Algorithm | Jumpstart

Jumpstart's algorithm

The Value of Jumpstart Jumpstart offers a deep learning platform to better evaluate residential solar customer default probabilities. Jumpstart’s model takes into account a number of customer data points to return a “probability of default” percentage, which can be used to determine which customers to serve.

Investors can utilize Jumpstart to make better loans on their current customer base. Jumpstart also presents an opportunity for investors to access low-FICO or non-credited households that may wish to install solar, because Jumpstart provides a creditworthiness calculation independent of FICO scores. This will allow solar investors to serve uncredited and under-credited customers across the country, and eventually, around the world, where FICO scores do not exist.

Jumpstart may also be used by utilities, to preemptively identify customers that are likely to default on future payments, so that the utility can work out additional precautions. This would also allow utilities to assess customers with deposit amounts more selectively, and minimize negative customer experiences.
Deep Learning Platform Jumpstart’s model is a powerful blend of data engineering and machine learning. The model was inspired by anomaly detection/rare-event predictive algorithms, where negative events happen infrequently and are highly undesirable.

Jumpstart’s model involves dozens of customer inputs, including:

Some feature engineering steps include polynomial feature processing, mean encoding, and synthetic minority oversampling.

Modeling Techniques Given the nonlinear nature of loan defaults, tree-based models were naturally the strongest performers, but non-tree based models such as deep forward neural nets and highly regularized regressions greatly improved performance when added in a multi-layered ensemble. The ensemble was optimized using Bayesian-optimized hyperparameter tuning. Modeling Performance In an F1-optimized model (minimizing the harmonic mean of recall and precision), an F1-score of 0.75 is obtained, and an AUC-ROC of 0.85 is achieved. Using a Kolmogorov-Smirnov test to test for similarity of distributions of the predictions and the real-world results, a p-value of 0.2249 was obtained, indicating that the two independent samples are likely drawn from the same continuous distributions. The following table shows the output dataframe from the Jumpstart algorithm.

Partnering with Jumpstart Jumpstart charges a monthly subscription. A free trial period for the first month may be available upon request.

Interested in partnering with Jumpstart, or have questions? Please reach out to us at to get started!