INNOVATION IN BANKING

Let’s talk about innovation in retail banking.

What comes to mind?

• The ATM? (Definitely a game changer)
• BPay? (An online registry of billers, probably less used now days due to the commoditisation of online payment acceptance and the growth of credit card rewards)
• Tap & Pay? (By the way Octopus Card launched it’s tap and pay card in 1997 so contactless payments has been around for a while!)
• Apple Pay? (Yes! but how come some banks left it so late, sticking a small card to your phone is ridiculous?)
• Mobile Banking App? (Trying to think about how this has changed my life but I suppose it’s just one of those things you must have)
• Buy now pay later (BNPL) services? (You used to have to sign up to each store’s Lay By programme, now you just need to sign up once with the likes of AfterPay and Zip)
• Budgeting?

A few International examples come to mind:

• PayPal, definitely a game changer as it addressed a key pain point in processing online payments for ecommerce
• Online FX / International payments (e.g TransferWise, OFX)
• Venmo - social payment system
• Uber app, wallet and payment ecosystems such as WeChat, WePay, AliPay etc.

OBSESSION WITH TRANSACTION CATEGORISATION

I recall seeing an article recently about UBank adopting AI powered tools to help millenials budget.

This is great, artificial intelligence is going to help us manage how much we have to spend each day.

This show’s how far we’ve come since…….. Wait didn’t we already have something that was an online pivot table which classified spend called pocketbook in 2012? So the idea of spend categorisation isn’t a new one, so what’s changed in the last decade?

Last year Experian acquired the Australian Fintech Look Who’s Charging. Look Who’s Charging provides ANZ and NAB customers a feature that informs them which categories their credit card transactions relate to.

So they solve the following problem for the consumer, when the merchant name differs from the public trading name, they match the two together. Sounds like a big look up table, but none-the-less, there was a gap in the market and the team came up with a solution.

It’s a no-brainer that the Banks are interested in this data.

Through your credit card transactions banks know this about you.

• When you buy
• Where you buy
• How often you buy

For example, how many times did you transact with Qantas or Virgin is a good indicator of how often you fly, and how often you fly is a good indicator of your level of discretionary spend.

The one thing Bank’s don’t have yet is access to “what you buy” This comes from the receipt and contains what we called SKU (Stock-Keeping-Unit) or PLU (Price-Lookup-Code) level data.

This is pretty much the holy-grail when it comes to consumer behavioural analytics.

Understanding a person’s life stage is a good indicator of what financial services that person will need. Are they starting out so they need a credit card, are they saving up to buy a car, are they saving for a house or having children.

Also for a retailer, to know what you buy, how often you buy and even the brands you buy, would be immensely powerful.

In recent years we have seen a very slow adoption of digital receipts. Firstly we got SMS receipts from the likes of Bunnings and JB-Hifi.

But then merchants started using what we call integrated POS from vendors such as Square were able to link the card with the in-store transaction. Meaning that once you tapped your card and entered your contact details once, they could send you the receipt no matter which store you purchased from. The receipt even has features such as rate your experience.

There are some interesting digital receipt services (e.g. Slyp) getting ready to hit the market so watch this space.

In the near future, retail behaviour analytics will start to enter into the world of product-level insights. But given the large variety of products out there (millions of products) how do we categorise them?

PRODUCT CATEGORISATION R&D PROJECT

In probing this challenge, Cognitivo partnered with UNSW’s Finance@IT research group to investigate the use of natural language processing and deep learning in building an ontology for the categorisation of products.

“The project was an excellent opportunity for UNSW not only to engage with industry but also conduct multidisciplinary work between different research teams for a single goal.  Several areas of further work have been identified and we are very keen to explore them via future grants.” (Professor Fethi Rabhi)

Project research focus areas included:

• Development of a deep learning model design to tag product / SKU data with product categories.
• An app prototype allowing the user to scan receipts and identify aggregate spending habits according to product categories
• An exploration into using semantic modelling techniques to represent custom categories enabling users to define their own product categories    

There were two parts of the challenge;

1. Items that had a pre-defined product code which was printed on receipts, which is issued by GS1, the global barcode organisation. We could screen scrape product categories and match these
2. Those that didn’t have a product code and were input by the merchant. For example you might see something like “open food” as an item when you next visit a cafe.

For the second part, our research effort create a self-organising map using a convolutional neural network categorise items into GS1 categories.

The product categorisation Model is a multiple layer neural network trained using typical receipt features. The model takes product level items as input (e.g. "Coca Cola 600ml") and outputs the probability that the item is belongs to each element of multiple product categorisation frameworks.

The convolutional neural network is composed of the following layers:

1. Vocabulary layer (extracted words from receipt represented as bag-of-words vector)
2. Hidden layers - two hidden layers (3000 and 1000 units produced best results)
3. Output layer - the probability that the item belongs to each of the product categories

From an technology perspective, the project utilised Python and Spark on Databricks hosted on Micrososft’s Azure Cloud.

WHAT’S NEXT?

We see potential applications of our algorithms in the following areas:

1. ATO classification of Work Related Expenses
2. Item category benchmarking (e.g. price comparison)
3. Company analysis of business spend (e.g. all those receipts we have been uploading to Xero)
4. Enhancements to personal financial management and budgeting applications.

‍