Show HN: Credit reports about German companies

[gab_]

Hello,

In addition to my studies in computer science, I have been working on a side project. I obtain data from the Unternehmensregister, a register where every German limited company is required to publish their financial statements. These statements are published as HTML files and are completely unstructured. While financial statements often look similar, companies are not required to follow a specific structure, which often leads to inconsistently formatted statements.

The use of the Unternehmensregister is completely free, so you can check out some examples.

I wrote code that converts the unstructured financial statements into structured data using the ChatGPT API. This works well. Of course, there are some problems that have not yet been solved, but data extraction works well for the majority of companies.

I than coded a Random Forest algorithm to estimate the probability of default for a company based on its financial statement from the Unternehmensregister. I built a website to present the structured data along with the scores. Essentially, I create a credit reports for companies.

Currently, there are four companies in Germany that also create credit reports (Schufa, Creditreform, Crif, and Creditsafe). Other companies resell the data from these four providers. I provide the same services as these companies, but without including personal information such as directors or investors. The market for this service is quite large; for example, Creditreform sold over 26 million credit reports about companies in 2020.

My probability of default prediction performs quite well, achieving an AUC score of 0.87 on my test data. An AUC of 0.87 means that there is an 87% chance that the model ranks a randomly selected company that defaults higher than a randomly selected company that does not default. Additionally, there are many more companies to crawl for my database.

Currently, I am focusing on companies that are required to publish their profit and loss statements. For testing purposes, there are currently 2,000 companies available on my website.

At the moment, the website is only available in German, but you can use Google Translate, which works ok for my website.

Thank you very much for your feedback!

[WhatsName]

Given you are operating in Germany, where sending people cease-and-desist letters (Abmahnungen) purely for monetary gain, I would highly reccomend you to take good care of compliance topics like having a proper Impressum (mandatory contact page).

Unless of course if you plan on never ever growing it into a business, then you might get away with Njalla and cloudflare as invisibility cloak...

[leobg]

The irony is… if you don’t have an impressum, where should they send their abmahnung? :) In 2008, they could still look up your WHOIS. But this doesn’t work anymore since GDPR.

[gab_]

Yes, of course, but at the moment the site is only there to try out my idea and evaluate its potential.

[gab_]

I just made some major updates: 1. I can now also create company reports for companies that do not publish a profit and loss statement. 2. For approximately 16.000 companies, I am able to create a monthly credit report. I can do this because, in addition to the financial statements, I include data such as interest rates, unemployment rates, CPI, and so on. This has also made my models a lot better.

[PeterStuer]

When I consulted in European financial services ICT (credit scoring and automated descisioning for asset based finance, AML etc.) the German data had an explicit regulatory restriction that data could only be obtained for one specific transaction and that that consent was not transferable. We obtained the data on company X explicitly for transaction Y. We could not pass on the data, nor simply reuse it for another purpose.

Has that changed?

[gab_]

I only work with public data and not personal data, even if it is publicly available. Anyone can look at the Unternehmensregister, the data that is uploaded there is not for a specific purpose. This data is there to inform e.g. customers, suppliers, creditors, employees etc. about the company's activities.

[drchaos]

This is pretty interesting, I think you should do two things right now:

1. add a message box stating that it is experimental and has only a very small set of companies right now

2. add an option to get notified when you have a more complete dataset (just use a Google form to collect email addresses)

Reason: Searched for my company, no result, ok, we're too small. Searched for some DAX companies, no results either => site looks broken.

Additional ideas:

* Add information from insolvenzbekanntmachungen.de, it's a major PITA to find someone there * Provide a (paid) API so it can be integrated into shop systems etc.

A Creditreform membership is quite expensive, probably worth it for larger shops, but for small enterprise your solution might come in handy.

[gab_]

Thank you! I've just fixed the first point. For the second point, I have to look for an alternative to Google Forms. At the moment I am not yet obtaining any data from Insolvenzbekanntmachung. Insolvencies are also stored in the Unternehmensregister. However, this could certainly be integrated quite well.

[costco]

It's cool that you were able to get the data even though it's not perfectly structured. Maybe you'll be the Dun & Bradstreet of Germany :)

[gab_]

Thank you, maybe :)

[davedx]

Excellent stuff! I’ve worked in this area. Have you considered applying ratios like the Altman z-score?

I’m also curious how you back tested to get the final scoring.

[gab_]

Thank you! Not yet, but it sounds interesting. The possibilities are endless. Currently, I am testing some methods from survival analysis.

The data is very imbalanced; there are very few insolvent companies compared to solvent ones. Therefore, I work with synthetic data in my training dataset. To get the final score, I need to scale the predictions to achieve a heavily right-skewed distribution.

Currently, I am using the method Platt scaling.

[riedel]

Hi, if you want to make this into a viable product, you might want to consider that credit worthyness assessment will be considered high-risk under the AI Act. This means that you will have some compliance headaches if you want to go beyond pure R&D. As the AI Act goes beyond GDPR, it won't matter I would guess if it is about company data, which btw can absurdly also be considered personal data by some data protection bodies. Still very nice.

[gab_]

Thank you! I will inform myself about this.