SMB-modell: Kredittmodell for små- og mellomstore bedrifter

Small and medium-sized businesses (SMBs) represent one of the pillars of today's economy, but struggle with limited access to capital to fund their operations and growth. FundingPartner focuses on filling this gap and contributing to their growth and development by improving their access to capital in the form of loans. It is essential for the interest rate of SMB loans to accurately reflect the risk inherent in the borrower's business model and operations, and the likelihood that they will be unable to fulfil their financial commitments. This project focuses on developing a machine learning model that serves as a key part of FundingPartner's credit risk estimation process, enabling us to better quantify the likelihood that the borrower will default on their obligations. This number, usually referred to as PD (probability of default), combined with LGD (loss given default), enables FundingPartner to set a fair and accurate interest rate which is attractive and reasonable both to the borrower as well as the investors who fund the loans. In addition, the credit model helps automate the risk estimation process, which decreases the time our analysts spend on assessing potential borrowers. These savings can be passed on to SMBs as lower effective costs of financing. The models we focus on are especially tailored for loans backed by the borrower's cash flow (as opposed to real estate), including venture backed technology companies which often have sustainability and innovation deeply ingrained in their business model. So far, we have collected several large datasets which comprise nearly 200 000 Norwegian SMBs, using their historical accounting data to produce key financial ratios as input to the model. In addition, numerous flags have been added to the datasets, which encode quantitative and qualitative information that tells a more complete story about the company not captured by accounting data. We have also added several variables that measure the prior experience of key people such as the CEO and the board chairperson, all of which are factors that have been shown to correlate with on a borrower's probability of default. Numerous machine learning and deep learning models have been tested on the resulting datasets, with only the most robust models being deployed to production. We have so far deployed three machine learning model versions to our internal software platform. These models have been assisting our credit analysts for several months already, helping them make more accurate assessments of a borrower's credit risk. Our initial results from the existing model versions indicate a consistently high prediction power, and further development of the models is being guided by feedback from credit analysts and other subject matter experts. The focus for the next year is on expanding the data set to include additional variables and reflect the knowledge we have about companies applying for loans, as well as building models based on datasets across the Scandinavian countries. Moreover, internal data sources will be used to a larger degree, which is expected to yield a unique competitive edge.

Dette prosjektet tar sikte på å utvikle en kredittmodell for små- og mellomstore bedrifter. Små- og mellomstore bedrifter utgjør 99% av alle bedrifter i Norge, og 47% av alle ansatte i privat sektor. Totalt omsetter disse bedriftene for nær 700 milliarder kroner. Samtidig har de over de siste årene, og særlig siden finanskrisen sakket akterut i tilgangen på kapital, sammenlignet med store selskap. I 2006 utgjorde lån til SMB-er nesten 70% av alle nye bedriftsbanklån. I 2015 hadde andelen krympet til 50%, en relativ nedgang på nesten 30%. Vi tror en del av forklaringen på denne nedgangen skyldes mangel på kostnadseffektive og presise kredittmodeller og metoder for å evaluere små- og mellomstore bedrifter. De siste årene har mengden tilgjengelig informasjon om selskaper økt betraktelig, og utviklingen vil fortsette over de neste årene. Med utgangspunkt i denne utviklingen søker dette prosjektet å utvikle en effektiv og treffsikker kredittmodell beregnet på små- og mellomstore bedrifter. Kredittmodellen utvikles i samarbeid med Norsk Regnesentral, og skal ha som hovedma°l a° regne ut sannsynligheten for at et selskap misligholder forpliktelsene sine. Det unike med modellen vil være en innovativ tilnærming til hvilke variabler og informasjon som blir inkludert i modellutviklingen. Ved hjelp av NLP (Natural Language Processing) og Deep Learning og Bayesiansk statistikk kan store mengder data systematiseres og analyseres. Dette vil være særlig nyttig i vurderingen av små- og mellomstore bedrifter, hvor kvalitativ informasjon ofte er av vesentlig karakter. Modellen vil utvikles i tett samarbeid med kredittanalytikere, og anvendes i daglige kredittvurderinger. Det vil gi den gode forutsetninger for å kontinuerlig testes og etterprøves. Sluttresultatet forventes å øke treffsikkerheten til kredittanalyser, samtidig som tidsbruken per analyse reduseres. Dette vil bedre tilgangen på finansiering for SMB-er innen blant annet bærekraft, innovasjon og teknologi.