Poor credit scoring is a key reason SMEs fail to win the trade finance they need. AI may provide the answer says Michael Boguslavsky, Tradeteq’s head of AI and author of a newly-released whitepaper titled ‘Machine Learning Credit Analytics for Trade Finance’.
Companies involved in trade often need small loans for short periods to tide them over. This might be a short-term cash injection to pay a supplier, book space on a container ship or bridge an overhead until they get paid for previous shipments. Yet, despite their requirements often being relatively modest, many find it hard to obtain the cash they need: frequently they are failed by the cumbersome and inflexible machinery banks and other financiers bring into play to process their applications and assess their credit scores.
In fact, lacking just one vital piece of information may scupper an application entirely. For smaller and medium-sized companies (SMEs) – who comprise the majority of trading firms as well as applicants for trade finance – this is a constant and disabling concern. Little wonder, then, that the well-documented trade finance gap (between the demand and supply of trade finance) – compiled by the Asian Development Bank – stands at $1.5trn.
Given this, any new approach aimed at encouraging lenders to finance more companies involved in trade, especially SMEs – as well as increase the uptake of this asset class by institutional investors – should be welcomed, not least for its potential contribution to global prosperity. Greater transparency is the key, as is the stronger understanding of SME credit risk, both for lenders and for investors. This calls for more accurate credit analysis and prediction, especially for SMEs involved in trade around the globe.
Traditional approach too rigid
One of the main reasons companies are refused the funding they need is that they cannot satisfy the rigid requirements of traditional credit rating techniques such as the Altman Z-score. First introduced in 1968, but still widely relied on, especially in its in improved versions, the Z-score uses linear discriminant analysis based on a highly-selective number of accounting entries. Meanwhile, it ignores a lot of potentially-valuable accounting and non-accounting information about companies. Any company unable to produce just one of the required accounting entries is simply not rated at all. Worse still, the accounting data on which it is based are taken from company accounts filed annually, and are consequently always out of date.
Over the years, attempts have been made to improve traditional credit scoring – perhaps by adding new financial ratios or by replacing the Altman Z-score linear approach with other models. Yet none has succeeded in curing its basic flaw: its inability to provide accurate credit-scoring for SMEs.
Innovation is required – with the most ground-breaking recent developments occurring in the field of machine learning, or AI. Using neural networks (i.e. by mirroring the neural networks of the human brain to assess inter-relationships between different factors) it is possible to predict the likelihood of credit events occurring for different types of company with far greater accuracy – not least by drawing on a far richer data base. Importantly, AI can draw on more diverse data types, and more detailed and more up-to-date company data.
Broader and deeper data sets
This is the approach outlined in the recently published whitepaper by Tradeteq, titled Machine Learning Credit Analytics for Trade Finance. The whitepaper explains that, by using a broad set of available and emerging data sources, it is possible to exploit categories of company information previously uncaptured by the more traditional credit scoring models.
For instance, combining registration and accounting information with geographical and socio-economic information provides deeper data – examples include mapping a company‘s registration address to socio-economic area classification and census data. In fact, there is a wealth of non-financial information available, including data on a company’s industry, as well as the company’s history of renaming and mortgage charges. All can be included and made relevant for a company’s credit status.
Secondly, unlike the Altman Z-score, the new model does not restrict its data breadth by placing a hard requirement on the availability of certain accounting entries while excluding all companies from training and test sets that fail to provide them (as well as of course refusing to provide them with a credit score). Instead, the new model accommodates varying data availability across companies, utilising all available data and noting the absence of any data so that the model learns from the pattern of absences.
Finally, the new 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 to three years out of date). Improving prediction quality requires higher frequency and more granular data on companies, which is obtained from counterparties such as 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.
In summary, the data used in new credit scoring models are richer in three fundamental ways than those used traditionally, and consequently result in a far more finely calibrated understanding of SME behaviour and credit risk.
Using machine learning to drive up accuracy of credit scoring
The application of machine learning, using neural networks, to the data sets is a second key ingredient to this new approach to credit scoring – one that is truly transformative in its effect. The combination of this with the richer data base is able to predict the likelihood of credit events occurring for different types of company with significantly greater accuracy than the Altman Z-score or similar traditional models.
Applying machine learning specifically to trade finance data yields a lot of valuable insights, for example allowing the identification of features not visible in simple cross-sectional analysis, such as irregularities in repayment patterns relative to other customers of a given supplier, or changes in trade payments relative to other similar companies from the same industry or region.
And it is this application of AI to broader and deeper data that allows a far greater number of SMEs to be captured and credit-scored. The latest published large-scale test of the Altman Z-score and its variations in 2014 covered private limited companies in 35 countries, with around 340,000 companies in the UK included – or just 13% of all the limited companies registered at Companies House. Smaller companies were excluded on the basis that their financial ratios were too unstable for a failure prediction model.
Meanwhile, Tradeteq’s current live UK limited company model, by comparison, combines inputs from Companies House, the London Gazette (the official public record of the UK government) and the Office of National Statistics and provides credit scores for all 3.4m UK active limited companies. Its best performing model is a deep neural network with four hidden layers.
Applying AI to broader and deeper data sets also results in measurably more accurate predictions than those made by more traditional models. On the above-mentioned dataset, the Altman Z-score’s key “AUC” metric (used to measure model performance) was between 0.70 and 0.74, whereas Tradeteq’s Neural Network UK limited company credit model version 2 compared very favourably with an AUC of 0.92.
A virtuous circle: benefits for borrowers, lenders, originators and investors
The joint use of better quality data – drawn from a wider number of sources – as well as improved prediction techniques using machine learning, allows neural network models to outperform the traditional Altman Z-score (and similar models) even on pure registration data, without using any accounting inputs. They deliver a better understanding of SMEs’ credit risk, dramatically improving the quality and timeliness of credit-event prediction. And they result in fewer loan rejections as well as improved credit decisions. They also have much wider coverage than the Z-score models – critically, never rejecting a company on the basis of its assets being too low.
Of course, better transparency into SME credit risk is likely to stimulate lending and widen distribution of trade finance assets to institutional investors – releasing badly-needed additional cashflows into trade finance and fuelling a virtuous circle of growth in trade and prosperity – all thanks to taking a fresh look at the data required for credit analysis, and enlisting the help of AI.