Post by the CEO: Redefined scoring logics from 5 to 15,000 data points

Our CEO and founder explains Kreditech’s business model from scratch: We as a financial consumer technology group use 15,000 data points in order to provide access to finance for the 4bn people with no credit score.

What is a credit score?

A credit score is a statistical number outlining the creditworthiness of a private individual in a linear number, e.g. 680 out of 1000 possible. 4bn people worldwide do not have such a score, but only positive raw data – examples would be “has 2 active credit cards”, or a negative entry like “has not paid utility bill two years ago”.

How do Equifax, Schufa, Transunion and others compute this number?

Traditional credit bureaus operate with networks of telco companies, utility providers, banks and other third party providers that inform the credit bureau about positive or negative incidents referring to individuals. Examples are primitive: Actively using a credit card for at least two years raises the score while no active payments lower it.

Why these scores are limited: Low availability and low quality

Credit scores only exist in developed markets for people who interact with companies within the network around the credit bureau. A person with a score would have to have a credit card, a cell phone contract and a bank account, for instance. In most developing markets these networks do not exist. Building such networks takes massive amounts of capital. It also takes time for setting up the network and time for “recording data”, so receiving positive and negative data from their partners. And then, the same problem reappears for people who have not used credit in the past, such as expats or immigrants. So people who are new in a country would not get a credit even though they have a fix income, e.g. in UK or the US. People who have never been active in the country’s financial system would – ironically – have to take out a credit first in order to prove credibility. This is a vicious circle based on extremely primitive and historic “rules”.

The second enormous problem is the quality of the data. A credit bureau only receives signals about good or bad incidents from the network partners. It does not have a deeper understanding about the attributes of an individual. Also, it does not include granular behavioral data. The credit bureau only learns that somebody it not creditworthy the moment the individual misses the last warning.

How is Kreditech different?

Kreditech’s approach is 100% unbiased when it comes to credit scoring. The proprietary machine learning models learn from experience, adjusting as they go by correcting the mistakes they make and avoiding them in the future. So the Kreditech Group profits from this investment long-term: while the algorithms are less efficient in the beginning, the more data they analyze, the faster they improve and the more accurate the prediction becomes. It is comparable to a “room to learn”, like if you let a small child crash with his toy car a few times, so that once he drives a real car he knows how to brake. The algorithm hence learns from every single bit and sets data patterns that optimize decision-making. Opposite to that, traditional banks just ride a taxi symbolizing the credit bureau score. The Kreditech-approach has a big advantage in comparison: Algorithms that make mistakes in the beginning, learn to become much better in identifying good customers. Even when traditional data is not available, they can score with a reliable result as they have a much broader set of data.

Isn’t Big Data a huge buzzword – how do you back yourself?

Big data is en vogue, so everyone who was using data before is now using big data. Many people and organizations confuse “a lot of data” with “big data”. Some lenders and credit bureaus have started looking into using alternative data. However, every scoring logic Kreditech is aware of uses traditional credit bureau data as major data source, attributing for approximately 95 percent of the actual decision. Only less than five percent come from alternative data sources. This means:

People with no or a bad traditional credit score still do not get any credit, even if the 5 percent alternative data are the strongest possible “+” signal. In the meantime, alternative data helps for border cases, e.g. people with a score of 799 with 800 being the knockout score. A real big data algorithm can only be built based on observation vs. gradually improving a pre-biased sample size. This is because every granted loan implies a positive credit score. For example, for someone who started his investment banking career in New York City two weeks ago: The traditional model would not find out whether the person is creditworthy as the credit bureau automatically rejects the person without looking at alternative data. What big data does better than traditional data is that it measures granular data points and repayment behavior more detailed. Beyond the surface this is not binary but comes in shades of partial repayments for instance.

What makes Kreditech’s scores better than traditional ones?

Kreditech has proven the predictive power of its proprietary algorithm by recently performing a back-test on two million scored customers to show that big data combined with traditional data produces default rates lower than the industry standard: An internal back-test looking at all issued loans between 2012 and 2014 measuring the return on investment after defaults, processing cost for data queries and such, as well as marketing cost found that:

1. Giving out credit with no scoring is a bad economical idea.
2. Giving out credit with traditional scoring is highly risky, and can even be unprofitable in markets with low credit bureau quality.
3. Our score works credit bureau independent and generates a high positive ROI
4. Our score in combination with credit bureau yields the highest ROI

Explanation:
Baseline = no scoring
Credit bureau only = reject all customers with bad credit bureau history
Big data only & CB data combined with big data = tried both to find out which works best