Machine Learning for Solar Accessibility: Implications for Low-Income Solar Expansion and Profitability
The solar industry in the US typically uses a credit score such as the FICO score as an indicator of consumer utility payment performance and credit worthiness to approve customers for new solar installations. Using data on over 800,000 utility payment performance and over 5,000 demographic variables, we compare machine learning and econometric models to predict the probability of default to credit-score cutoffs. We compare these models across a variety of measures, including how they affect consumers of different socio-economic backgrounds and profitability. We find that a traditional regression analysis using a small number of variables specific to utility repayment performance greatly increases accuracy and LMI inclusivity relative to FICO score, and that using machine learning techniques further enhances model performance. Relative to FICO, the machine learning model increases the number of low-to-moderate income consumers approved for community solar by 1.1% to 4.2% depending on the stringency used for evaluating potential customers, while decreasing the default rate by 1.4 to 1.9 percentage points. Using electricity utility repayment as a proxy for solar installation repayment, shifting from a FICO score cutoff to the machine learning model increases profits by 34% to 1882% depending on the stringency used for evaluating potential customers.
As is customary in economics, the authors are listed alphabetically. This paper has benefited from conversations with the Solstice Initiative. Financial support from the Department of Energy is gratefully acknowledged. Onda's work was supported by the Kimmelman Family E-IPER Fellowship and the Satre Family Fellowship during his doctoral work at Stanford University. The views expressed herein are those of the authors and do not necessarily reflect the views of the National Bureau of Economic Research.