The Data Analytics Blog

Our news and views relating to Data Analytics, Big Data, Machine Learning, and the world of Credit.

All Posts

All You Need To Know About Randomisation In The Credit Lifecycle

February 2, 2017 at 4:48 PM

If you’re involved in credit risk and existing customer marketing you’ll know that random numbers are frequently used when deploying different strategies.  As strategies grow more complex and numerous, so the role of the random number grows more important. In this blog post, I’ll cover what randomisation is, why you should do it, when you should do it and what else to consider.

In credit and existing customer marketing, a critical component in deploying a new strategy is ensuring that results can be clearly measured against a previous strategy. This is done through a scientific test-and-learn process known as champion / challenger testing.  Here one random population receives one action (champion) and another random population receives another (typically new) action (challenger).  The process is similar to how pharmaceuticals test new drugs known as “randomised double-blinded placebo-controlled” trials.

A key principal in champion / challenger testing is to ensure that when one is comparing two strategies, one is comparing like-versus-like.  That means the two populations should look very similar.  We ensure this through a process of randomly allocating the population to two different groups (“randomisation”)

Typical examples of Champion / Challenger strategies across the credit life-cycle include:

  • Originations: New scorecards and policy rules, sensitivity to interest-rates, lower / higher, loan / limit amounts, automated / manual decisions.
  • Account Management: Credit-limit increase / decreases, authorisations, card re-pricing, and pre-delinquency.
  • Marketing: Testing cross-sell / up-sell and targeted campaigns, and testing channels.
  • Collections: Testing different account prioritisations, settlement optimisation, right-time-to-call, and agency-allocation.

In complex environments, such as account management where multiple tests are required, the use of more than one random number is becoming crucial.  As strategies get more complex (for example through deploying mathematical optimisation where many combinations of strategies might be tested), so the need for champion / challenger testing grows.

How is randomisation done in originations?

In originations we don’t know what type of application we will get when, so typically random numbers are allocated sequentially or through the use of a random-number generator. These random numbers are normally integers between 00 and 99. 

Multiple random numbers might be allocated if multiple originations tests are required. If multiple random numbers are allocated a more complex system is required (see below). 

How is randomisation done in other decision areas (or complex originations)?

In account management the process is different.  A new account entering the account processor for the first time will be allocated a random number or multiple random numbers.  For “multiple random numbers” the numbers also need to be independent in relation to each other.  For this, a random number generator is used. This generator produces sets of two, three or four numbers that are independent to each other and also, as each set is allocated to accounts, they result in each number having a random distribution between 00 and 99.  Principa has built such a system for clients to produce up to four random numbers to be deployed in a COBOL, JAVA or .Net environment. 

Ensuring randomisation remains effective: Re-randomising your book

Once random numbers are issued on an existing book, over time one finds that the random groups lose their “smoothness”.  If, for example, one uses one set of Random Digit Groups, or RDGs, for a conservative strategy of limit increases and another set for an aggressive strategy, over time the second group will likely end up with higher average balances and limits than the first group.  This would mean that any new champion/ challenger tests could break the like-versus-like rule and therefore not be “true” tests. Re-randomisation is therefore important - ideally it should be an annual event, but one should always be aware of any additional tests that might be currently running.  It should ideally be run at the end of a test-and-learn process when the new champions are “crowned” – i.e. the results of the tests are measured and a new strategy is adopted.

How do you re-randomise?

Re-randomising is an iterative process that involves the following key steps:

1. Understand your testing environment

  • What tests are being planned over which decision areas?
  • How many random numbers can the account management system support?
  • Can the reporting identify champion/challenger groups?
  • How many random numbers will need to be produced based on answers to the above? 

 2 Identify your exclusions

Certain accounts you may wish to exclude from tests, particularly if they are treated or have been treated differently, for example:

  • Staff accounts
  • VIPs
  • Accounts with unusually high limits or balances (these will skew the smoothing process)

3. Extract the full book to be re-randomised

4. Allocate random numbers to entire book

(These numbers need to be independently random from each other)

5. Test each random group across each key metric to ensure smoothness (see box below)

6. Either:

  • If distributions are uneven, then go back to step 4
  • If distributions are “smooth” then upload new random digits to accounts in the host (account processing) system

Steps 4 - 6 can take some time, but the process is important.

Typical examples of Champion / Challenger strategies across the credit life-cycle include:

  • Similar number of accounts in each Random Digit Group (RDG),
  • Similar distribution of limits, balances, delinquent accounts, over-limit accounts , utilisation%, dormant accounts across RDGs,
  • For marketing communications – checking customer cluster distribution would also be critical, and
  • Similar distributions across each combination of RDGs.

With a fully randomised (“smoothed”) population, credit and marketing analysts can have confidence that running tests using the RDGs will produce true tests.  Randomisation is an important process to understand the best actions to take for the most profitable decisions.

Principa’s key randomisation services include.

  • Random number generators,
  • The set-up of a random allocation scheme, and
  • Re-randomisation.

To find out how Principa can help with your randomisation needs get in touch for a chat!

Debtor Profile and Treatment Framework infographic

Image credit: Shutterstock

Thomas Maydon
Thomas Maydon
Thomas Maydon is the Head of Credit Solutions at Principa. With over 17 years of experience in the Southern African, West African and Middle Eastern retail credit markets, Tom has primarily been involved in consulting, analytics, credit bureau and predictive modelling services. He has experience in all aspects of the credit life cycle (in multiple industries) including intelligent prospecting, originations, strategy simulation, affordability analysis, behavioural modelling, pricing analysis, collections processes, and provisions (including Basel II) and profitability calculations.

Latest Posts

How chat is revolutionising the digital onboarding experience

Principa’s onboarding chatbot solution; Atura allows lenders to engage a customer effectively through an application process while accessing necessary data and decisioning calls using Principa’s SmartSuite software. The digital revolution “Digital” has been a financial services buzz-word for some time. Most South African lenders Principa works with have been working hard to adapt to a digital existence for several years. Some have been successful, others are still working on the challenge - and most have only partially adapted.

How to choose the correct collections chatbot

Principa has a wealth of experience in building and deploying chatbots for the financial services industry. Our custom-built solution is flexible and fully customisable which allows your bot to assume your brand’s persona. We can also seamlessly integrate with existing systems. Click here to find out more. 

Model validation and adjustment

The time is NOW for model validation and adjustment. One of the major premises used in credit scoring is that “the future is like the past”. It’s usually a rational assumption and gives us a reasonable platform on which to build scorecards whether they be application scorecards, behavioural scores, collection scores or financial models. That is reasonable until something unprecedented comes along. You can read about this black swan event in our previous two blogs here and here.