“Only when the tide goes out do you discover who has been swimming naked.”
Those who act fast and adapt to turmoil will survive and succeed. In the context of the Coronavirus pandemic and the lending business, this means that financial institutions should have instruments to continuously analyze on a permanent basis what is happening and react immediately if needed.
As experts in credit scoring utilizing artificial intelligence (AI) and machine learning (ML), we have prepared this publication to offer an in-depth understanding of the problems the consumer finance business is currently facing. We focus on counteraction mechanisms, emphasizing the potential of automated decision-making in general. The dramatic change in people’s lives, the social distancing measures and the heterogeneous impact of the crisis on particular segments of the population, especially employment, as well as the need for additional financing led to a shock for the consumer finance business. As a result, some businesses have drastically reduced their operations and switched their credit scoring processes to manual loan approvals. This significantly reduced the speed and efficiency of their lending process as a whole. Some companies even chose to shut down part of their business abroad due to fears of significant deterioration of their loan portfolio and the negative effect of the imposed legal measures on debt repayment and collateral liquidity.
Other consumer finance businesses have restructured their operations, and some have focused entirely on optimizing their strategies for risk control, and long-term business development. At times like these, when consumer behavior and the economic situation are rapidly evolving, the way credit businesses react is key not only to their future development and the redistribution of market shares across the sector, but also to their survival.
This publication focuses entirely on the consumer finance sector, which services the most crisis-sensitive segment of the population. This segment includes borrowers who for the most part are unable to access financing from the banking sector because of their precarious financial situation. It is structured as follows: first, we review the changes affecting the risk level and provide concrete examples, next we explore the challenges and options for timely, rapid and effective business response, which is needed now more than ever, and finally we conclude by discussing the changes in the standard approach to risk analysis, necessary in order to provide adequate business assistance in such a crisis-sensitive sector.
What are the economic changes impacting credit risk?
COVID-19 has led to a rapid increase in unemployment, which has increased the risk assigned to borrowers working in areas, like travel, leisure and entertainment. On the other hand, the income of the retired part of the population is covered by the state and therefore it is not expected to be affected by the COVID-19 epidemic. Another group of changes caused by the pandemic is related to people’s lifestyle. For example, the demand for short-term travel or holiday loans is declining due to transportation restrictions imposed by the different governments as one of the measures to fight the spread of the virus. Loans intended for luxury goods purchases are also expected to decline due to falling incomes and the increase of uncertainty regarding the outlook of the economy. Furthermore, the need for additional funding to support the family budget – for food and utility bills, is expected to grow. On top of this, potential legal changes limiting interest rates and affecting the processing of claims must also be taken into account by risk assessment systems in the early stages. Due to the inertia of the economy, these and other changes in consumer behavior have yet to manifest to their full extent. Therefore, it is imperative for consumer finance companies to adapt their operations quickly to the changing market conditions during the upcoming months, if they want to succeed in the new reality caused by COVID-19.
Consumer loans amid COVID19
Among the hardest lessons that the financial sector learned from the 2008 financial crisis is that credit models can deteriorate quickly, and borrowers with identical credit scores can perform in a dramatically different fashion based on when in the credit cycle a loan was originated. Recalibrating the business processes for risk assessment around black swan events is something that needs to be done in a timely manner, especially in the consumer finance sector. Nowadays, there are state-of-the-art analytical solutions based on Machine Learning methods that provide banking and non-banking financial institutions with the ability to quickly and automatically generate scorecards and calibrate the lending strategy. This is a capability that did not exist during previous crises and downward economic cycles. In addition, these analytical solutions use data streams to self-learn effectively and to adapt to the new environment. It has been observed that the impact of the introduction of data-driven analytical solutions is much more pronounced during the recession phase of the economic cycle. This is definitely a good time for the businesses to strengthen their ML practices by investing in such technology with an eye to the future. On top of that organizations should not neglect the use of alternative data sources when assessing consumer credit. Such additional data can provide significant information regarding the vulnerable part of the population, which will allow for better decision-making and faster models adaptation.
What are the challenges consumer financing is currently facing?
The emergency COVID-19 situation has changed both the behavior of the credit applicants, as well as the whole environment that the business operates in. In order to survive and especially to expand their market share, consumer finance companies face the following challenges:
- companies have to be able to quickly update how they measure the creditworthiness of their clients;
- companies have to be able to update the definition of a good borrower in terms of his behavior over time;
- the approval process of loan applications in case of statistically insufficient data about their current behavior.
The rapid alignment of the credit strategy to the dynamics of the environment is a major challenge for any credit institution. The reason for this is due to the fact that the classification of a customer as either good or bad requires an extensive repayment history of this customer. In addition, since having more data helps to improve the model’s accuracy the longer credit history would help to more accurately classify a consumer as either good or bad. In reality this this statement holds true only when the economic and social conditions are relatively stable. A short credit history can have an advantage in that it leads to a faster adaptation to the current risk revels However, the period should not be too short, as a time period for the so called “Bads maturing” is needed.
In brief, one of the most important challenges facing all the players which operate in the consumer finance industry would be finding the delicate balance between the speed at which new data is incorporated into the decision-making process and the accuracy of the risk assessment. Updating how risk is measured is usually reduced to picking a value for the maximum number of overdue days, above which the consumer is considered to be too risky and is classified as a bad borrower. In a crisis situation the challenge is to recalibrate the definition for high risk customers taking into account the current economic situation while still using the pre-crisis customer payment history. In other words, businesses would have to make a decision to what extent a pre-crisis consumer can still be considered a good borrower without having sufficient data to support this decision.
The human factor and analytical applications in periods of crisis
As mentioned above, some consumer finance institutions have stopped relying on data-driven analytical systems and have instead switched to manually reviewing each loan application. The reasons driving this modification are the rapidly changing environment and the necessity for higher flexibility with regards to the application of the various business rules in the loan approval process. The disadvantages of such an approach are the significant increase in processing time, which in combination with more restrictive rules and the inherent human tendency towards inconsistency can lead to a significant shrinkage of the customer base.
Apart from the need to quickly respond with new classification rules, another reason to exclude part of the historical data and the ML approach as whole from the loan assessment process is that past data is already outdated and does not reflect accurately the current environment. During the first couple of months of the pandemic outbreak, we found that this statement applies mostly to certain consumer segments, most notably the travel and leisure industries, as well as to people employed in sectors, which provide non-discretionary goods and services.
Yet, one indisputable strength of the data-driven approach is its ability to detect customers that have either proven to be good borrowers or are new applicants with ”good” profile. Any company in the consumer finance sector would like to retain or attract such customers since at times when the financial system is exposed to considerable stress, these customers are an important source of income without unnecessarily burdening the financial institution’s balance sheet. Moreover, employing a data-driven approach, helps to easily classify high-risk candidates as such and in doing so it allows the non-banking financial institutions to be better able to manage the credit risk of their loan portfolios. It is no surprise then that these benefits have led some of the players in the sectors to allocated additional resources and invest heavily in data-driven analytical solutions.
Is a market share redistribution possible amid COVID-19?
Based on the above, the symbiosis between loan officers, with their business knowledge and a data-driven analytical tool in the risk assessment process is the right solution for fast and accurate business response.
The truth is that in many countries the market for consumer finance is already saturated. The pandemic outbreak in such countries provides an unprecedented opportunity for significant market share redistribution, which could not be achieved with any advertising or marketing campaigns. The change in how companies behave would inevitably lead some long-term customers to look for new opportunities due to the more conservative approach employed by some of the market players. So, the speed and accuracy of the credit assessment decision-making that are the key to the stable expansion of market share. The symbiosis mentioned in the previous topic can be achieved by using the invaluable experience of the loan officers to define clear business rules, which could be integrated into risk assessment analytical applications. This is an extremely important step since the accumulated data on consumer behavior during the COVID-19 crisis is still limited. On the other hand, historical data allows ML applications to aggregate information about the degree of credit risk that a company finds acceptable and thus support the decisions made by the loan officers or completely ignore them.
Deterioration of risk assessment during a crisis
One of the hardest lessons the financial sector has learned from the 2008 financial crisis is that assessments produced by credit risk models can rapidly deteriorate and borrowers with an identical credit score can perform dramatically different, depending on when in the credit cycle the loan was originated. This makes the models highly dependent on changes in the economic environment. This means that vintage analysis, which assesses credit quality based on the loan origination date, is the most reliable tool for understanding whether actual outcomes are aligned with expectations. Furthermore it can be used as an early warning indicator, if models begin to fail during the credit cycle downturn.
As seen during the 2008 financial crisis, credit defaults were highly concentrated in the vintages closest to the downturn, and as the economic deterioration spread from the housing sector to other parts of the economy, many of the diversification assumptions build into the models proved not to be well founded. By the time most banks discovered these modeling flaws, it became way too late.
The solution: Evaluation process adaptation
Beyond the symbiosis of the human factor with the analytical applications, other improvements to the standard approach to modeling credit risk are needed. Such improvements are related to:
- using credit contracts with at least one historical payment date
- updating the “good borrower” definition, replacing it with a more conservative one
- taking into account the behavior of borrowers during the crisis and giving it more weight in the construction of the risk models
- integration of appropriate business rules in case of statistically insufficient data.
The research which this note is based on uses data reflecting the specifics of the consumer finance sector. In order to assess the effect of the difference in the economic conditions and the potential of the modified analytical approach to adapt to this change, the last 112 days of data have been artificially modified to reflect the change in the behavior of borrowers during the crisis.
Three studies were conducted, in each of which 7 scorecards were built one after another, artificially changing the date of their automated generation, over a period of 2 weeks, starting 28 days after the beginning of the pandemic outbreak. Thus, the first scorecard is built using only data before the crisis and hence no applications originated during the state of emergency were involved. The reason behind that is that the minimum required 30-day period, including the first installment of crisis applicant, has not yet expired. The second scorecard applications were made during the first couple of weeks of the COVID-19 crisis. The third scorecard is built with data generated in the first 4 weeks of the crisis, etc. The studies also introduce a new definition of a good borrower, which takes into account the negative effect of the crisis on global lending. In order to do this, the allowable overdue period has been reduced compared to the standard definition of consumer loans. This takes into account the change in the risk strategy of the financial company to a more conservative one, assuming that borrowers with loans overdue with less than the usual number of installments are too risky during a crisis. In addition, the applicants, which were not classified as acceptable in the past are also considered as bad. This simplifies the problem and the analysis presented in this publication. It is the assumption that the rejected candidates are bad that leads to the higher values of the quality indicator of the models used. The Gini statistics is used to track quality, which reflects the discriminatory power of the models. Gini varies between 0 and 100%, with the Gini = 0 means zero discrimination, and Gini = 100% – full recognition of good and bad borrower (i.e. perfect model). In order to draw conclusions about the adaptation of scorecards, Gini is calculated on the basis data derived during the crisis, which was available at the time of modeling. Only for the first scorecard, which is built on data accumulated before the crisis, Gini is calculated with data before the changes. This makes it possible to account for the degree of deterioration of the next models. The changes in the credit score derived by the models, which is the estimated risk borne by the credit applicants, are also monitored.
The first study tracks the deterioration of scorecards generated by the standard model strategy relying on credit applications generated up to 1 year before the data export. This is caused by the need to have 12-month payments history to be able to create a proper model. This way, the standard approach would take into account the effect of the crisis 1 year after its occurrence.
The second study utilizes an extended observation interval compared to the standard one, and the latest applications with an available payment history. This allows to report the change in consumer behavior as quickly as possible.
The third study puts an additional weight to the new observations, which gives the models the potential to adapt quickly to any change in the market environment. Observations during the crisis are weighted more heavily in order to emphasize their effect on the final models.
The results of the research are presented in the table below.
As expected, the pandemic has caused deterioration in the quality of the scorecards in all of the studies as there is a considerable shift in consumer behavior. Only with the accumulation of new observations and the diminishing use of old data are the models expected to regain their ability to classify applicants correctly. The difference between the scenarios is mainly exhibited by the time, which each model needed in order to be able make proper classification. The results show that applying the standard data modeling approach (scenario 1) during the crisis, the ability of the scorecards to correctly discriminate the candidates naturally weakens, as the decline is about 33% compared to the quality of the scorecards assessed with pre-crisis data. In the second scenario, in which observations from the crisis period are introduced, with all observations equally weighted, Gini in the initial crisis model decreases sharply (as in the first scenario), but then gradually improves and the final deterioration is 15%. In the third scenario, where weights were introduced, the Gini indicator declines just 7% from the period before the crisis. This shows that the change in strategy ensures a rapid model adaptation to the new crisis situation. The change in risk level in each of the three scenarios can be measured by looking at the points that retirees get in each model. The higher a candidate’s score, the lower the risk according to the model. As already mentioned, pensioners’ income is not expected to be impacted materially by Covid-19 epidemic and therefore the relative credit risk of retirees should decrease compared to other consumer credit applicants. This is embedded in the data generated during the crisis period. The results show that in the first scenario, the variation in the points that retirees receive when applying for a loan fluctuates between 35 and 40, driven by the change in the set of variables which are part of the respective model. On the other hand, in the other two scenarios the increase is more stable, i.e. the more conservative lending will make each subsequent scorecard more favorable to this segment of the population. In the third scenario, the points more than double.
The latter score analysis directly presents the ability for faster scorecards adaptation to the new situation, achieved with the second and especially with the third modified modeling strategy.
The main conclusion of the study is that a suitable change in the way risk models are built allows them to achieve a rapid adaptation to the current customers behavior. In addition, by integrating updated business rules into the analytical application, the side effect of the deterioration in the quality of decision-making is further reduced.
JPMorgan CEO Jamie Dimon says in a letter to the shareholders he expects bad recession combined with some kind of stress similar to the global finance crisis of 2008. “As we have seen in past crises of this magnitude, there will come a time when we will look back and it will be clear how we — at all levels of society, government, business, health care systems, and civic and humanitarian organizations — could have been and will be better prepared to face emergencies of this scale.”
A4Lending is credit scoring solution developed by A4Everyone, which not just automates the decision making process for consumer finance businesses but also successfully provides a higher level of business security during the COVID-19 pandemic outbreak. By utilizing such a solution, consumer finance businesses are able to stand out from the competition in successfully managing the outbreak dynamics.
Visit us for more decision automation solutions by A4Everyone at https://www.a4everyone.com