By Genify on 13/03/2023, 21:24:46

What is a recommender system?

In 2006, Netflix, nine years old at the time, announced “The Netflix Prize” whose mission was to make the firm’s recommendation engine 10% more accurate than it already was. The competition instantly spread everywhere in the world. This problem was so crucial that Netflix offered a $1 million prize. An important point to note is that most of the attention was dedicated to the data set under study. Indeed, the data set comprised of 100 million ratings, of 17,770 movies, from 480,189 customers — a spectacular number at the time.

We all would like to know how companies like Netflix — valued at $216.38 billion — achieved such high levels of success and manages to recommend movies to its users with high accuracy — more than 75% of viewer activity is based on Netflix’s recommendations [1].

Nowadays, Netflix knows almost everything about its users — their favorite movie genre, their visit frequency, whether they prefer Spanish TV shows such as “Casa de Papel” (“Money Heist”) or American comedy movies such as “21 Jump Street”. One may wonder where this success comes from, and the answer is pretty simple: a powerful recommender system based on a humongous user database — more than 193 million users worldwide!

Figure 1 — Netflix recommendations based on a user’s habits
Figure 2 — Global companies all use recommender systems, that is why they are highly valuable

The famous Chinese social media for millennials — TikTok (ByteDance) — also understood that recommender systems would pace the way for its success. TikTok uses personalization essentially by unearthing and leveraging patterns in its consumers’ behavior. Given that the videos on this platform are very short (about 15 seconds), the 800+ million users generate an enormous number of training data! In fact, TikTok uses reinforcement learning to maximize time spent on the platform, and the action space is the videos to be recommended.

To simplify, the recommendation engine essentially collects, sorts and analyzes the data gathered. Then, it will “play” with the data in order to predict the future tastes of its users. The outcome obtained will form the personalized recommendations that a user will be offered when clicking on a webpage browsing a specific website.

Why should a bank use a recommendation engine?

Recommender systems lead to several benefits for banks

First, they increase direct and indirect revenues

Let’s try to understand the general concept of a recommender system. For instance, when an individual purchases a flight ticket worth $500, the recommender system will automatically understand that this person is flying abroad and suggest that they also buy an insurance travel bundle. A more relevant example in our current pandemic times could be suggestions concerning one’s health, such as masks.

Thanks to transaction data, recommender systems let banks assess when it comes to tailoring college loans or investing in real estate. This is a great advantage as it will boost banks’ revenues given the personalized activity offered to the client.

By using recommender systems, banks enable the recommendation of savings rules (take the example of Ubank) to increase adherence. The recommendation is based on transaction data, enabling adapted personalization to the client. A generic example could be: Customer 1 has a transaction history of “A B C” while Customer 2 has a transaction history of “D E F”. Recommender systems essentially learn that sequential patterns associated with “A B C” are more likely to need banking service “X” rather than another service, “Y”.

The higher the savings rate, the higher the potential return for banks. Thus, banks improve their revenues directly through digital sales from successful recommendations and indirectly through increases in customers’ savings — which can then be leveraged to sell more financial products, invest into other assets, and more.

Lastly, a recommender system can be viewed as a business tool that can boost a company’s income by up to 30%. Nowadays, a user does not want products offered on the Internet that they have already purchased or that do not interest them. That is why recommender systems aim to understand user behavior, making the users’ life easier, and the site or application gaining their trust [2].

And it decreases direct and indirect costs

Given that is it is easier to sell to existing customers than it is to sell to prospective customers, recommender systems enable cost minimization. Indeed, selling to existing customers is much easier since the probability of converting an existing customer is 60% to 70%, whereas the probability of converting a new prospect is only 5% to 20% [3].

Lastly, a recommendation engine also improves the overall customer experience

In very competitive sectors such as banking, customer experience is an essential lever. With plentiful supply, it is very easy for an unhappy customer to quickly change providers.

The importance of customer experience is even more evident today because traditional market players face competition from pure digital players which rely entirely on the customer experience with a 100% self-caring customer relationship. Faced with these competitors available online outside of traditional agency hours, the incumbent players must review their ways of interacting with their clients. Customers’ needs for immediacy and autonomy also invite companies to rethink user experience and customer relationship.

The personalization of recommendations consists of offering the customer the right product at the right time, based on their personal profile data, and above all, on the products with which the customer has interacted or for which they have shown interest. Recommendations allow banks to adapt to customer expectations, reduce the complexity of their choices, increase loyalty and, finally, increase purchase and consumption frequency.

Let’s name several trends in the uses of recommender systems in banking

Recommender system technology is already out there!

There has been tremendous development done up to this day, and more is likely to come. In Recommender systems: principles, methods and evaluation (2015, F.O.Isinkaye), the author agrees that recommender system technology is already in use for certain banks and thus ready to be improved even more in the coming decades to — once again — taking personalization of user experience one step further. Let’s name a few institutions as evidence to the fact that banks are increasingly adopting recommendation engines [2].

Santander hosted three public competitions on building a recommendation engine: “Santander Product Recommendation” in 2016 paired products with people, “Santander Value Prediction Challenge” in 2018 predicted the value of transactions for potential customers, and “Santander Customer Transaction Prediction” in 2019 identified the transaction’s protagonist. Prize money awarded was between $60K and $65K, proving the tremendous unmet need for bank recommendation systems. These three challenges relate precisely to the recommendation engine problem in the sense that a recommender system is a powerful tool capable of product recommendation, customer transaction prediction, and transaction identification.

Figure 3 — Santander’s recommender system timeline

Liv powered by Emirates NBD — based in Dubai, UAE — and is currently developing its own product recommendation engine. Thanks to Genify’s research, which consisted of a team of three contacting banks around the world for a month, found that another bank, BBVA from Spain, has already been using its recommendation engine for several years.

Nowadays, major advancements in predictability power of models have allowed for more information to be stored and processed in less time.

Banorte also built promotions in thousands of establishments with the use of Banorte Promotions. CitiBanamex via Citi Wealth Builder provided recommendations based on one’s needs and goals which was very handy for customers. Lastly, HSBC developed a new financial product of recommendation — Money-Wise.

Personalization is poised to become the future of banking

Banks whose technical abilities are sufficient will be able to enjoy the use of such algorithms, for instance to proceed with administrative tasks more quickly or to improve customer experience via personalization.

By using recommender systems, banks will have the ability to explain complex and confusing topics to their customers, especially by implementing app personalization which will revolutionize online customer experience. In other words, recommender systems adapt to one’s needs and improve interaction with the bank. As a matter of fact, customers like it best when rewards, offers or even products are tailored to their wants and needs.

Thanks to Natural Language Processing, banks will be able to offer even more personalized recommendations using top notch chatbot interfaces [4]. Finally, a recommendation engine will surely beget more engagement from the user. Amazon generated 35% of its revenue through their recommendation engine.[5]

Other applications of recommender systems in banking

An important example of loan protagonists nowadays are SMEs — especially startups — which can sometimes be in demand for consequent loans (e.g $30,000) to create a successful business. Retail is also an important loan actor. Recommender systems enable loan customization in a much easier and fashionable way than what was done in the past, making it more personalized and more convenient for the client [6]. A new trend is likely to appear that will lead to an increased ask for personalized, need-based, specific loans by SMEs, all thanks to recommendation engines’ implementation in financial services companies.

Robo-advisory is a new trend which provides financial advice online with minimal human intervention. This is very practical for banks as it can adequately advise customers in stock investments, planned home purchases, estate planning, diverse financial plans, etc. Another great aspect of recommender systems applied to banks is personalized financial product recommendations. It will boost the banks’ revenues through sales from successful recommendations, especially in times of economic crises during which most operations occur online, due to social distancing. However, let’s note a visible difference between this kind of recommendation engine and Genify’s, as applied to banking products.

Recommendation engines also enable personalized savings rule

An increasingly high number of banks will use personalized recommendation of savings rules for customer experience reasons. Indeed, this new trend will increase customer adherence and revenues for banks due to the wide diversity in clients. In fact, depending if the client is an 18-year-old female or, on the contrary, a retired male, the savings rule will look different for either situation. Hence, with personalized recommendation, banks can adequately advise their clients.

Some banks chose to create their own smart savings rules (for instance Ubank), however most choose to outsource the work to third-party companies (such as Santander).

Human presence would still matter in banking

To make emulation happen, it is essential for a bank to know how to create a recommender system algorithm capable of complementing the human presence as adequately as possible. Experts and people capable of maintaining and monitoring technologies will also be fundamentally important.

Indeed, people need to be reassured during crises or difficult markets when they see their investments plummet, but also when they need to make the correct investments which can oftentimes be correlated with emotional considerations and trust in one’s banker. In other words, persuading clients to take specific actions will a priori remain a physical banker’s role.

More reasons why a human presence will still be indispensable in the banking sector can be read at this link.

OK but how do recommender systems really work?

Recommender systems rely on filtering, which is essentially the process of choosing a smaller part of a data set and using this particular subset for analysis or display. There are several types of filtering for the recommendation — it can be collaborative or content-based.

Collaborative filtering

This type of filtering is a family of algorithms where there are multiple ways to find similar users or items and multiple ways to calculate ratings based on ratings of similar users [7].

It focuses on similarities in tastes and preferences between users without giving weight to the objects of interest. In other words, it is a technique used to predict the items that a user might like based on other people’s ratings [8]. Let’s illustrate this approach by considering this example : Person A rated “The Office”, “Friends” and “How I met Your Mother” very highly. Person B rated “The Office” and “Friends” very highly as well. User-based recommendations would push “How I met Your Mother” to Person B based on similarities with Person A. From a more technical point of view, similarity scores — based on users’ ratings of items — are computed between users themselves before generating the final recommendation

Content-based filtering

In this approach, the filtering is based on the user’s purchase history — and not on similar users’ preferences. For example, if one takes the case of Netflix, let’s consider User 1 preferring action and romance movies, amongst all the other Netflix movie categories. Netflix will recommend movie and TV show titles to User 1, from the same sections that this user prefers — that is to say action and romance movies. In other words, this approach no longer relies on other users’ opinions to make recommendations to User 1 but on the history of User 1.

Another example could be Person C often purchasing books on Amazon from author D. Whenever a new book from author D is available on Amazon, Person C will be notified because of his taste for such an author. Item-based recommendation is really powerful given the relative easiness with which an item from a certain category can be recommended to a user, based on their taste [9].

The positive aspects of the content-based filtering method are that there is no need for acquiring data on other users. One can recommend products that are not popular or new, and one is capable of making recommendations to users with unique or rare taste. Unfortunately, finding the right “features” the User seeks to have is not always easy… [10]

Figure 4 — Content-based filtering vs. Collaborative filtering

Challenges and data usage

The main challenges when adopting a recommender system are time and lack of data.

How to minimize the impact of these obstacles?

  • Since one can imagine that even if the algorithm holds more or less over a few hundred lines of code, behind it, it must be integrated into a coherent ecosystem. This is the complexity of recommender systems. The time spent developing the algorithm itself is just the tip of the iceberg. Therefore, Genify’s advice to solve this timing issue would be that banks and developers start by building simple algorithms which could then be improved over time — in terms of complexity — if the needs are there.
  • Furthermore, let’s focus on the second issue: for algorithms to work, a bank needs data! So, when developing a recommender system, one needs to take into account a tracking system. The data has to be accessible quickly so that the algorithms can learn from it and thus provide up-to-date recommendations. Thus, solutions to a lack of data would include using a simpler classifier model less susceptible of over-fitting, consider transfer learning or data augmentation (i.e. overfitting) [11].

Lastly, in reality, a range of recommendation models should be created since certain specificities inherent to the content or to Internet users can disrupt the recommendations. An ensemble of models will lead to reduced overfitting — based on gradient boosted decision trees and on the fact that each neural network layer learns to predict the error of its neighbors.

This result is a powerful common regression approach that was for instance, used to win the Zillow prize on Kaggle ($1,2 million prize money in 2017). The winning solution beat the Zillow Benchmark Model score of 0.14084 by over 13%, with a final score of 0.12110.

A well-designed recommendation engine should suggest content in all circumstances. This is why having several algorithms is useful and essential for both common and specific situations.

To conclude

A recommendation engine is a powerful tool available to banks and companies today, which will only continue to grow. Its applications are infinite. One may even argue that it is a technology whose developments have seen tremendous pace in the past ten years and have yielded billions in extra revenue to Amazon.

However, companies have only scratched the surface of the true possibilities the recommendation engine can offer and traditional banks have lagged behind with its implementation, which has given rise to new digital competitors such as Genify. Banks should therefore consult with these newcomers and partner with them in order to reach for even greater opportunities than present ones.

The recommendation engine’s main role is to recommend personalized activities or items to a customer based on similar users’ data and past actions. We have seen earlier that recommender systems lead to changing trends in today’s society, especially in the banking sector — for instance with improved personalization. Nonetheless, a human presence will still be needed to complement customer experience.

Also, collaborative and content-based filtering are the two categories of recommendation engines available to this day. To sum up, collaborative filtering deals with similarities in tastes and preferences between users, whereas content-based filtering focuses on the user’s purchase history, and not on similar users’ preferences.

One needs to have in mind that an entity lacking user data, time, and technology will almost always fail to adequately implement a recommendation engine. This is why it is essential to develop such a technology with the best possible approach. This is why Genify is also trying to make classification of data more accurate thanks to its Transaction Classifier product.

Figure 5 — User experience made possible by Genify’s Transaction Classifier

Bank needs can differentiate themselves from their competitors by implementing a better and more accurate recommendation engine, to compensate for the increasing competition, from neobanks and other fintech incumbents.

If you’ve read this far, it probably means that you found this opinion piece interesting. Keen to discuss further? Drop us a line on LinkedIn!


[1] Netflix recommendations, source: Netflix’s Twitter account

[2] McKinsey article, “Special Edition on Advanced Analytics in Banking”

[3] Customer retention and acquisition, source: Retention Science

[4] “Towards Knowledge-Based Personalized Product Description Generation in E-commerce”, Qibin Chen et al., Tsinghua University, Alibaba research, 2019, source

[5] Amazon recommendation systems’ impact on business, source:

[6] “Recommender Systems for Mass Customization of Financial Advice”, InCube, source

[7] Collaborative filtering, source:

[8] User based recommendation, source:

[9] Item based recommendation, source:

[10] Oyebode et al., “A hybrid recommender system for product sales in a banking environment”, 2020, source

[11] Data augmentation, source:

About the author

Hi! I’m Aymeric Bouvier, a graduate student researcher in Autonomous Driving and AI at Tsinghua University in Beijing, China. I am also currently interning as a Business Development Strategist at, where I focus on emerging markets such as the UAE or Egypt.

Give it a Try - It's Free

  • Explore sample enriched data & insights.
  • Query the API in a sandbox environment.
  • Explore the extended API documentation.
  • Test with up to 1000 transactions for free.
  • Access API keys.