MoneySmart Group, a Singapore-based fintech, is using AWS to provide personalised, relevant and contextual recommendations to its customers.
Founded in 2009, MoneySmart helps consumers compare and purchase loans, insurance, credit cards and other financial products.
It has over 30 financial product categories and operates in Singapore, Hong Kong, Taiwan and the Philippines. The company has a separate insurance brand Bubblegum, currently available only to Singapore customers.
MoneySmart’s head of technology, Prateek Baheti, said at the recent AWS ASEAN Summit that the firm is using Amazon Personalize as a platform to “build” a product recommendation engine to help customers find the right financial products.
According to Baheti, the project kicked off in October last year in Singapore and went live with the first MVP (minimum viable product) in December. The product was made available in Hong Kong by the end of January 2023.
Since the launch, over 50,000 customers have received recommendations experience, out of which 14 percent have interacted with the suggestions, Baheti said.
The recommendations engine has directly contributed towards generating over 4,500 applications for various financial products, he added.
Prior to the tie-up with AWS, the platform used to provide a very “static user experience” to customers, Baheti said.
"Regardless of the customers' preferences, financial situation, and past interactions, they were presented with the same generic content and recommendations."
This one size fits all approach limited the firm’s ability to effectively engage with customers, Baheti said.
In addition to a comparison platform, MoneymSmart offers content articles and personal finance and lifestyle topics tailored to users’ interests to help them make informed financial decisions.
Evaluating options
Baheti noted that prior to the partnership with AWS, MoneySmart was sitting on a wealth of customer interaction data aggregated from multiple sources in its data lake and data warehouse.
He added that the company faced a gap in artificial intelligence-machine learning (AI-ML) expertise and resources. This was hampering the aggressive timeline set by it to bring a product recommendations engine to the market.
Baheti said the team evaluated three options:
First, developing a pure rules-based recommendation engine, that would have harder to manage and maintain with a large data set.
Second, developing custom machine-learning models from scratch which would require significant AI-ML expertise and resources.
Third, look for partners in platforms-as-a-service-solution (PaaS).
Baheti said as MoneySmart's entire platform was already hosted on AWS and the team has extensive experience with the service and tools; this familiarity made it easier for the company to integrate Amazon Personalize into existing infrastructure and workflows.
AWS’ ability to work effectively with large volumes of interaction data, pre-built algorithms and managed services enabled the team to develop and deploy the recommendations engine quickly, he added.
“It allowed us to focus on building a product recommendation engine without the need for extensive upskilling,” said Baheti.
Challenges
In personal finance, customers tend to be more cautious and deliberate, as the products have long-term implications on their financial decision, which is in contrast with e-commerce platforms where customers sometimes make quick and impulsive decisions, he noted.
To address this challenge, the platform leveraged “custom domain sets” instead of specific domain sets like e-commerce or video recommendations within AWS Personalize, Baheti said.
It allowed the platform to capture the unique characteristics of personal finance products and customer behaviour, he added.
Baheti added that the platform faced a cold start problem - when a large number of users visiting were new and had no prior interactions, making it difficult for a product recommendations engine to generate personalised suggestions.
He added another challenge was what he called a “low category diversity” for popular products.
This occurred because interaction datasets were skewed towards credit cards - major performers on the platform.
To that end, MoneySmart incorporated “curated lists” created by subject matter experts, which provided a handpicked selection of high-quality financial products across various categories.
This ensured that the recommendations offered both diversity and relevance to users regardless of their low interaction history, said Baheti.
Final solution architecture
Baheti said that in the final solution architecture, all interaction data from the website experience and back-end servers are pushed into AWS' Kinesis data streams that are processed in AWS Lambda and stored in S3 Data Lake.
This raw data is then processed daily in the ETL (Extract, transform, and load) jobs, scheduled via Airflow to populate the Redshift data warehouse, he added.
A relevant subset of data is then prepared and fed into Amazon Personalize in the form of user, item and interaction data sets. The model is then retrained regularly on the new data.
This process was done manually at first, but the firm is working towards automation in retraining the model, Baheti said.
He added that a key part of this architecture was to stream live customer interactions into Amazon Personalize which “modifies” the product recommendations results in “near real-time” with no need to retrain the model.
There's also a strong business rules layer built into the recommendation service, which utilises Amazon Personalize features and post-processing to make the recommendations more contextualized and relevant.
This includes promoting products as per trends and offers and leveraging readership insights from a customer's consumption of content and profile attributes, Baheti said.
The team noticed that ML-based recommendations were performing better than the recommendations by subject matter experts in terms of engagement and conversion.
Future plans
Baheti said the team wants to expand personalisation efforts beyond financial products to include personalised recommendations of content and deals across MoneySmart.
It plans to leverage deeper customer profile attributes in recommendations, including demographic, financial and behavioural attributes, to create a more comprehensive view of the needs and preferences.
“We plan to explore solutions like AWS Sage Maker to build custom models that can solve unique challenges, allowing the platform to further optimise and enhance personalisation capabilities,” Baheti said.
