It’s capable of tracking down scammers and defrauders, testing traditional models and verifying your identity, but can it decide about your loan? Mariusz Cholewa, PhD, the President of the Polish Credit Bureau (BIK), on artificial intelligence in the financial sector, in conversation with Tomasz Jurczak

Tomasz Jurczak: What exactly is automation in the credit process?

Mariusz Cholewa*: Nowadays almost all decisions on extending small credits are made automatically. Whether our credit application is approved or rejected depends on an algorithm.

Supervision authorities have been putting a lot of effort into regulating this domain. Why ?

Because laying down the rules to control algorithms is a big challenge. The automated decision-making process should allow for human intervention whenever necessary. On the other hand, each algorithm has to be transparent and verifiable.

I’m not sure if it’s always possible in the case of AI algorithms.

You’re right. Everything depends on the type of an AI model we want to apply. But, under the GDPR regulations, an average Joe has the right to know what factors have impacted a decision regarding his person. As far as the so-called statistical artificial intelligence models are concerned, i.e. those that have been trained and implemented in the “frozen” form, explaining such a decision is still possible, which is not the case of autonomous AI models. In the case of models based on artificial intelligence and machine learning which can learn and modify their parameters, it is often hard to say why they’ve produced this result and not any other.

What can we do about it?

This is why autonomous AI is not directly used for crediting services. However, models created on the basis of artificial intelligence and machine learning are used to make prototypes of production algorithms. An AI model can reveal a new relevant correlation which haven’t been used in traditional credit risk assessment models. If that happens, the conclusions can be used in models developed in a traditional way. According to the guidelines of the European Banking Authority and the Polish Financial Supervision Authority it is necessary to diligently document all models and check their quality and stability. Although such analyses are not always possible in the case of AI models, they can be performed in the case of traditional models based on conclusions from the use of artificial intelligence.

By the end of June 2020 the Intelligent Antifraud Platform of the Polish Credit Bureau helped the banking sector to foil fraudulent attempts amounting to PLN 252 million

Does that mean there’s no turning back from automated credit services?

Automation is a sign of our times. Today, after logging to our e-bank account, we often see offers that best suit our needs, e.g. card or account limits. Having access to the history of their customers’ accounts, banks can provide tailor-made offers, which allows them to increase their chances for a successful transaction. Of course, before any funds are made available to a customer, the bank will thoroughly check their credit score. Many modern financial institutions profile their customers and adapt their offers to make it easier for their individual clients to make a decision, while still searching for additional business opportunities.

To assess their customers, banks use data on the borrowers, their creditworthiness and financial standing. The decision, however, is made with the use of advanced algorithms.

What is the quality of data you’re using?

To estimate as accurately as possible whether a customer will or will not pay back their loan we refer to their credit history. With our tools [see frame] this information is visible to every bank and to each natural person logged to www.bik.pl.

The credit history can be displayed by the Polish Credit Bureau thanks to the data regarding credits granted provided by the banks two times a week. The reports include the information on the number and value of credits each person has, on the type and amount of installments each person is obliged to pay, and on whether such installments are paid as agreed in the contract. For example, if an amount was due yesterday and a client is in arrears, the information on the late payment will become visible the following day. Similarly to the account history, these are hard data which are difficult to falsify by potential scammers.

On top of that, there are also the so-called payment data. Generally speaking, they provide information on payments, e.g. if a phone bill was paid when due. Although they are second class data, they are important for assessing the credit risk level. The third category may include alternative data, e.g. from social media, which are gathered from devices customers use. These data are the least helpful in our evaluations…

… so you rarely refer to them.

Yes. Financial institutions always use hard data falling into the first of the above categories. The second group has no particular added value unless a customer has no credit history. If a borrower is new to the credit market, he or she has no official record and is anonymous to a lender. This means that not being included in our database is not so good. It would be better if the system had our credit history and if we already had a scoring.

Do those weaker data have any value?

They do but only if appropriate models are used. Data from the first and the second group are pretty well correlated with the credit risk and other types of business risk. This is borne out by the results of the projects carried out by banks and insurance companies or by banks and telecoms. Today, basing on banking data, we can predict if a customer who has purchased an expensive pay monthly phone will pay for his plan or not. The same goes for insurance services. For now, in Poland, only prototypes of similar models have been implemented and few insurance companies use them. But in the USA the credit-based insurance score is applied by almost all insurance companies.

All customers have to be protected against adverse effects of automatic decisions. Otherwise they might be stigmatized by AI algorithms used in the financial sector

And what about Facebook data?

Social media data are not very resistant to market trend changes and manipulations. Apart from that, there are also ethical issues and the need for transparency. Consult our register and read the report about yourself to know what information we have, who analyzed your data and for what purpose. It’s fully transparent. However, in the case of SI and ML models developed on the basis of social media no such transparency can be ensured, as it’s hard to establish what rules will be applied by an algorithm based on autonomous machine learning.

Does this mean that using artificial intelligence is not safe?

It is. We often test models used in the Polish Credit Bureau with AI solutions to check if we’re able to come up with an algorithm that would be better than the ones currently available.

Unfortunately, such models are not useful over the long haul as they quickly become obsolete. They also require large data sets, so it takes a lot of time to reconstruct them. If we add to this the fact that the structure of such models has to be diagnosed (transparency!), it will turn out that their use in practice is quite challenging.

What does an average Joe asking for a loan have to do with automation, artificial intelligence models and machine learning?

EU regulations focus on a consumer seen as an individual. All customers have to be protected against adverse effects of automatic decisions. Otherwise they might be stigmatized by AI algorithms used in the finance sector.

For example?

Some statistical models are based on geolocation. If someone is born in a place that is perceived by an analytical system as “disadvantaged” or in an area where credit installments are not timely paid, then the “bad” location history will weigh heavily against such a person as long as he or she decides to change their place of residence. Until that happens, the customer will have to pay more for their insurance policy, pay a higher apartment deposit or pay bigger interests on their credit.

And what about gender related stigmatization?

In Poland, but also for example in the USA, no gender category can be used for this kind of models. However, our data clearly show that women are more diligent when it comes to paying back their debts than men. It’s a universal phenomenon.

Is artificial intelligence going to have its day in credit services?

Yes. AI models are already used in automated processes. However, they have to be used in a transparent manner, properly documented and meet all the standards of traditional models.

If we want to apply AI in more serious financial operations, we will need hard data. Even the best algorithms won’t be enough if the quality of the data is poor

Some fintech companies that are trying to get their foothold in the credit market lure their customers with deferred payments. They often rely on incomplete data that are not connected with the customer’s credit profile and try to use this information to assess the credit risk. Then Mr. Smith, who badly needs the money, allows the financial institution to use extensive information about his person, giving up his privacy. As far as small loans and petty amounts to be paid back over a short period of time go, weaker models based on scarce data are enough to predict whether a customer is able to pay the money back or not. However, if we want to apply AI in more serious financial operations, we will need hard data. Even the best algorithms won’t be enough if the quality of the data is poor.

Polish Credit Bureau and artificial intelligence: what are your ideas for using AI?

We have been using AI to test and “challenge” already existing models to check and to improve their efficiency.

Artificial intelligence has also proven essential in countering scams and extortions. Apart from the traditional credit history, we also have a lot of solutions that help financial institutions to protect them against frauds, for instance the BIK Antifraud Platform, which is a tool for systemic security of the customer credit processes. The algorithms of the platform that are connected to the biggest database in Poland (storing over 8.5 million complete credit applications) search for potentially risky correlations and generate reports and warnings for its users.

AI and ML are just perfect for such operations. They make it possible to efficiently identify early symptoms of planned scams. From the end of 2017, when we launched the BIK Antifraud Platform, to the end of June 2020 we helped the banking sector to foil fraudulent attempts amounting to PLN 252 million.

Another example of using smart algorithms in practice in our industry is passive biometry. With ML models, the system is able to analyze the way a user uses his or her device and therefore learn about that user. When the system discovers an unusual way of logging in, the bank will call the customer and ask him or her to confirm their identity.

Does the Polish Credit Bureau know everything about the Poles?

I think we know really a lot about the financial sector already. We have gathered information about over 147 million accounts held by 25 million natural persons and 1.4 million enterprises. This means that we have data about almost every adult Pole and a considerable part of companies. But we don’t know everything yet.

---

*Mariusz Cholewa, PhD, President of the Polish Credit Bureau (BIK) and the Association of Consumer Credit Information Suppliers (ACCIS).
He is a graduate of the Faculty of Management of the University of Gdansk. In 2005 he took his PhD in economics. He also completed his post-graduate banking and finance studies (majoring in investment banking) organized by the London Guildhall University, the Gdansk Academy of Banking and the University of Gdansk.

Between 1993 and 1998 he collaborated with the Gdansk Academy of Banking, i.a. as a member of the Management Board, and participated in the studies on financial restructuring of companies and banks performed by the Gdansk Institute for Market Economics. In 1995-2003 he delivered lectures at the Faculty of Management, University of Gdansk. From 2007 to 2010 he was the President of the Management Board of Bank Rozwoju Cukrownictwa SA.

Between 1998 and 2013 he worked for Bank Handlowy w Warszawie SA (Citi Handlowy), where he held different managerial posts and, since April 2013, was employed as the director of the Strategy Department. He was a member of supervisory boards of various companies operating in the banking, investment, insurance and leasing industries, including companies listed on the stock exchange. From 2012 to 2013 he was a member of the supervisory board of Biuro Informacji Kredytowej SA.

He has been the president of BIK SA since June 2013. He is also the chairman of the supervisory board of Biuro Informacji Gospodarczej InfoMonitor SA. In May 2020 he became the president of ACCIS.

• black box problem
• finance
• insurance