levels—but it doesn’t have to be time consuming or expensive.
Addressing the issue of prospective biases in lending is an urgent issue for the financial services industry—and if institutions do not do it themselves, future regulation will determine what bias mitigation will look like. If other industries provide a roadmap, financial services should be open and transparent about the biases that technology may either amplify or introduce. We should be forward thinking and reflective as we confront these new global challenges, even as we continue to actively leverage digital finance for financial inclusion.
Women’s World Banking intends to be part of the solution. Thanks to our partnership with data.org, a project of Mastercard and the Rockefeller Foundation, Women’s World Banking is joining with University of Zurich and two of our own Network members to include gender awareness in credit scoring algorithms. This next phase of our workstream on algorithmic bias will help us think about not only how to address bias in algorithms, but how to use technology to analyze new and emerging sources of data to increase inclusion.