News & Insights Archives

Imagine creating a model that reviews thousands of transactions every second and aims to flag the transactions related to money laundering. This is something that has become possible thanks to advances in artificial intelligence in recent years, and even appears to be a necessity for banks that are flooded with huge amounts of daily transactions. Hence the challenge of fighting money laundering is growing and RiskQuest is thrilled to be able to help the banks in the process of designing and creating Anti-Money Laundering (AML) models.

During the fight against money laundering other financial crimes like financing of terrorism, fraudulent transactions and corruption can also be captured. Even better, banks can decide to not only monitor money laundering but to also actively fight other financial crime. New models may be designed each aimed to detect a different specific fraudulent behavior. Examples are counter human trafficking, illegal wild life, corona fraud or healthcare fraud. In this blog the latter will serve as an example on how to develop such a transaction monitoring model and thus fight financial crime!

Healthcare fraud in the Netherlands

Before creating a model to fight a specific form of financial crime there must be sufficient reason to believe that this behavior occurs often and that this behavior could be captured by transactional data. Let’s tackle the first subject and dive into the health expenditures landscape of the Netherlands.

In 2015 the healthcare system of the Netherlands was decentralized, meaning that the responsibility of certain government duties were shifted from the government towards municipalities. Municipalities are among other things responsible for work and income, youth, long-term sick and elderly care. This decentralization was the starting signal for an increase in fraudulent health care declarations, often through healthcare providers (‘zorgbureaus’). These ‘providers’ for instance receive insurance money from home care declarations which is never provided, or they cheat with personal budget (PGB), the so-called PGB-fraud. Unfortunately the patient is often victimized, since they do not receive the care they are entitled to. However, sometimes the (fake) patient is engaged in the fraud as well.

Since the decentralization, it is fairly easy to become a healthcare provider nowadays; one does not need a diploma or patients to register. Subsequently it is very difficult for the municipalities to check whether everything is correct and non-fraudulent. Some even say that fraud is risk-free for such a healthcare provider at the moment and there is no sanction.

In the meantime, the Public Prosecutor (‘Openbaar Ministerie’ (OM)) estimates that hundreds of millions of Euros per year are used for fraud with healthcare funds. Partly because organized crime focuses on this ‘field’. The government and insurers are gradually seeing valuable healthcare funds disappear towards sports cars and hotel stays in Ibiza. This leak of hundreds of millions of Euros each year should be fixed!

Investigative journalists from Follow The Money (FTM) thought it was time for action. They built a model that checked annual financial statements of health care providers and filtered out those companies with high margins. Margins above 10% and high dividend payments were found to be indicators of fraud.

However, while developing our own tool RiskNavigator we found that using annual financial statements for analyses of a client is time-consuming, old fashioned and not up to date. Besides, it appears that annual financial statements are sensitive to fraud, which is a big problem in the healthcare fraud scenario. On the contrary, client transactions do not lie. Hence creating a model based on client transactions is more accurate and less fraud sensitive.

Reason enough to start making a transaction monitoring model, but how?

Time for some (model) action

When creating a transaction monitoring model, one should first decide on the scope of the clients. In the case of the healthcare fraud detection model one could for instance decide on only including those accounts that are registered under a certain SBI code, or maybe including those accounts that receive funding from insurance companies. If you stay focused on the behavior you would like to capture, a good prior selection may achieve a better performing model.

The second step in creating a transaction based model is finding the right data. Not only transactional data, but also client or company information, network information and other enrichments may be useful. Now is also the time to find your label (or target) information, in the case of the healthcare fraud detection model try to gather all historically known fraud cases.

Once you gathered all the useable sources it is time to create our data pipeline. In the data pipeline we combine all the input sources and create understandable features that our model can train on. Regarding the healthcare fraud detection model, good features to start with are for instance margins and dividend features, since FTM already found those are good risk indicators. Since the quantity of transactional data can be immense, one could also consider aggregation of transactions. Consider aggregating transactions for each client. This can for instance be a daily, weekly, monthly, quarterly or yearly aggregation.

Make sure to create your data pipeline in a structural way, such that development and production environment both can make use of the same underlying library. Git tooling is a DevOps add-on that may be convenient during development of the data pipeline. It allows the data pipeline to have multiple versions and allows contributors to work together on the same project. Even one step further is to train the final model on the production environment, such that no alignment test needs to be done between production and development environment.

Once you have your features, you need to combine them with historic label information and decide whether sufficient labels are available to perform supervised machine learning or that we should aim at unsupervised machine learning techniques. Sometimes explainable non-machine learning models or even clever business rules may be preferred to capture certain behavior. It may also be preferred to use a combination of techniques (an ensemble). This part of the project requires a lot of testing, for instance testing different features, different models, different hyperparameters and different scopes. To make a considered choice of the final model one can use MLflow. This tooling may be convenient during development, since it helps to register all findings and performance of tested models, but also during production, since it helps with registry of the model and automatic deployment.

After choosing a final model, one could also determine whether automatic retraining is preferred each time the model is executed. When new labels are available why not use them as soon as possible?

Detecting healthcare fraud

During model development analysts are often available to give feedback on the model performance. In the case of creating a healthcare fraud detection model one could try to retrieve specific feedback on the chosen client scope or certain features used in the model. Precious information is for instance behavior that analysts never find risky, or behavior that is always risky.

But we need to give something back to the analysts. Using machine learning techniques comes with a price. As previously discussed in the blog ‘What can Credit Risk modelling learn from Counter Financial Crime (Models) and vice versa?’ , the models themselves are somewhat of a ‘black box’. Hence we should make sure that our models and outcomes are understandable and explainable for analysts.

Documentation is a big part of this and spreading the word is as well: explaining how the models work, showing the performance of the models and provide insight in the use of models. Another important aspect is training the analyst in order to correctly interpret model alerts. Feature importance or transaction importance per alert may be a big help for the analyst.

Another thing that can’t be forgotten is bias quantification, for more information on this topic please read our previous blog ‘Are you mindful of discrimination in (your) models?’ and wait for our upcoming blog on Algorithmic Fairness.

Once we checked the performance of the healthcare fraud detection model, and made sure our outcomes are understandable and non-discriminating it is time to put the model into production. Let’s find those fraudsters!

Conclusion

In this blog we aimed to inform you on how to develop a transaction monitoring model. This blog used the example of the healthcare fraud detection model, however many other transaction models can follow this setup. At RiskQuest we have gained experience in creating models that fight financial crime and are fully aware of the importance of widespread knowledge sharing within an organization. We are open to share our knowledge and support you on the challenges you face in this area. Let us fight financial crime together!

For more information on this topic contact Emma Immink (Consultant), or Sven de Man (Partner)

If you want to join our RiskQuest team, please check our current job openings here

Counter Financial Crime (CFC) modelling is a relatively new field which has seen a lot of rapid progression in the past years. This rapid progression has been spurred on by the Dutch Central Bank (DNB) after it became clear that banks were insufficiently deterring and detecting financial crime, which resulted in severe fines from DNB. Banks have made a great effort to implement various ways of countering financial crime, most notably in anti-money laundering but also in terrorism financing and fraud. Advanced modelling techniques are used as creative methods to find any criminal activity.

Large progression in credit risk modelling has also been made. One of the main reasons for this progression is the introduction of many new regulatory standards concerning credit risk models published by the European Banking Authority (EBA), leading to updated requirements and methodologies. Focus of these new regulatory standards is on the robustness of required data, definitions and models. These have led to a need to update all credit risk models.

Both types of models differ significantly from each other in terms of purpose, but the challenges faced are similar. At RiskQuest, we have experience in both types of models. In the remainder of this blog, the lessons learned from credit risk models will be applied to Counter Financial Crime models and vice versa.

Machine learning methods

Counter financial crime models make use of advanced machine learning techniques with the goal to detect irregular or suspicious behavior that could indicate money laundering, whereas the industry standard in Credit Risk Modelling is the use of Ordinary Least Squares (OLS) regression and Logistic regression. The main benefit of the methods used in credit risk modelling is that the relationship between the dependent and the independent variables is very clear and can be shown with a single equation. Instead of these more traditional methods, machine learning techniques can be used for credit risk models. In general, these techniques can be used to achieve higher model performance on a chosen metric, since they can capture a greater variety of relationships using the same datapoints.

Unfortunately, there are two major downsides to the application of machine learning techniques for Credit Risk Modelling. The first downside is that it is questionable that machine learning models meet all regulatory requirements set and reviewed by EBA, ECB and AFM. This is related to the second downside, which is inherent to the use of machine learning techniques, is that machine learning models are often thought to be opaque in terms of their exact operation. While the model inputs and output are clearly defined, the models themselves are a ‘black box’.

This problem, however, does not just apply to credit risk modelling but is present for all uses of machine learning models. Hence, machine learning has seen significant progress in the development of explainability tools. These tools attempt to identify the relevant risk drivers as inferred by the machine learning models. The relevant risk drivers are those risk drivers that had the biggest impact on the model score. With these tools the relationship between the model score and the most relevant risk drivers becomes clearer. Counter Financial Crime modelers have used these tools to help analysts understand the models and how their output was generated.

The output of the machine learning models is used by analysts to determine whether the alerted behavior proves to be suspicious. In order for the analysts to fully understand why specific behavior is flagged, they need to understand how the model output is generated. These efforts to improve the understandability are very effective in reducing the operational costs and time spent on these alerts.

Similar techniques could also be used in credit risk modelling. One could apply these machine learning methods to a credit risk modelling problem and attempt to extract the most important relationships from this model. This relationship could then be recreated using Ordinary Least Squares regression or Logistic regression. This approach allows for full explainability of the model, with the added benefit that more complex relationships can be discovered through machine learning techniques. Additionally, machine learning methods can also be used to deploy a challenger model. This challenger model can be used to assess if the traditional credit risk models can still be improved, as well as provide a performance benchmark.

Use of transactional data

Transactional data plays a key role in the identification of money laundering, since cash flows reflect the behavior of the client. These cash flows can lead to a trigger of suspicious behavior. In credit risk models, most explanatory variables are related to behavior of the client with respect to delinquency and utilization of limits, or their financial statements. While delinquency and utilization information is mostly up-to-date, financial statements can be outdated, especially in crisis periods, where previous financial statements are not a good representation of current financial strength. This is where transactional data can be used to improve models.

Transactional data can be used to enrich credit risk models. The biggest advantage is that transactional data can be updated near-instantly and be reflected in the credit risk model scores. Furthermore, transactional data also contains information on debts, fixed expenses and provides insight in other cashflows that could be used as Significant Increased in Credit Risk (SICR) triggers, such as missed tax payments and interactions with debt collection agencies. Additionally, transactional data is of very high quality and highly standardized, making it ideal for incorporation in a credit risk model. The availability of transactional data, however, is limited to the accounts that are available to the creditor. Since large companies often have numerous accounts with several different institutions, transactional data should always be used in combination with a companies’ overall cashflows for a complete overview.

What can counter financial crime learn from credit risk?

While the counter financial crime sector is making great strides in terms of innovation and the use of sophisticated techniques, the model governance still has a lot of catching up to do. Although models in Counter Financial Crime have to go through numerous steps of model validation and other types of stakeholder approval, this entire process has not been refined yet. This is not a surprise, since CFC is a relatively new field and has a lot of freedom to introduce models and needs to find creative ways to detect a range of different kinds of financial crime. A natural consequence of this environment is that the model governance is constantly challenged to adapt to these new modelling approaches.

Despite the successful efforts to develop explainability tools for machine learning models, there are still challenges in the model governance process. Model validation often lacks a framework for the validation process for these kinds of models. As a result, the validation is a long process involving lots of communication back and forth with the model developers. This means a lot of operational time and money is wasted, which could be spent better if the process was more streamlined. These challenges are not only limited to model validation.

Requirements from DNB demand each of the bank’s business lines to cover and mitigate their own risks. From an organizational perspective, these business lines are rather detached from the modelers who build the machine learning models, not to mention the gap in knowledge. As modelers are continually developing and releasing new versions of their models, there are implications for the risk mitigation efforts of each individual business line. They are stakeholders of the models but are often not involved in the development process or not fully aware of the model specifics. This can result in a lot of unnecessary pushback and inefficiency. Creating an environment of successful cooperation is beneficial to the developers as well, since modelers will always need business knowledge to optimally counter financial crime.

Fortunately, within banks, there is already a field with a long history in model governance and stakeholder management. A field where each step of the process is guided by an extensive set of standards: credit risk. Credit risk is heavily regulated and for each step of the model governance process there is a clear set of predefined standards. Having a similar set of frameworks would be very beneficial for the counter financial crime department and increase its efficiency. Even though it is important that CFC keeps its innovative spirit and not be restricted too much, the knowledge and experience from the credit risk domain would still be transferrable and prove to be valuable for the model governance.

Conclusion

Although counter financial crime and credit risk are two fields that serve vastly different purposes, the challenges they face are similar. As the two fields excel in different aspects, it could be very beneficial to instigate more cooperation between the service lines. This will lead to increased knowledge sharing and improved model performance. Credit risk can learn from CFC to use machine learning techniques and apply transactional data in their models, while CFC can benefit from the extensive knowledge of credit risk experts in the model governance domain.

At RiskQuest we have gained experience in both fields and are fully aware of the importance of widespread knowledge sharing within an organization. We are open to share our knowledge and support you on the challenges you face in this area.

For more information on this topic contact Sjoerd Keijsers (Consultant), Mark Croes (Consultant) or Sven de Man (Partner)

If you want to join our RiskQuest team, please check our current job openings here

In recent years, new regulations with respect to credit risk models were published at a fast pace. Consistency between data sources, definitions and model methodologies were important topics for a number of guidelines from the regulators. This led to an intensified relation between the models for the two most important fields of application with respect to credit risk models: RWA calculation and provisioning.

Many banks decided to develop AIRB and IFRS9 models combined. Currently, the market seems to evolve into developing one probability of default (PD) base ranking model that will be calibrated differently for AIRB (through the cycle, TTC) and IFRS9 (point in time, PIT) purposes. Banks that decided to make two separate models for these purposes had a hard time to explain their choice to the regulator.

One could argue that a low-risk client under AIRB will probably be a low-risk client under IFRS9 as well. Therefore, multiple ranking models, with, likely, different ranking outcomes would lead to counterintuitive results. However, is the combination of the development of both models as straightforward as it seems?

This blog will focus on one specific inconsistency between the AIRB and IFRS9 requirements that led to challenges during our latest model development and many sleepless nights for me personally.

On the one hand we have the development of the IFRS9 multi-year PD model. The model is required to estimate the lifetime PD of the loan given the current and future state of the macroeconomy. This is a complex requirement, especially given limited data history. Therefore, it is market practice to calculate one-year PIT migration matrices and multiply them to obtain the multi-year PIT PD. However, the matrices should be Markov in order to achieve correct multi-year horizon estimates. The Markov property implies, in this context, that the probabilities of a set of possible rating migrations only depends on the current rating and is therefore independent on the full rating migration history.

On the other hand, there is the heterogeneity requirement for pooling of the AIRB models. This means that all pools should have a significantly different observed PD. However, the significant difference will be tested on the AIRB horizon, which is one year. This mismatch in horizon is exactly where the shoe pinches. Two facilities with the same one-year PD could have a different multi-year PD. However, due to the requirement of heterogeneity they will be merged into one pool, which leads to non-Markov properties of the migration matrix. This non-Markov behavior leads to problems in the development of the IFRS9 model.

The description of the inconsistency can be illustrated by the following example. Assume the following start population and the (true) migration matrix:

Note that this is in line with reality where most of the clients are healthy and have a low PD. Note that these numbers have been chosen to minimize the period of convergence of the observed default frequency (ODF), see the figure below:

With this start population we can calculate the number of defaults that we will observe over a five years period. In the first year 19 defaults will be observed. The distribution of the facilities over the ratings can be calculated in the same way. In year 2, rounded to an integer, 20 defaults will be observed and in years 3-5 every year 18 defaults will be observed (also rounded to an integer). In total we observe 93 defaults over the first five years. This number can also be obtained by multiplying the migration matrix 5 times with itself:

The expected number of defaults in five years is calculated as: 1000 × 0.071+120 × 0.134 + 60 × 0.103 = 93(rounded). Note that this three-state process is observed and will be estimated if all information is available.

However, the one-year default probability of rating 2 and rating 3 are equal. Therefore, based on the heterogeneity requirement from AIRB, these ratings will be merged into one rating. This leads to the following matrix (given the start population of rating 1, rating 2, rating 3 and default [1000,120,60,0]):

If we look at the number of estimated defaults in the first and the second year, we see an exact match with the three-state matrix (both estimated and observed). However, from the third year onwards, the number of estimated default for the two-state matrix is higher. After 5 years we estimate 99 defaults, instead of 93. Therefore, merging these two states leads to non-Markov behavior.

It is interesting to see that a one state model would be Markov, due to the chosen equilibrium population. Note that this would not be the case in real life, because there would be no equilibrium. However, usually the default probabilities are linked to the macroeconomy. Therefore, the one state model can be used for multi-year PD, without failing the Markov test. Note that a ranking model, with only one rating would be Markov, but would not meet any other requierements and is therefore not a solution.

Note that the example seems odd, but in practice these types of state traveling happen a lot. For example if there are separate ratings for facilities that currently have arrears, facilities that had arrears in the past 12 months and clients that did not have arrears in the past 12 months. In reality the observations are a combination of all kinds of patterns.

Preferably, as a modeler I would like to have:

The same ranking model for both AIRB and IFRS9
Heterogenic buckets under AIRB requirements
Pass the Markov test on the migration matrix for IFRS9

From the example, it can be concluded that it is impossible to fulfill all three requirements. Therefore, we have to choose two. In our vision it depends on the specific situation at the client what would be the best choice. It is market practice to go for 1 and 2 and apply scaling to compensate for non-Markov matrices. This could be a good approach given that multi-year IFRS9 is only used for stage 2 facilities. Furthermore, the majority of the ECL will be based on the risk in the first years where the deviation is small and we are able to improve by scaling. However, if the bank uses the IFRS9 models also for extensive stress testing then not only the default probability is relevant, but also the performing-to-performing migrations, the bank might want to choose another strategy.

We as RiskQuest are always open to discuss the best strategy for your bank.

For more information on this topic contact Janneke van Schijndel (Manager) or Sven de Man (Partner)

If you want to join our RiskQuest team, please check our current job openings here

RiskNavigator powered by PSD2 (RiskQuest & Invers)

What is RiskNavigator?

RiskNavigator is a risk management solution that enables credit issuers to efficiently and effectively assess loan applicants
It provides credit issuers with a complete credit risk management dashboard based on real-time banking transactional data powered by PSD2
It is an integrated end-to-end solution that can be adopted instantly

What can RiskNavigator do for you?

We bring credit issuers and customers together. RiskNavigator delivers an integrated and customized credit risk management dashboard (e.g. customer credit-profile) by using enriched banking transactional data powered by PSD2 (supplied by Invers, a FinTech enabler with a PSD2 license). The tool operates in the following four key steps:

Retrieve all banking transactional data once customers grant RiskNavigator access to their personal and / or company-specific financial transaction history
Enrich banking transactions with predefined classifications and categories
Profile customer credit status (e.g. key metrics, lending capacity and early warning signals)
Prepare extensive financial analysis in a user-friendly risk management dashboard that is ready-to-use

Why should you use RiskNavigator ?

RiskNavigator is developed by RiskQuest, a consultancy firm that specializes in risk management solutions in the financial services industry and has extensive knowledge in the credit risk domain.

RiskNavigator is a cloud based ready-to-use solution that can be adopted instantly
RiskNavigator helps credit issuers not only to speed up credit approval processes but also to gain deeper and more meaningful insights into the customer’s credit status in a user-friendly and intuitive risk management dashboard
RiskNavigator’s dashboard provides detailed credit insights such as repayment capacity, lending capacity, debt-servicing capacity, early warning signals and in due course the credit risk score. These credit insights are all based on customer’s banking transactional data
RiskNavigator’s dashboard will enhance and complement the existing traditional risk assessment based on annual report and its respective financial ratios
RiskNavigator provides ongoing and real-time risk monitoring functionality via early warning signals. It enables the credit issuers to monitor and update their credit assessments in real time for current loans

Who are RiskQuest and Invers?

RiskQuest (developer of RiskNavigator)

RiskQuest is a consultancy firm in the Netherlands that specializes risk management solutions in the financial service industry. RiskQuest has over 15 years of extensive knowledge in credit risk models and credit assessments that help the financial institutions to develop its risk strategy, define suitable risk appetite and its respective credit solutions. Risk models and machine learning algorithms developed by RiskQuest have helped financial institutions gain a deeper and more meaningful understanding of their key risk factors that are embedded in the business portfolios. This provides the financial institutions with guidance for decision making.

Invers (provider of the enriched PSD2 transaction data)

Invers has 25 years of experience in customer insights in the field of financial situation of the individual consumers and SMEs. Invers is specialized in customer insights by analyzing cash flows patterns of individual consumers and SMEs through three core areas: 1) retrieving bank transactions, 2) enriching banking transactions with predefined categories, 3) creating descriptive financial analyses of individual consumers and SMEs. These analyses are used by banks, lenders, financial service providers, e-commerce and software suppliers. They integrate Invers’s web services via API’s with their own services for its respective daily operations, client servicing and financial institutions’ financial analysis.

RiskNavigator contact details

Hans Heintz
RiskQuest B.V.
Herengracht 495
1017 BT Amsterdam
info@RiskQuest.com
Office number: +31 20 693 29 48
Mobile number: +31 6 81 50 90 88

In 2009, the Swedish Central Bank (SCB) introduced negative interest rates for the first time in history (Christophe Madaschi, 2017). The European Central Bank (ECB) followed in 2014 and introduced negative interest rates to boost the economy during the aftermath of the global economic crisis. Since changes in interest rates affect a bank’s profitability, it is of utmost importance to have well performing interest rate models. However, not all interest rate models are able to deal with negative interest rates. In this blog, we stress the importance of good interest rate models, review various candidate models and provide solutions in case traditional models do no longer function with negative interest rates.

Negative interest rates

The idea behind the introduction of negative interest rates was twofold. First, a negative interest rate stimulates economic consumption, since consumers will be less inclined to save money. Secondly, funding for companies becomes cheaper and therefore more attractive, and will stimulate economic production. Combined, these two effects should lead to an economic growth stimulus.

Importance of appropriate interest rate models

For all banks, but in particular for banks with a lot of retail deposits, interest rate models are needed that can deal with a negative interest rate policy (NIRP). Banks with such models can create a competitive advantage because appropriate interest rate models contribute to sound risk management, and also because they are able to better adapt their business to changes in the interest rate environment.

The importance of appropriate interest rate models is also recognized by regulators. In July 2018, EBA guidelines on IRRBB (EBA/GL/2018/02) were finalized as part of Pillar II of Basel’s capital framework. These guidelines discuss Interest Rate Risk in the Banking Book (IRRBB), i.e. the risk that adverse movements in interest rates affect the earnings and the economic value of an institution. The guidelines state that in low interest rate environments, institutions should also consider negative interest rate scenarios. Therefore, it is not only the competitive advantage aspect, but also the regulatory requirements that stress the need for interest rate models that can deal with a NIRP.

It is important that the interest rate models are aligned with its purpose. For the IRRBB, different interest rate models are needed than for example the trading book. For the trading book it might be preferred to use a more complex interest rate model in order to accurately capture the volatility smile. The drawback of these complex models is that more simulations are needed to compute the present value of derivatives. When institutions hedge on a frequent basis, the question arises whether these complex models are the most convenient. A less accurate model, which is easier to use, can be preferred in this situation.

What interest rate models are commonly used and what are their limitations in a negative interest rate environment?

To forecast interest rates, in practice often short-rate models are used. Short-rate models are mathematical models that model the future evolution of the short-rate over time. However, in the current negative interest rate environment, some of these models lose their functionality or their predictive power. The specifications of these models can for example be restricted to allowing only positive current short-rates or assume a log-normal distribution for the interest rates, which results in a lower bounded interest rate of 0%. These model characteristics prevent the interest rate models to predict negative interest rates in a negative interest rate environment. However, some of these restrictions can be circumvented, as we will discuss.

One of the short-rate models that has problems with modelling the evolution of interest rates in a negative interest rate environment, is the Cox-Ingersoll-Ross (CIR) model. The CIR model uses the square root of the current short rate as input to forecast interest rates. Although imaginary interest rates sound like an interesting concept, practically speaking the square root of negative short rates is not defined.

The horizon for which short-rate models provide accurate predictions, can be relatively short. Therefore, for predictions on longer horizons, other types of interest rate models might be preferred. An often preferred model is the LIBOR Market Model[1], which models a set of forward rates for different periods that are directly observable in the market. A downside of the LIBOR Market Model is that it assumes that the forward rates are log-normally distributed. The log-normal distribution restricts the forecasted interest rates to be positive, resulting in the fact that negative interest rates can’t be predicted by this model.

What adjustments can be made to the models to make them suitable under a NIRP?

Some interest rate models, such as the LIBOR Market Model and the Cox-Ingersoll-Ross (CIR) model, are not able to forecast negative interest rates. For these interest rate models, the model specification needs to be adjusted before they can be used to properly model the interest rate movements in a negative interest rate environment.

The non-negativity limitation of the LIBOR Market Model can be solved by introducing a shift in the log-normal distribution, which allows for negative interest rates. For the CIR model an additional parameter can be added to the model specification to ensure that the current short-term interest rate is shifted to a positive value, such that the square-root transformation is defined (Orlando, Mininni & Bufalo, 2019).

A potential drawback of this adjustment for the LIBOR Market Model is that, since the log-normal distribution is shifted, the distribution is still capped by a negative boundary. How to implement the shift also introduces expert judgement and possibly frequent recalibrations in a declining interest rate environment. This also is a drawback for this adjustment related to the CIR model, since the shift must be large enough to ensure that the current short rate becomes positive and in a declining interest rate environment this shift needs to be constantly larger.

What model alternatives are better equipped to deal with a NIRP?

A popular choice in the current low interest rate environment is to use shadow rate term structure models. These models use a shadow short rate rather than the actual short rate . The actual short rate is defined as:

that is, the short rate equals the shadow rate, if this is above the lower bound , while the actual short rate remains at the bound if the shadow rate is below the bound. In order to account for negative interest rates, the lower bound can be set to a negative time-dependent value. Shadow rate models can capture the decline in yield volatility far better than a popular benchmark model that ignores the presence of a lower bound. (Lemke & Vladu, 2017).

Other popular choices are regime switching interest rate models. By means of regime switching techniques, different models are used depending on the environment at hand: negative, low, normal or high interest rates. This has the advantage of always using a suitable model, whatever the level of the interest rate may be. A drawback of regime switching models is that the model increases in terms of complexity. Traditional short rate models are easier to interpret and are often arbitrage free and provide closed form formulas. The increase in complexity of the regime switching interest rate models makes it more difficult to calibrate the model on the market data.

Conclusion

The NIRP affects the profitability of a bank, and therefore suitable IR models are needed that can deal with this situation. Not only is this desired from a competitive advantage perspective, but it is also a regulatory requirement. While some models are still valid, other models need adjustments either by changing their specification, or combining models into one model using regime switching techniques.

At RiskQuest, we are well aware of the effects of the NIRP, both on banks and on interest rate models. We are also mindful of the importance to align the complexity of the model with its purpose, where complexity should be balanced with frequency of use. We are ready to advise and provide the best possible IR models for each situation and tailor this for specific needs.

For more information about IRRBB, negative interest rates or specific interest rate models, please contact Lino Kragtwijk (consultant) or Nick Verbaas (consultant)

If you want to join our RiskQuest team, please check our current job openings here

[1] Note that the LIBOR Market Model can also be used to model other interest rates than the LIBOR

The share of non-banking financing options among small and medium sized-enterprises (SMEs) in the Netherlands has increased significantly over the past years. With this increase, the importance of up-to-date and efficient customer’s risk assessments is also rising. Such risk assessments can be performed in multiple ways. As of January 2018, financing providers may also consult the Payment Services Directive 2 (PSD2), a revision of the European guidelines adopted in 2007 by the European Commission. Under PSD2, customers can grant access to their transaction data for banking as well as non-banking institutions. The data remains the property of the customer, which means it may be shared with other parties at the same time. This should, among others, improve competition, enhance innovations and create a level playing field in the European payment market. In this blog, we will zoom in on the trends in non-traditional financing possibilities and how PSD2 can help non-banking financing providers to properly assess their risk position in relation to their profitability.

How does PSD2 work?

With the introduction of PSD2, customers can give (one-time) access for a maximum of 90 days into their personal or company-specific transaction history to banks as well as third parties with a DNB permit. When the 90 day term has expired, the third party access is automatically revised unless the customer specifically agrees to extend. If and only if such access is granted by the customer to an external company, the company is able to provide its services to the customer. Typical services that may be offered by a third party are cashless payment services, aggregations and visualizations of historical transactions to indicate expense or income patterns for financial planning and building quantitative or qualitative models to indicate customer risk levels. At the moment access is granted, the external company can get full insight in the historical transactions and extract typical information like account balances, potential contacts with debt collectors, outstanding active debt levels, etc. The available historical transaction period varies per bank but is limited to the most recent two years.

Market trends in SME financing

Over recent years a significant shift is observed from banking (i.e. traditional) to non-banking (i.e. alternative) financing in the SME segment. This shift is mainly observed for financing up to one million euros and is largely impacted by the possibilities provided under PSD2, where alternative financing providers can easily enter the financing market. Stichting MKB Financiering (SMF) projects that this trend continues over the upcoming years. According to their projection as shown in Figure 1, it is expected that alternative financing will become the dominant option for SMEs with financing needs up to one million euros as of 2023.

Figure 1: Projections of traditional versus alternative financing over time for financing needs up to one million euros. Source: stichtingmkbfinanciering.nl*

Traditional banking financing options have a long history and are a well-known financing form for the SME segment, whereas the alternative financing options are somewhat newer. Among the realm of alternative financing options one could consider crowdfunding, credit unions, direct lending, SME exchanges, real estate financing, lease and factoring. When visualizing the developments of alternative financing over recent years, one may conclude (from Figure 2) that a steep positive trend is observed over time for all of these alternative financing forms.

Figure 2: Alternative non-banking financing options presented in outstanding financing over time. Source: stichtingmkbfinanciering.nl*

Advantages of PSD2

As there is clear evidence of an increased market for alternative financing demands in the near future, there is also the need to efficiently facilitate the financing process. The latter requires a thorough and quick assessment of the applicant’s risk level. Currently, it is observed in the market that financing companies are not using all available information to its fullest by only assessing for example the Bureau Krediet Registratie (BKR), outdated balance sheets, buying external data or even manual checks of transactions. These are ways of working that are no longer sustainable and profitable as the market for alternative financing keeps on growing. Using PSD2 in this context has a large variety of advantages:

Up-to-date insights: it enables the financing company to get accurate and up-to-date insights in the current financial state of the financing applicant. In the current setting with Corona for example, it would show whether the applicant used one of the support measures offered by the government, which the financing company would have never spotted when solely looking at the balance sheets.
Standardized data: the data provided is very standardized and transparent and hence decreases the possibilities of fraudulent applicant behavior (as long as you know how to indicate such behavior).
Efficient process: when PSD2 access is provided by the applicant and further semi-automated processes are already in place, the process is very efficient and clear insights could be obtained on the spot.

In addition to alternative financing providers, the usage of PSD2 might also be of interest to other parties like banks for their mortgage/loan acceptance as well as monitoring processes and insurance companies to assess their client’s behavior to identify moral hazard issues and to provide customer-specific insurance premiums. Finally PSD2 can be used to look for fraudulent activities on a large scale by means of highly-technical machine learning models.

The new way of working with PSD2 provides the market with endless opportunities to get insights into the transaction history of customers. In a way it ensures that all ingredients for a proper risk assessment become available to everyone who gets explicit access from the customer. However, to properly use the ingredients, additional treatments are required that are not always straightforward. Two large workflows that may add value to the ingredients are 1) processing, aggregation, categorization and enrichments of the transactions and 2) creating visualization and monitoring products as well as quantitative or qualitative risk assessment models (e.g. debt capacity analysis).

Added value RiskQuest

RiskQuest focusses on adding value in the latter workflow. We can provide an interactive Dashboard that has the sole purpose of getting quick and clear insights in a customer’s transaction history, obtained through practical experience, to better organize the application acceptance process. This Dashboard provides insightful and meaningful metrics as well as the opportunity to quickly see any indications of, for example, missed tax payments and debt collector or regular debt commitments. In addition to this interactive Dashboard, RiskQuest is a market leader in building tailormade quantitative risk models based on available datasets. This enables us to create models together with our clients that could help in further automation of the decision making process.

*The numbers shown throughout the blog are for illustrational purposes to show the general market trend. No reliance can be placed on these numbers.

For more information on this topic contact Rick Stuhmer (senior consultant) or Hans Heintz (partner)

If you want to join our RiskQuest team, please check our current job openings here

With Covid-19 financial institutions are once again at the center of attention. In sharp contrast with the great recession banks are now considered part of the solution. The Corona Monitor by the Nederlandse Vereniging voor Banken (NVB) shows that 26.500 commercial clients have been provided with new loans, 5.800 with state-guaranteed facilities and 129.000 commercial clients are making use of a payment holiday. These numbers reflect the strong capital buffers that banks have been building up since 2008. Like the Dutch Delta works require continuous maintenance to protect the Netherlands against extreme floods, so do banks need to continuously maintain their credit risk models to be able to actively monitor the adequacy of their buffers. In this blog, we will zoom in on how Covid-19 may impact these models, with a particular focus on probability of default (PD) models.

PD-models form one of the key ingredients to estimate the capital that a bank should hold. On the surface a PD-model predicts the probability that a client will go into default within a certain period (typically a year). However, underlying this prediction is a ranking of clients from worst to best credit rating. Essentially such a ranking model assigns each client a qualitative risk profile. For example, such a relatively simple ranking may be obtained by assigning clients to three different buckets, corresponding to a low, medium, and high credit risk. By considering the historic default frequencies per bucket, one can then transform such a qualitative rating to an estimate for the probability of default. This is schematically illustrated in the figure below.

While Newton’s laws will still be valid tomorrow, the behavior of both companies and clients will change both due to external and internal factors. As a result, the PD model of last year, may not be fit for the next year. This may be especially true in the present situation as the ‘new normal’ may be very different from the ‘old normal’. At least, so it seems.

When one is purely interested in the ranking of clients from bad to good, the impact may actually be relatively mild. To a large extent this will depend on the risk drivers that underpin the ranking. Typical risk drivers may include a client’s payment behavior such as credit utilization, the number of arrears, or refused transactions, but also other parameters such as the age of a company. Now the good thing is that even though the precise relationship between those input parameters and the probability of default may have changed due to Covid-19, the nature of the relationship has not. Still it seems safe to assume that a client with maximal credit utilization is more likely to go into default than a client that has not used its limit at all (on average that is). In other words, as long as the underlying risk drivers are sensible, then so will be the resulting ranking.

On the other hand, the present situation is so different from the past that it doesn’t seem reasonable that the probability of default estimates for each bucket are still correct. By the very nature of a probabilistic prediction, some disagreement between the estimate and the observed default frequencies is to be expected. However, when the observed discrepancy becomes very unlikely (say that such an extreme discrepancy can only be attributed true randomness with a chance of 5%) then it is typically decided that the PDs assigned to the various buckets need to be recalibrated. To which extent this is necessary will crucially depend on whether one is dealing with a point-in-time (PIT) estimate for the probability of default, or a through-the-cycle (TTC) estimate. As TTC-models assign to each bucket the observed default frequency averaged over a full credit cycle, the impact of Covid-19 will be relatively low as the current observations are damped by the past. Instead, for PIT-models the situation will be way more complex. Both the macro-economy and the correlation between the default rate and the macro-economy can change. Especially, the government aid could lead to less defaults or a delay of default observations.

Covid-19 will certainly have a big impact on banks, however, we expect that the PD-models for estimating capital buffers will be relatively robust due to their inherent structure. Now, this message should certainly not be taken as a recommendation for banks to sit back and relax when it comes to models. To the contrary, we believe that models can play a pivotal role in minimizing future credit losses. In addition to the reassessment of models used for provisioning; models can be used for other purposes as well. For example, due to Covid-19 the number of clients at FR&R is likely to increase, and may potentially overwhelm its capacity. In such a case, models can be used to perform the financial analogue of triage. Furthermore, we believe that banks may benefit from models that explicitly account for current support measures in order to get an understanding of how clients would fare in the absence of these measures. Such a model could then be used in the pricing for new loans as one day the support measures will stop.

For more information on this topic contact Guido van Miert (Consultant) or Janneke van Schijndel (Manager)

If you want to join our RiskQuest team, please check our current job openings here

Financial institutions cannot simply accept every application for e.g. a mortgage loan or non-life insurance product. To avoid harmful commitments for both bank and client, acceptance is based on selection criteria, often with the help of mathematical models. In these processes, discrimination should be prevented, as we consider it unethical and it is also strictly prohibited. Until recently, it was relatively easy to prevent model discrimination: by omitting variables mentioned in Article 1 of the Dutch Constitution from the model, discrimination could be avoided.

However, the use of advanced machine learning algorithms is becoming more widespread. While such models achieve high performance e.g. as selection criteria, their complexity reduces their transparency. In becoming somewhat more of a ‘black box’, the chance of ‘latent’ discrimination increases. For example, neural networks with different layers can identify very complex patterns.

In order not to make this blog post overly complex, regulations other than Article 1 of the Dutch Constitution (such as the GDPR) are not taken into account. Furthermore, several methods are known in the literature to guarantee ‘algorithmic fairness’. These will be discussed in a future blog.

Discrimination

Discrimination is (pre)judging an individual based on the group to which the individual belongs. There should be equal opportunities for all without bias.

Mind you, there may also be latent discrimination. For example, pregnancy is inextricably linked to the female sex (a direct relationship, see figure below) and shopping in a halal supermarket is strongly linked to religion (a derived characteristic in the figure). It shows that latent discrimination can appear in unexpected forms.

It seems that a simple way to avoid discrimination is to exclude potentially directly or indirectly discriminating variables from the model. However, advanced machine learning algorithms can find very complex patterns which could result in hidden, or latent, discrimination.

Input control

While discrimination is obviously prohibited, an individual may be subject to a decision on the basis of his or her behavior insofar as this involves a free choice for which they could be held accountable. A free choice is here defined as a choice which cannot be traced back to protected characteristics.

When deciding whether to accept a new client, it is important to motivate which criteria are relevant, i.e. show a causal (and not just statistical) relationship with the purpose of the decision. For example, the historical payment behavior of an applicant is very relevant to a mortgage application.

Thus, two dimensions are important in determining variables that form the input of a model: voluntariness and relevance.

Below are some examples of the different groups of variables:

A. Not a free choice characteristic that is relevant:
The sex of an applicant for a car insurance policy could potentially be relevant for predicting future car damage[1], but this is clearly not allowed.

B. Partly free choice characteristic that is relevant:
This category is most interesting for the discussion, because the variable is considered relevant and in some sense as a behavioral characteristic of the individual applicant. Consider, for example, the street where someone lives. This often appears to be relevant, but at the same time may betray a certain cultural background of the applicants. It is not always a completely free choice.

C. Free choice characteristic that is relevant
The fact that someone is a member of a mountaineering association can potentially be relevant for the purpose of a decision. Furthermore, such a membership can be labeled as a free choice. The same goes for whether someone smokes. That is a free choice that is very relevant for a life insurance policy.

D. Free choice characteristic that is irrelevant
The number of hours someone spends watching TV per week is a completely free choice. However, it does not seem very relevant to assessing his or her behavior as a motorist. The motto would therefore be to omit such a variable, since a decision based on it would not be justifiable.

E. Not a free choice characteristic that is irrelevant
Variables in this category are not allowed nor are they relevant for the purpose of the selection. For example, political affiliation does not seem to be causally related to the risk of accidents. Apart from the legal prohibition to include these variables, there are also no justified grounds for inclusion, since we cannot argue for a causal relationship.

In fact, variables in category B should lead to the most discussion, because they are considered relevant and are (partly) the result of free choice and partly traceable to a protected characteristic from Article 1. Is the street where someone lives a completely free choice? Or the country where people go on vacation? Is it fair to hold someone accountable for these choices?

Output of the model

One could argue that if protected or non-behavioral variables are excluded from the model a priori, the risk of discrimination is smaller. After all, what does not go in, cannot come out. However, this is not necessarily true. Deep learning algorithms can find very complex patterns that may still correlate with protected characteristics.

It could therefore happen, even if all protected data is excluded, that the use of behavioral variables still leads to a model which discriminates. For examples, suppose all Cretans lie[2] about their income statement. This could result in a mortgage acceptance model which rejects all applications of Cretans.

If the model works correctly and all Cretans lie, there could be justified grounds to reject their applications. At the same time, this result can be undesirable and call for a debate about algorithmic fairness. Also, the bank could suffer reputational damage, regardless of justifiability.

Conclusion

Financial institutions may assess applications based on the individual behavior of the applicant and possibly discriminate. The central question here is to what extent behavioral variables are the result of completely free choice and to what extent these can be traced down to characteristics as referred to in Article 1 of the Dutch Constitution.

The discussion about algorithmic fairness and which variables can lead to discrimination is certainly not black and white, but gray. We therefore encourage an industry-wide discussion on this topic to establish what is permissible and required.

The discussion is becoming increasingly relevant with the introduction of advanced machine learning algorithms, where model transparency can decline. It is therefore very important that financial institutions introduce proper monitoring and controls to prevent discrimination in their models.

RiskQuest has experience in quantifying and counteracting discrimination in models. If you would like to receive more information, please contact us.

Please note that at RiskQuest, we do not tolerate racism, discrimination, or violence of any kind. We prioritize and value the diversity of our people, and we are committed to fostering a culture where everyone feels they belong, while providing equal access to opportunity.

For more information on this topic contact Hans Heintz (partner) or Lars Haringa (senior consultant)

If you want to join our RiskQuest team, please check our current job openings here

[1] Research finds that men are more often involved in traffic accidents.

[2] From Epimenides paradox: “All Cretans lie” –said the Cretan

Machine learning and artificial intelligence are regularly in the news: from self-driving cars to clever algorithms that detect fraudulent transactions. However, within risk modelling machine learning has yet to make an impact. This is partly due to the strict regulations these models need to comply with and possibly also due to unfamiliarity with these types of models. Nonetheless, the strong performance of machine learning models cannot be ignored. At RiskQuest we believe these new techniques will be part and parcel of risk modelling within a few years.

Machine learning models are incredibly powerful, primarily because of their flexibility compared to standard statistical techniques. Machine learning is better at handling:

High number of input variables,
Non-monotonic data patterns and
Varying data types

These properties make machine learning models highly suitable for dealing with new data sources, such as PSD2 (sharing bank transaction data). For such complex, unstructured data we expect machine learning models to make the largest impact. Nevertheless, it can be a risky endeavor to develop innovative models because it is not always clear whether these techniques meet the broad range of regulations set and/or checked by the AFM, EBA and ECB.

A case in point is the Dutch allowance affair (‘toeslagenaffaire’) of the Dutch tax authority (‘De Belastingdienst’) in which many parents were erroneously forced to return their allowances due to biases in the machine learning models. Another reason why standard statistical techniques will never be entirely discarded within risk modelling is that they allow users full understanding of the relationship between each risk driver in the model and the impact on risk. Something that is difficult to achieve using machine learning.

The trade-off between model performance and transparency is a balancing act. Introducing machine learning can lead to a competitive advantage. However, fully switching to machine learning is not compatible with risk management’s needs given machine learning’s lack of transparency and the difficulty of understanding the relationship between the different risk drivers in these models. We envision hybrid models combining both existing statistical techniques and machine learning. For instance, machine learning can support model development by quickly identifying patterns over time and by pre-selecting relevant risk drivers that experts can accept/reject on economic grounds. During validation, machine learning can be very useful in benchmarking the explanatory power of the data.

Machine learning is here to stay. As the toeslagenaffaire committee stated: a system like De Belastingdienst cannot function without computers, algorithms, and automated processes. Another example is the discussion paper published by the Dutch central bank (DNB) last year on the “General principles for the use of Artificial Intelligence in the financial sector”, showing that the DNB is seriously preparing for machine learning and artificial intelligence applications. As an industry, we need to adapt to these changing circumstances. Investing time and resources in machine learning models will pay out when machine learning models become market practice in risk modelling.

How soon these changes will happen remains uncertain. At RiskQuest, we are well-prepared and ready to advise on current market practice models, as well as to open up the advantages machine learning can offer you.

For more information on this topic contact Tom Siebring (risk consultant) or Sven de Man (partner)