Model explaining anomalies in QIWI transactions

QIWI is a popular electronic payment system.

With the help of QIWI wallet users are able to make transfers amoung themselves, pay bills, and make online purchases

QIWI provides acquiring services to businesses.

Acquiring is a technology that allows companies to accept non-cash payments.

Payment passed or not

The main quality indicator in acquiring is conversion.

Payment failed - the client is dissatisfied

We try and prevent that from happening

On the side of the acquiring system

On the provider side

On the side of banking systems

There was a instants where for one mobile user account credits did not work between 1-3 am - its turned out the user had their time zones incorrectly set

Identify signs of anomalies which cause payments to fail

What we did:

We looked at transaction data, looked for patterns.

September - November 2018

Hypothesis:

With a fictitious wooden toy store, transactions will be more successful than with a fictitious online casino

The customer could offer different acquiring prices depending on the likelihood of the success of the transaction

We had trouble verifying transactions becasue the customer restricted the provider company information due to security protocol

Classic Data Science

Sergey Verentsov

Developer

Ramil Gizatullin

Developer

Received dumps of QIWI databases in PostgreSQL system

Collected preliminary markup from databases

Processed data using Python

Built a decision tree

You can go through it and see what set of features leads to unsuccessful transactions.

Clustered by provider group

Each node is a condition for a sign of a transaction

Orange nodes are where many transactions have failed.

Limited access to data

Complexity of the subject area

No guarantees of meaningful results

Communication problems

For example, feature importance is familiar to datasets, while analysts are more familiar with correlations and funnels.

Confirmed the client's hypotheses:

One user tried to put 3.5 billion rubles on the phone.
Of course, nothing came of it.

Transactions with large amounts with a high probability will not pass

There are problematic providers and provider groups

Unfortunately, there was no «magic pill» to fix one problem and transactions will grow. But extra confirmation is always good.

Learned to plan better

Understood the importance of working closely with the client

For a week we worked in Innopolis together with a QIWI analyst, exchanged competencies.

Georgy Konnov,

Head of QIWI Cashier, one of the business customers of the program

«... As a result, with the EORA team, we made a data analytics case, during which we analyzed the work of our providers and were able to identify new factors that reduce the conversion of payments. Working with them will lead to an increase in the turnover of our partners.»

«It was a unique experience for our data scientists: we got financial data on transactions that are not available in other ways. The volume of data was very large, and a whole infrastructure had to be deployed for them.

We worked side by side with specialists from QIWI, learned how the processes inside the corporation are arranged, learned how to interact with the security service.»

CEO EORA

Roman Doronin