1. Digital transformation in retail banking
Over the past few years, the banking industry has witnessed a new wave of digital transformation. Virtual banking, for example, has become more popular in many regions. Other digitalisation trends such as open banking, RegTech, AI and data-driven decision-making, to name a few, are in the headlines.
In addition, the Covid-19 pandemic is changing the way that banks and customers interact. Today, retail banking products are very much commoditised, with interest rates and other features of bank offers proving similar between them. FinTechs and TechFins have therefore emerged to bring new and better customer experiences to their incumbent counterparts. Customer expectations have also been changing, and customers are seeking digital, seamless, fast and integrated services more and more. Thus, on the other side of the table, banks are not left with much choice but to undertake necessary digital transformations to meet expectations.
Many retail banking transformations are taking place in the market. In the broad sense, we can categorise them into three types: (1) moving from product-centric to customer-centric (i.e. to have more and faster customer interactions, to offer more personalised services and advice, etc.), (2) automating end-to-end services (i.e. adoption of technology for on-boarding, e-KYC, risk management, internal controls, etc.), and (3) enabling big data analytics and data-driven business decision-making.
In this article, we focus on the second and the third categories. In particular, we dedicate this article to the discussion of credit scorecards, as one of the major tools for data-driven risk management and business decision-making. We will review traditional scorecard development methodologies and then discuss the latest trends.
2. Key makes it possible to unlock value
Retail banks typically use credit scorecards, which are mathematical models, to predict the behaviours of their customers. The most important behaviour to predict is whether the customers will default on or repay their borrowings. When it comes to such predictions, two types of scorecard are widely used: the application scorecard and the behavioural scorecard.
Table 1 – Comparison between application and behavioural scorecards
3. Traditional scorecard development framework
There are a number of tools that can be used for the development of retail credit scorecards. Historically, SAS was arguably the dominant programming language for retail credit risk management, including the development of credit scorecards. Over the past decade, open source programming languages, such as Python and R, have become more and more popular. While most banks are still using SAS now, many have started using open source languages in parallel.
Traditionally, a six-phase framework is adopted for credit scorecard development. As demonstrated in figure 1 below, the six phases are (1) data processing, (2) variable transformation and selection, (3) logistic regression, (4) performance inference, (5) scorecard scaling and (6) scorecard validation. Refer to appendix 1 for a more detailed discussion regarding the development procedures.
Figure 1 – Six-phase credit scorecard development processes
4. Challenges in traditional credit scoring
Traditional credit scorecards have been used by market practitioners for a few decades. However, they are not perfect when considered through the lens of big data. Below we highlight the major challenges faced by traditional credit scorecards today.
Figure 2 – Key challenges faced by traditional credit scoring
5. Key enablers to unlock value
We have already briefly mentioned the trends to tackle the challenges encountered by traditional credit scoring. These are certainly at the heart of retail credit decisioning in the era of big data analytics.
Figure 3 – Key trends in retail credit scoring
Trend one: Big data analytics and the use of alternative data
Looking at retail banking globally, we are seeing a strong focus on improving data and deeply understanding customer needs to create personalised experiences. Big data analytics and the use of alternative data have become one of the most prevalent trends in the industry’s transformation. With rising computing power and increasing access to advanced analytics tools, market practitioners are starting to realise the hidden value of data as well as to search for new data sources.
Retail banking has long been a data-driven business, where data is generated at every stage of the customer journey. However, historically, most banks did not have an efficient way to realise the potential of the data nor the IT infrastructure necessary to do so. Furthermore, traditional data as used in the past is just the tip of the iceberg; huge amounts of alternative data, in either structured or unstructured forms, are generated every second from various data sources, both internally and externally, in this digital era.
Over the past decade, thanks to advances in big data analytics, retail banks now have increasing capabilities to process traditional and alternative data efficiently; thus, they are able to build up the customer’s 360-degree profile digitally. With that in mind, banks are starting to provide a more tailor-made customer experience via their banking apps and digital platforms. In addition, upselling and cross-selling campaigns can now target specific customer segments based on insights from big data analytics. Developments in AI and machine learning also help banks and data providers to gain insights from unstructured data (e.g. using nature language processing (NLP) to gauge a customer’s sentiments).
Figure 4 – Traditional data and alternative data comparison
The use of alternative data not only improves the robustness of the scorecard model but also enables banks to assess the creditworthiness of untapped customer segments. This helps to extend financial services to the two billion unbanked adults globally.
With alternative data, retail banks are also able to develop more scorecards, such as income scores, propensity scores and marketing scores. These further help banks decide to whom to lend their money, how much to lend, in what time frame and through what channels.
Some FinTech firms and digital financial service providers have taken the initiative to make use of alternative data sources for credit scoring. Credit bureaus, such as Experian, can now add rent payment history to their credit scoring algorithms thanks to a tool developed by the UK PropTech firm CreditLadder. Lenddo,a software business in Singapore, has incorporated social media and mobile phone data to assess clients’ credit levels. By aggregating data from SMS footprints, electronic devices, emails and credit bureau reports, among others, Algo360, an alternative credit score solution provider, helps new-to-credit customers get loans. Small FinTech companies have used smartphone activity, including calls, GPS data and contact information, to execute credit scoring in microfinance. As alternative data accumulates, the output from predictive models is likely to become more reliable and accurate over time.
Alternative data is not only beneficial when credit scoring individuals, but also in the case of SMEs. Banks commonly consider SMEs to be high-risk clients since information about them is limited, causing difficulty in evaluating their creditworthiness. Because of the intrinsic qualities mentioned previously, alternative data, in conjunction with traditional data sources, will help to build a more comprehensive profile of SMEs, allowing lenders to make better decisions. Digital SME lenders (e.g. Kabbage, an Atlanta-based FinTech company) are making wide use of alternative data such as bank account money flows and balances, business accounting, social media, real-time sales, payments, trading, logistics, and credit reporting service provider data, as well as various other private and public sources of data, to improve risk assessment and to tap into a large market of underserved SMEs.
Moreover, the value of data can be further ‘mined’ if combined with AI and machine learning techniques, which brings us to the second major trend in the digital transformation of retail banking: AI and machine learning in modelling.
Trend two: AI and machine learning in modelling
Retail banking is one of the industries where the use of artificial intelligence (AI) and machine learning (ML) has become widespread. We have talked about the tremendous potential of data, and we believe that these new techniques are well placed to assist in unleashing this potential, particularly when it comes to credit decisioning.
Machine learning can be applied to strengthen traditional logistic regression credit scoring or a solely ML-based model can be developed for credit scoring. Below we highlight some ML techniques that can be applied by banks when developing their credit scoring systems.
Figure 5 – Common machine learning analytics applied for credit scoring system development
An ML-based model would have several advantages over a logistic regression model. First, it can capture the non-linear nature of risk factors, and thus if trained appropriately, can possess higher predictive power. Second, it is agile and dynamic enough to perform the timely assessment of customer credit quality based on greater amounts of relevant and recent data. Third, the model can be highly automated and self-improving, thereby lowering ongoing operational costs.
Higher predictive power
ML-based models are trained with much more data than their traditional counterparts. These include both traditional data and alternative data as discussed above. While traditional models are not designed to discover complicated relationships between large amounts of data, ML-based models are much stronger in this area. As such, it would not be surprising to see that ML-based models are more predictive than traditional models.
More agile and dynamic
ML-based models are continuously trained with the most up-to-date data, so that they are able to perform real-time assessments of customer creditworthiness. This allows the models to provide rapid feedback to model users for credit approval and other decision-making processes. Due to their agility, ML-based models are also more customer-centric and offer smoother assessments of customer creditworthiness. As a result, greater financial inclusion is possible.
Figure 6 – Risk assessment over time – ML model vs traditional model
ML-based models are designed to be self-improving over time and thus highly automated. Traditional models require users to recalibrate them (e.g. on a yearly basis) and redevelop them (e.g. every few years). ML-based models are able to update themselves based on updated data feeds. As such, operational costs for ML-based models are lower, especially in the long term.
With these benefits, it is no wonder that credit bureaus are aggressively using ML to evaluate large amounts of data and generate improved insights. Equifax, for example, provides its clients with tailor-made services by applying neural networks to an artificial intelligence credit scoring approach. Equifax is not alone in experimenting with ML; Experian boosts its analytics products with ML capabilities to provide richer, more insightful information. Even for ‘credit invisible’ clients with infrequently updated credit files, VantageScore incorporates ML to analyse risks and provide ratings. ML has also proved to be effective in detecting high-risk behaviours and providing more accurate credit scorecards by TransUnion and FICO. A blend of Tree Ensemble Modelling (a machine learning technique employed by FICO) and scorecards significantly improves predictive performance in credit assessment, compared with traditional scorecards.
Figure 7 – Machine learning outperforms scorecards
Source: FICO Blog (March 2022)
In addition to traditional credit bureaus, FinTech companies are also actively exploring possibilities in ML to run their businesses. For example, LendingClub, the world’s largest online platform connecting investors and borrowers, has created its credit-scoring algorithm based on ten years of LendingClub data, AI and ML technologies; Kabbage is developing next-generation ML and analytics stacks for credit risk modelling and portfolio analysis; and LendUp, an American online direct lender, employs ML algorithms to identify the top 15% of borrowers who are most likely to pay back their debts.
Notwithstanding the advantages of ML-based models, they possess some limitations to be resolved. First, ML-based models are not as trivial as traditional models, and the modelled results can be challenging to interpret. It can also be more challenging to explain the models to regulators and auditors. Second, the performance of ML-based models is highly dependent on the quality of the data used. When feeding huge amounts of data into the models, ensuring the quality of the data can be challenging.
Trend three: Process automation
The third trend is the increasing automation in almost every part of the business. In order to provide fast interactions and personalised customer experiences, automation in know your customer, credit approval, risk management and reporting has become highly important. For example, OppFi, a leading financial technology platform, effectively automates the credit scoring process by using AI models, real-time data analysis and proprietary scoring algorithms. Zest AI, an AI-empowered credit life cycle management organisation, provides banks with its automated services in data processing and documentation as well as compliance validation, deployment and integration. With the help of process automation, banks and FinTech companies are largely improving customer experience and greatly reducing operating costs by cutting loan application processes to a few minutes. Credit scoring is at the heart of credit approval and risk management, and its automation largely relates to the automation of data processing, modelling and validation.
With big data analytics, banks use both internal and external data to a great extent. Data collection and data cleansing are the major tasks to be automated. Data collection involves the collection of data from different sources, whether traditional or alternative, as well as its digitisation and standardisation. Data cleansing involves data validity checking, data backfilling, treatments for outliers and doubtful data, etc.
A large part of model development can be automated with proper governance and approval processes. For ML-based models, this is more trivial as the models are designed to improve themselves on an ongoing basis using the latest data. For traditional models, automation can be useful for recalibration and the generation of challenger models.
The validation of models can be entirely automated, whether for traditional or ML-based models. Model validation consists of calculating predefined performance metrics and comparing them with predefined thresholds. It is relatively straightforward to automate such processes and generate validation reports.
Figure 8 – Process automation with the help of ML and AI
What Accuracy does
For clients who need to navigate the digital transformation in the retail banking industry, especially in credit scoring, Accuracy is well placed to work with you on the following tasks:
• Perform an independent review and validation of your existing credit scorecards
• Develop credit scorecards using programming languages including SAS, Python, R, VBA, etc. The development process is semi-automated for easier repetition and maintenance
• Advise you on the adoption of alternative data for credit scorecard development, whether for traditional or machine learning models
• Develop machine-learning-based credit scorecards using open-source languages such as Python
• Perform automation on data processing, modelling and validation
• Perform the overall strategic shaping of retail banking digital transformation and adoption of big data analytics
At Accuracy, our financial services industry experts work with banks and non-bank financial institutions on mergers and acquisitions, strategic transformations, quantitative modelling and adoption of technology solutions. We have been working closely with global financial institutions as well as small and medium-sized institutions over the past two decades.
Appendix 1 – detailed procedures for retail credit scorecards development