Small businesses have long found it challenging to access credit on favourable terms as an absence of publicly available financial data makes it difficult for lenders to use traditional credit scoring methods. The end result is that often those businesses are denied credit even if they would have had no problem paying the cash back.
When we first launched Hokodo, we had to rely on external data from data providers such as credit rating agencies to be able to build our credit risk models. The problem with external data is that it is usually only effective for larger buyers because they typically have more widely available financial information. Smaller buyers—like the small and medium-sized enterprises (SMEs) that we work with—are not required to publish their financial accounts, so we can’t access financial data such as balance sheets and profit and loss accounts that we could for larger companies. For those smaller businesses, the lack of data meant their credit scores were usually very generic to their buyer profile which meant that some low risk buyers were given harsher credit ratings than they deserved.
Using transactional data
With a machine learning model, the more data you have, the better it will perform. We recognised early on the importance of having a highly functional and expansive internal data infrastructure that would allow us to collect and analyse data at scale. When developing our latest model, we have leveraged the power of our data platform to utilise transaction-level data alongside company-level data to train our model. Previously we were only able to use company-level information, for example things like what sector a company operates in, how a company is structured or how old it is. By using transactional data, we can see more granular detail such as repayment information, the timings of transactions, or how a company behaved between transactions—intelligence that is not possible to gather using third-party data.
When we were restricted to just using external data, every time a customer returned, their credit score would either be the same or very similar because the company-level data wouldn’t have changed. With the new model, every time a company makes a transaction, that feeds through into the score, making the model more dynamic than 1+ year old public financials.
By combining company-level information with transactional data, we can gain insights into a customer’s creditworthiness that would otherwise not be possible. All of that gives us more of an idea of how that buyer is going to behave when they use Hokodo, because the more information we have when we are scoring customers, the more predictive we can be.
Financing with confidence
That gives us a competitive advantage because it allows us to optimise our models to increase the number of invoices we are confident will be paid back. To set the threshold for financing, we perform an impact analysis that balances how many companies we are willing to finance versus potential non-payment risk. With the old model, the risk threshold for whether we accepted or rejected a customer was lower. With the new model that is underpinned by transactional data, we can increase the threshold because we are confident that a lower proportion of companies’ invoices would go unpaid, therefore increasing our potential customer base.
The more data we can collect and the more we work with other platforms and expand into different geographies, the greater and more accurate the models are that we can build. In addition, having all of this data gives us greater flexibility to innovate, such as trying new features and new types of models to see what works best. This enables us to do what we call feature engineering—taking the raw data and combining it in new ways that could have a positive impact on our bottom line and broaden the availability of credit to companies that have previously struggled to get accepted.