Richer data and AI drive up credit scoring accuracy, particularly for SMEs

Combining an entirely fresh look at the data used for credit scoring with cutting-edge AI provides a solution to one of banking’s most intractable problems: compiling accurate and reliable credit scores for companies asking for loans, in particular for thousands of SMEs involved in global trade, many of whom are rejected by lenders. Michael Boguslavsky, head of AI at Tradeteq and author of a newly-released whitepaper titled Machine Learning Credit Analytics for Trade Finance, explains how it works.

A perennial problem for small and medium-sized firms involved in world trade is failing to win often moderate amounts of finance exactly when they need it.

One of the main reasons is poor credit scoring by lenders, particularly of this sector, and part of the solution involves abandoning out-of-date ideas of what kinds of data are required to build an accurate credit score. When AI is then applied to new and richer data sources, there is a considerable improvement in the accuracy with which credit events can be predicted for smaller companies, making these high-quality credit scores a sound basis for lending.

Overcoming the traditional approach

Michael Boguslavsky

The traditional method used for credit scoring is the Altman Z-score, first introduced in 1968 but still widely relied upon, especially in its improved variations. It uses linear discriminant analysis based on a small number of specific accounting entries. Lenders using traditional methods like this may reject an application for credit just because it lacks a single vital piece of accounting information – even when a lot of other relevant accounting and non-accounting information on the company is available: instead, it is not rated at all. Yet the data on which these traditional approaches rely are taken from company accounts filed annually, with considerable delays, and are consequently already out of date.

Attempts to improve credit scoring still broadly following this traditional approach have failed to come up with an accurate gauge of SME credit-worthiness: this was clearly an area ripe for innovation. Now, a new approach promises to move credit scoring into the digital age and reliably widen its reach to encompass all small and medium-sized companies. It consists of just two crucial and innovatory steps outlined in Tradeteq’s new whitepaper.

First, credit scoring is freed from over-rigid ideas of which data are relevant to a company’s credit score. Secondly, the power of machine learning, through neural networks, is applied to increase the accuracy of credit event prediction to new levels.

Broadening and deepening the data

The data used in the new credit scoring models are richer in three ways than those used traditionally. First of all, there is no hard requirement for a limited number of accounting data, as there is under the Altman Z-score, excluding all companies failing to provide them from training and test sets (as well as refusing to provide them with a credit score). Instead, the new modelsaccommodate varying data availability across companies and utilise all data available as well as noting, and learning from, any pattern of absences.

Secondly, there is a lot of non-financial data available that can be made relevant to credit status. The new models exploit a broad set of available and emerging data categories previously uncaptured. For example, combining registration and accounting information, with geographical and socio-economic information,e.g. mapping a company‘s registration address to socio-economic area classification and census data, provides deeper data.

Finally, wherever possibleTradeteq’s models use data that are more up-to-date than those filed annually – or even less frequently – in a company’s accounts (making them potentially two-three years out of date), for example data obtained from counterparties ‒banks, large customers, electronic invoicing companies and digital marketplaces. Such private data will often be available in batches, with underlying credit exposures linked by common clients, suppliers or bank relationships. Data like these that are of higher frequency and more granular improve prediction quality.

Basing credit analysis on these broader, deeper and more up-to-date data results in a far more finely tuned understanding of how SMEs behave and of their credit risk.

Machine learning drives up accuracy of credit scoring

Improving data quality is only the first – albeit substantial‒ step. The second is applying machine learning, using neural networks (mirroring those of the human brain, to assess inter-relationships between different factors), to the data sets. The effect of this is transformative. New models are able to predict the likelihood of credit events occurring for different types of company far more accurately than the Altman Z-score or similar models. Its best performing model is a deep neural network with four hidden layers.

These models also capture and score more companies. Tradeteq’s current live UK limited company model gives credit scores to all 3.4m UK active limited companies. The latest published large-scale test of the Altman Z-score in 2014 covered just 13% of limited companies registered at Companies House–excluding smaller ones because their financial ratios made them too unstable for a failure prediction model.

Most importantly, applying AI to these richer data sets results in measurably more accurate predictions. On the above-mentioned dataset, the Altman Z-score’skey “Area Under Curve” metric (measuring model performance) wasbetween0.70 and 0.74–Tradeteq’sNeural Network UK model was 0.92.

Benefits for all:borrowers, lenders, originators and investors

Better quality data – drawn from a wider number of sources – combined with improved prediction techniques using machine learning, allowsTradeteq’s neural network models to outperform the traditional Altman Z-score (and similar models), even on pure registration data, with no accounting inputs. Dramatically improving the quality and timeliness of credit-event prediction, they result in better credit decisions and fewer loan rejections. Finally, they have far wider coverage than the Z-score –never rejecting a company because its assets are too low.

Enlisting the power of AI and rethinking data choice are thus set to stimulate lending and widen the distribution of more attractive trade finance assets to institutional investors, directly supporting global growth.

Tradeteq is now looking for partnerships and collaborations to work on transaction-level trade finance datasets, aiming to leverage its expertise in deep data analysis and the broad data sourced from partners, to produce state-of-the-art credit analysis for the trade finance community.

To learn more visit www.tradeteq.com

Combining an entirely fresh look at the data used for credit scoring with cutting-edge AI provides a solution to one of banking’s most intractable problems: compiling accurate and reliable credit scores for companies asking for loans, in particular for thousands of SMEs involved in global trade, many of whom are rejected by lenders. Michael Boguslavsky, head of AI at Tradeteq and author of a newly-released whitepaper titled Machine Learning Credit Analytics for Trade Finance, explains how it works.

A perennial problem for small and medium-sized firms involved in world trade is failing to win often moderate amounts of finance exactly when they need it.

One of the main reasons is poor credit scoring by lenders, particularly of this sector, and part of the solution involves abandoning out-of-date ideas of what kinds of data are required to build an accurate credit score. When AI is then applied to new and richer data sources, there is a considerable improvement in the accuracy with which credit events can be predicted for smaller companies, making these high-quality credit scores a sound basis for lending.

Overcoming the traditional approach

Michael Boguslavsky

The traditional method used for credit scoring is the Altman Z-score, first introduced in 1968 but still widely relied upon, especially in its improved variations. It uses linear discriminant analysis based on a small number of specific accounting entries. Lenders using traditional methods like this may reject an application for credit just because it lacks a single vital piece of accounting information – even when a lot of other relevant accounting and non-accounting information on the company is available: instead, it is not rated at all. Yet the data on which these traditional approaches rely are taken from company accounts filed annually, with considerable delays, and are consequently already out of date.

Attempts to improve credit scoring still broadly following this traditional approach have failed to come up with an accurate gauge of SME credit-worthiness: this was clearly an area ripe for innovation. Now, a new approach promises to move credit scoring into the digital age and reliably widen its reach to encompass all small and medium-sized companies. It consists of just two crucial and innovatory steps outlined in Tradeteq’s new whitepaper.

First, credit scoring is freed from over-rigid ideas of which data are relevant to a company’s credit score. Secondly, the power of machine learning, through neural networks, is applied to increase the accuracy of credit event prediction to new levels.

Broadening and deepening the data

The data used in the new credit scoring models are richer in three ways than those used traditionally. First of all, there is no hard requirement for a limited number of accounting data, as there is under the Altman Z-score, excluding all companies failing to provide them from training and test sets (as well as refusing to provide them with a credit score). Instead, the new modelsaccommodate varying data availability across companies and utilise all data available as well as noting, and learning from, any pattern of absences.

Secondly, there is a lot of non-financial data available that can be made relevant to credit status. The new models exploit a broad set of available and emerging data categories previously uncaptured. For example, combining registration and accounting information, with geographical and socio-economic information,e.g. mapping a company‘s registration address to socio-economic area classification and census data, provides deeper data.

Finally, wherever possibleTradeteq’s models use data that are more up-to-date than those filed annually – or even less frequently – in a company’s accounts (making them potentially two-three years out of date), for example data obtained from counterparties ‒banks, large customers, electronic invoicing companies and digital marketplaces. Such private data will often be available in batches, with underlying credit exposures linked by common clients, suppliers or bank relationships. Data like these that are of higher frequency and more granular improve prediction quality.

Basing credit analysis on these broader, deeper and more up-to-date data results in a far more finely tuned understanding of how SMEs behave and of their credit risk.

Machine learning drives up accuracy of credit scoring

Improving data quality is only the first – albeit substantial‒ step. The second is applying machine learning, using neural networks (mirroring those of the human brain, to assess inter-relationships between different factors), to the data sets. The effect of this is transformative. New models are able to predict the likelihood of credit events occurring for different types of company far more accurately than the Altman Z-score or similar models. Its best performing model is a deep neural network with four hidden layers.

These models also capture and score more companies. Tradeteq’s current live UK limited company model gives credit scores to all 3.4m UK active limited companies. The latest published large-scale test of the Altman Z-score in 2014 covered just 13% of limited companies registered at Companies House–excluding smaller ones because their financial ratios made them too unstable for a failure prediction model.

Most importantly, applying AI to these richer data sets results in measurably more accurate predictions. On the above-mentioned dataset, the Altman Z-score’skey “Area Under Curve” metric (measuring model performance) wasbetween0.70 and 0.74–Tradeteq’sNeural Network UK model was 0.92.

Benefits for all:borrowers, lenders, originators and investors

Better quality data – drawn from a wider number of sources – combined with improved prediction techniques using machine learning, allowsTradeteq’s neural network models to outperform the traditional Altman Z-score (and similar models), even on pure registration data, with no accounting inputs. Dramatically improving the quality and timeliness of credit-event prediction, they result in better credit decisions and fewer loan rejections. Finally, they have far wider coverage than the Z-score –never rejecting a company because its assets are too low.

Enlisting the power of AI and rethinking data choice are thus set to stimulate lending and widen the distribution of more attractive trade finance assets to institutional investors, directly supporting global growth.

Tradeteq is now looking for partnerships and collaborations to work on transaction-level trade finance datasets, aiming to leverage its expertise in deep data analysis and the broad data sourced from partners, to produce state-of-the-art credit analysis for the trade finance community.

To learn more visit www.tradeteq.com