Application of machine learning algorithms to the analysis of financial traditional data and additional specific risk factors extracted from digital Alternative Data.
Dedicated to Credit & Risk Management, Asset Management, Family Offices
CREDIT SCORE BY MACHINE LEARNING
Application of machine learning algorithms to the analysis of financial traditional data and additional specific risk factors extracted from digital Alternative Data. In this way, the number of variables considered simultaneously increases significantly (over 1,000) and greater precision is demonstrated by backtesting. Some variables remain decisive, for example the geographical location of the company which, however, acquires a greater level of detail with a breakdown of the Italian territory according to the different zip codes.
With our proprietary NLP technologies we can also analyze and turn textual documents in data, such as the explanatory notes.
3 specific risk factors
We also measure 3 specific risk factors for which we use digital Alternative Data. The addition of these factors contributes to completing the estimate of the financing or investment risk in SMEs or micro-enterprises. The main reason lies in the fact that the digital data are dynamic and collected continuously, unlike the balance sheet data necessarily static as they refer to the relevant year and, moreover, compiled, published and deposited subsequently.
Additional risk factors can also be very useful to insurance companies and debt collection companies.
A leader company in Business & Credit Information chose FinScience to improve its 18-months default risk assessment model for unlisted companies.
The previous analysis was based on a maximum of 16 variables extracted from traditional financial data. The scoring was attributed considering the regional membership of the companies.
Leveraging Artificial Intelligence, from the same traditional data FinScience has extracted 1,400 variables and added to them the digital risk factor. The scoring is attributed considering the municipality of origin or the ZIP code if the company headquarter is located in a big city.
Results from the backtesting analysis
After only three months of work, FinScience was able to improve the reliability of the model by 13%, going from 80% to 93%. Thanks to the new scoring model, on a subset of 135,000 companies FinScience identified:
The added-value of Alternative Data
Find out more about our investment strategies to beat the market and get alpha