AI in FIs: Foundations of Machine Learning in Financial Risk

This blog series will focus on Financial Institutions as a premier business use case for Machine Learning through the lens of financial risk.

Today, opportunities exist for professionals to delegate time-intensive, dense, and complex tasks to machines. Machine Learning (ML) has the ability to automate Artificial Intelligence (AI) and is becoming much more robust as technological advances ease and lessen resource constraints.

Financial Institutions (FI) are constantly under pressure to keep up with evolving technology and regulatory requirements. Compared to what has been used in the past, modern tools have become more user-friendly and flexible; they are also easily integrated with existing systems. This evolution is enabling advanced tools such as ML to regain relevance across industries, including finance.

So, how does ML work? Imagine someone is learning to throw a football. Over time, the to-be quarterback is trained to understand how to adjust the speed of the ball, the strength of the throw, and the path of trajectory to meet the expected routes of the receivers. In a similar way, machines are trained to perform a specific task, such as clustering, by means of an algorithm. Just as the quarterback is trained by a coach, a machine learns to perform a specific task from a ML algorithm. This expands the possibilities for ways technology can be used to add value to the business.

What does this mean for FIs? The benefit of ML is that value can be added in areas where efficiency and accuracy are most critical.  To accomplish this, the company aligns these four components: data, applications, infrastructure, and business needs.

 

A flow chart showing the relationship between technology and data

 

The level of data maturity of FIs determines their capacity for effectively utilizing both structured and unstructured data. A well-established data governance framework lays the foundation for proper use of data for a company. Once their structured data is effectively governed, sourced, analyzed, and managed, they can then employ more advanced tools such as ML to supplement their internal operations. Unstructured data can also be used, but the company must first harness the tools and computing power capable of handling it.

Many companies are turning to cloud computing for their business-as-usual processes and for deploying machine learning. There are options for hosting cloud computing either on-premises or with public cloud services, but these are a matter of preference. Either method provides scalable computing power, which is essential when using ML algorithms to unlock the potential value that massive amounts of data provides.

Interested in reading more? Subscribe to the FRG blog to keep up with AI in FIs.

Hannah Wiser is an assistant consultant with FRG. After graduating with her Master’s in Quantitative Economics and Econometrics from East Carolina University in 2019, she joined FRG and has worked on projects focusing on technical communication and data governance.

 

 

 

Is a Pandemic Scenario Just a Recession Scenario?



Recently, I wrote about how a pandemic might be a useful scenario to have for scenario analysis. As I thought about how I might design such a scenario I considered: should I assume a global recession for the pandemic scenario?

A pandemic, by definition, is an outbreak of a disease that affects people around the globe. Therefore, it is reasonable to think that it would slow the flow of goods and services through the world. Repercussions would be felt everywhere – from businesses reliant on tourism and travel to companies dependent on products manufactured in countries hit the hardest.

For an initial pass, using a recession seems sensible. However, I believe this “shortcut” omits a key trait needed for scenario development: creativity.

The best scenarios, I find, come from brainstorming sessions. These sessions allow challenges to be made to status quo and preconceptions. They also help identify risk and opportunity.

To immediately consider a recession scenario as “the pandemic scenario,” then, might not be advantageous in the long run.

As an exercise, I challenged myself to come up with questions that aren’t immediately answered when assuming a generic recession. Some that I asked were:

  • How do customers use my business? Do they need to be physically present to purchase my goods or use my services?
  • How will my business be impacted if my employees are not able to come into work?
  • What will happen to my business if there is a temporary shortage of a product I need? What will happen if there is a drawn-out shortage?
  • How dependent is my business on goods and services provided by other countries? Do these countries have processes in place to contain or slow down the spread of the disease?
  • Does my business reside in a region of the country that makes it more susceptible to the impact of a pandemic (e.g., ports, major airports, large manufacturing)?
  • How are my products and services used by other countries?
  • How can my company use technology to mitigate the impacts of a pandemic?
  • Is there a difference in the impact to my company if the pandemic is slow moving versus fast moving?

These are just a few of the many questions to consider for this analysis. Ultimately, the choice of whether to use a recession or not rests with the scenario development committee. To make the most informed decision, I would urge the committee to make questions like these a part of the discussion rather than taking the “shortcut” approach.

Jonathan Leonardelli, FRM, Director of Business Analytics for FRG, leads the group responsible for model development, data science, documentation, testing, and training. He has over 15 years’ experience in the area of financial risk.


RELATED:

Do You Have a Pandemic Scenario?

VOR Scenario Builder: Obtaining Business Insight From Scenarios

 

Do You Have a Pandemic Scenario?



A recent white paper I wrote discussed the benefits of scenario analysis. The purpose of scenario analysis is to see how economic, environmental, political, and technological change can impact a company’s business. The recent outbreak of COVID-19 (“Coronavirus”) is a perfect example of how an environmental event can have an impact on the local and, as we are finding out, global economy.

As the world watches this virus spread, I suspect there are some companies who are thankful they created a pandemic scenario. Right now, they are probably preparing to take steps to enact the procedures they created after running the scenario. I also suspect there are other companies who might be in a bit of panic as they wonder how much this will impact them. To those companies I suggest they start considering the impacts now. While we hope this will not reach full pandemic level, the future is unknown.

 

Jonathan Leonardelli, FRM, Director of Business Analytics for FRG, leads the group responsible for model development, data science, documentation, testing, and training. He has over 15 years’ experience in the area of financial risk.

CECL Preparation: How Embracing SR 11-7 Guidelines Can Support the CECL Process

The Board of Governors of the Federal Reserve System’s SR 11-7 supervisory guidance (2011) provides an effective model risk management framework for financial institutions (FI’s). SR 11-7 covers everything from the definition of a model to the robust policies/procedures that should exist within a model risk management framework. To reduce model risk, any FI should consider following the guidance throughout internal and regulatory processes as its guidelines are comprehensive and reflect a banking industry standard.

The following items and quotations represent an overview of the SR 11-7 guidelines (Board of Governors of the Federal Reserve System, 2011):

  1. The definition of a model – “the term model refers to a quantitative method, system, or approach that applies statistical, economic, financial, or mathematical theories, techniques, and assumptions to process input data into quantitative estimates.”
  2. A focus on the purpose/use of a model – “even a fundamentally sound model producing accurate outputs consistent with the design objective of the model may exhibit high model risk if it is misapplied or misused.”
  3. The three elements of model risk management:
    • Robust model development, implementation, and use – “the design, theory, logic underlying the model should be well documented and generally supported by published research and sound industry practice.”
    • Sound model validation process – “an effective validation framework should include three core elements: evaluation of conceptual soundness, …, ongoing monitoring, …, and benchmarking, outcomes analysis, …”
    • Governance – “a strong governance framework provides explicit support and structure to risk management functions through policies defining relevant risk management activities, procedures that implement those policies, allocation of resources, and mechanisms for evaluating whether policies and procedures are being carried out as specified.”

The majority of what the SR 11-7 guidelines discuss applies to some of the new aspects from the accounting standard CECL (FASB, 2016). Any FI under CECL regulation must provide explanations, justifications, and rationales for the entirety of the CECL process including (but not limited to) model development, validation, and governance. The SR 11-7 guidelines will help FI’s develop effective CECL processes in order to limit model risk.

Some considerations from the SR 11-7 guidelines in regards to the components of CECL include (but are not limited to):

  • Determining appropriateness of data and models for CECL purposes. Existing processes may need to be modified due to some differing CECL requirements (e.g., life of loan loss estimation).
  • Completing comprehensive documentation and testing of model development processes. Existing documentation may need to be updated to comply with CECL (e.g., new models or implementation processes).
  • Accounting for model uncertainty and inaccuracy through the understanding of potential limitations/assumptions. Existing model documentation may need to be re-evaluated to determine if new limitations/assumptions exist under CECL.
  • Ensuring validation independence from model development. Existing validation groups may need to be further separated from model development (e.g., external validators).
  • Developing a strong governance framework specifically for CECL purposes. Existing policies/procedures may need to be modified to ensure CECL processes are being covered.

The SR 11-7 guidelines can provide FI’s with the information they need to start their CECL process. Although not mandated, following these guidelines overall is important in reducing model risk and in establishing standards that all teams within and across FI’s can follow and can regard as a true industry standard.

Resources:

  1. Board of Governors of the Federal Reserve System. “SR 11-7 Guidance on Model Risk Management”. April 4, 2011.
  2. Daniel Brown and Dr. Craig Peters. “New Impairment Model: Governance Considerations”. Moody’s Analytics Risk Perspectives. The Convergence of Risk, Finance, and Accounting: CECL. Volume VIII. November 2016.
  3. Financial Accounting Standards Board (FASB). Financial Instruments – Credit Losses (Topic 326). No. 2016-13. June 2016.

Samantha Zerger, business analytics consultant with FRG, is skilled in technical writing. Since graduating from the North Carolina State University’s Financial Mathematics Master’s program in 2017 and joining FRG, she has taken on leadership roles in developing project documentation as well as improving internal documentation processes.

 

CECL Preparation: The Power of Vintage Analysis

I would argue that a critical step in getting ready for CECL is to review the vintage curves of the segments that have been identified. Not only do the resulting graphs provide useful information but the process itself also requires thought on how to prepare the data.

Consider the following graph of auto loan losses for different vintages of Not-A-Real-Bank bank[1]:

 

While this is a highly-stylized depiction of vintage curves, its intent is to illustrate what information can be gleaned from such a graph. Consider the following:

  1. A clear end to the seasoning period can be determined (period 8)
  2. Outlier vintages can be identified (2015Q4)
  3. Visual confirmation that segmentation captures risk profiles (there aren’t a substantial number of vintages acting odd)

But that’s not all! To get to this graph, some important questions need to be asked about the data. For example:

  1. Should prepayment behavior be captured when deriving the loss rates? If so, what’s the definition of prepayment?
  2. At what time period should the accumulation of losses be stopped (e.g., contractual term)?
  3. Is there enough loss[2] behavior to model on the loan level?
  4. How should accounts that renew be treated (e.g., put in new vintage)?

In conclusion, performing vintage analysis is more than just creating a picture with many different colors. It provides insight into the segments, makes one consider the data, and, if the data is appropriately constructed, positions one for subsequent analysis and/or modeling.

Jonathan Leonardelli, FRM, Director of Business Analytics for the Financial Risk Group, leads the group responsible for model development, data science, documentation, testing, and training. He has over 15 years’ experience in the area of financial risk.

 

[1] Originally I called this bank ACME Bank but when I searched to see if one existed I got this, this, and this…so I changed the name. I then did a search of the new name and promptly fell into a search engine rabbit hole that, after a while, I climbed out with the realization that for any 1 or 2 word combination I come up with, someone else has already done the same and then added bank to the end.

[2] You can also build vintage curves on defaults or prepayment.

 

RELATED:

CECL—Questions to Consider When Selecting Loss Methodologies

CECL—The Caterpillar to Butterfly Evolution of Data for Model Development

CECLData (As Usual) Drives Everything

CECL Preparation: Questions to Consider When Selecting Loss Methodologies

Paragraph 326-20-30-3 of the Financial Accounting Standards Board (FASB) standards update[1] states: “The allowance for credit losses may be determined using various methods”. I’m not sure if any statement, other than “We need to talk”, can be as fear inducing. Why is it scary? Because in the world of details and accuracy, this statement is remarkably vague and not prescriptive.

Below are some questions to consider when determining the appropriate loss methodology approaches for a given segment.

How much history do you have?

If a financial institution (FI) has limited history[2] then the options available to them are, well, limited. To build a model one needs sufficient data to capture the behavior (e.g., performance or payment) of accounts. Without enough data the probability of successfully building a model is low. Worse yet, even if one builds a model, the likelihood of it being useful and robust is minimal. As a result, loss methodology approaches that do not need a lot of data should be considered (e.g., discount cashflow or a qualitative factor approach based on industry information).

Have relevant business definitions been created?

The loss component approach (decomposing loss into PD, LGD, and EAD) is considered a leading practice at banks[3]. However, in order to use this approach definitions of default and, arguably, paid-in-full, need to be created for each segment being modeled. (Note: these definitions can be the same or different across segments.) Without these definitions, one does not know when an account has defaulted or paid-off.

Is there a sufficient number of losses or defaults in the data?

Many of the loss methodologies available for consideration (e.g., loss component or vintage loss rates) require enough losses to discern a pattern. As a result, banks that are blessed with infrequent losses can feel cursed when they try to implement one of those approaches. While low losses do not necessarily rule out these approaches, it does make for a more challenging process.

Are loan level attributes available, accurate, and updated appropriately?

This question tackles the granularity of an approach instead of an approach itself. As mentioned in the post CECL – Data (As Usual) Drives Everything, there are three different data granularity levels a model can be built on. Typically, the decision is between loan-level versus segment level. Loan-level models are great for capturing sensitivities to loan characteristics and macroeconomic events provided the loan characteristics are accurate and updated (if needed) on a regular interval.

Jonathan Leonardelli, FRM, Director of Business Analytics for the Financial Risk Group, leads the group responsible for model development, data science, documentation, testing, and training. He has over 15 years’ experience in the area of financial risk.

 

[1]FASB accounting standards update can be found here

[2] There is no consistent rule, at least that I’m aware of, that defines “limited history”. That said, we typically look for clean data reaching back through an economic cycle.

[3] See: Capital Planning at Large Bank Holding Companies: Supervisory Expectations and Range of Current Practice August 2013

RELATED:

CECL—The Caterpillar to Butterfly Evolution of Data for Model Development

CECLData (As Usual) Drives Everything

CECL Preparation: The Caterpillar to Butterfly Evolution of Data for Model Development

I don’t know about you, but I find caterpillars to be a bit creepy[1]. On the other hand, I find butterflies to be beautiful[2]. Oddly enough, this aligns to my views on the different stages of data in relation to model development.

As a financial institution (FI) prepares for CECL, it is strongly suggested (by me at least) to know which stage the data falls into. Knowing its stage provides one with guidance on how to proceed.

The Ugly

At FRG we use the term dirty data to describe data that is ugly. Dirty data typically has these following characteristics (the list is not comprehensive):

  • Unexplainable missing values: The key word is unexplainable. Missing values can mean something (e.g., a value has not been captured yet) but often they indicate a problem. See this article for more information.
  • Inconsistent values: For example, a character variable that holds values for state might have Missouri, MO, or MO. as values. A numeric variable for interest rate might have a value as a percent (7.5) and a decimal (0.075)
  • Poor definitional consistency: This occurs when a rule that is used to classify some attribute of an account changes during history. For example, at one point in history a line of credit might be indicated by a nonzero original commitment amount, but at a different point it might be indicated by whether a revolving flag is non-missing.
The Transition

You should not model or perform analysis using dirty data. Therefore, the next step in the process is to transition dirty data into clean data.

Transitioning to clean data, as the name implies, requires scrubbing the information. The main purpose of this step is to address the issues identified in the dirty data. That is, one would want to fix missing values (e.g., imputation), standardized variable values (e.g., all states are identified by a two-character code), and correct inconsistent definitions (e.g., a line indicator is always based on nonzero original commitment amount).

The Beautiful

A final step must be taken before data can be used for modeling. This step takes clean data and converts it to model-ready data.

At FRG we use the term model-ready to describe clean data with the application of relevant business definitions. An example of a relevant business definition would be how an FI defines default[3]. Once the definition has been created the corresponding logic needs to be applied to the clean data in order to create, say, a default indicator variable.

Just like a caterpillar metamorphosing to a butterfly, dirty data needs to morph to model-ready for an FI to enjoy its true beauty. And, only then, can an FI move forward on model development.

 

Jonathan Leonardelli, FRM, Director of Business Analytics for the Financial Risk Group, leads the group responsible for model development, data science, documentation, testing, and training. He has over 15 years’ experience in the area of financial risk.

 

[1] Yikes!

[2] Pretty!

[3] E.g., is it 90+ days past due (DPD) or 90+ DPD or in bankruptcy or in non-accrual or …?

 

RELATED:

CECL—Questions to Consider When Selecting Loss Methodologies

CECLData (As Usual) Drives Everything

CECL Preparation: Data (As Usual) Drives Everything

To appropriately prepare for CECL a financial institution (FI) must have a hard heart-to-heart with itself about its data. Almost always, simply collecting data in a worksheet, reviewing it for gaps, and then giving it the thumbs up is insufficient.

Data drives all parts of the CECL process. The sections below, by no means exhaustive, provide key areas where your data, simply being by your data, constrains your options.

Segmentation

Paragraph 326-20-30-2 of the Financial Accounting Standards Board (FASB) standards update[1] states: “An entity shall measure expected credit losses of financial assets on a collective (pool) basis when similar risk characteristic(s) exist.” It then points to paragraph 326-20-55-5 which provides examples of risk characteristics, some of which are: risk rating, financial asset type, and geographical location.

Suggestion: prior to reviewing your data consider what risk profiles are in your portfolio. After that, review your data to see if it can adequately capture those risk profiles. As part of that process consider reviewing:

  • Frequency of missing values in important variables
  • Consistency in values of variables
  • Definitional consistency[2]
Methodology Selection

The FASB standard update does not provide guidance as to which methodologies to use[3]. That decision is entirely up to the FI[4]. However, the methodologies that are available to the FI are limited by the data it has. For example, if an FI has limited history then any of the methodologies that are rooted in historical behavior (e.g., vintage analysis or loss component) are likely out of the question.

Suggestion: review the historical data and ask yourself these questions: 1) do I have sufficient data to capture the behavior for a given risk profile?; 2) is my historical data of good quality?; 3) are there gaps in my history?

Granularity of Model

Expected credit loss can be determined on three different levels of granularity: loan, segment (i.e., risk profile), and portfolio. Each granularity level has a set of pros and cons but which level an FI can use depends on the data.

Suggestion: review variables that are account specific (e.g., loan-to-value, credit score, number of accounts with institution) and ask yourself: are the sources of these variables reliable? Do they get refreshed often enough to capture changes in customer or macroeconomic environment behavior?

Hopefully, this post has started you critically thinking about your data. While data review might seem daunting, I cannot stress enough—it’s needed, it’s critical, it’s worth the effort.

 

Jonathan Leonardelli, FRM, Director of Business Analytics for the Financial Risk Group, leads the group responsible for model development, data science, documentation, testing, and training. He has over 15 years’ experience in the area of financial risk.

 

[1] You can find the update here

[2] More on what these mean in a future blog post

[3] Paragraph 326-20-30-3

[4] A future blog post will cover some questions to ask to guide in this decision.

 

RELATED:

CECL—The Caterpillar to Butterfly Evolution of Data for Model Development

Avoiding Discrimination in Unstructured Data

An article published by the Wall Street Journal on Jan. 30, 2019  got me thinking about the challenges of using unstructured data in modeling. The article discusses how New York’s Department of Financial Services is allowing life insurers to use social media, as well as other nontraditional sources, to set premium rates. The crux: the data cannot unfairly discriminate.  

I finished the article with three questions on my mind. The first: How does a company convert unstructured data into something useful? The article mentions that insurers are leveraging public information – like motor vehicle records and bankruptcy documents – in addition to social media. Surely, though, this information is not in a structured format to facilitate querying and model builds.

Second: How does a company ensure the data is good quality? Quality here doesn’t only mean the data is clean and useful, it also means the data is complete and unbiased. A lot of effort will be required to take this information and make it model ready. Otherwise, the models will at best provide spurious output and at worst provide biased output.

The third: With all this data available what “new” modeling techniques can be leveraged? I suspect many people read that last sentence and thought AI. That is one option. However, the key is to make sure the model does not unfairly discriminate. Using a powerful machine learning algorithm right from the start might not be the best option. Just ask Amazon about its AI recruiting tool.[1]

The answers to these questions are not simple, and they do require a blend of technological aptitude and machine learning sophistication. Stay tuned for future blog posts as we provide answers to these questions.

 

[1] Amazon scraps secret AI recruiting tool that showed bias against women

 

Jonathan Leonardelli, FRM, Director of Business Analytics for the Financial Risk Group, leads the group responsible for model development, data science, documentation, testing, and training. He has over 15 years’ experience in the area of financial risk.

Does the Liquidity Risk Premium Still Exist in Private Equity?

FRG has recently been investigating the dynamics of the private capital markets.  Our work has led us to a ground-breaking product designed to help allocators evaluate potential cash flows, risks, and plan future commitments to private capital.  You can learn more here and read about our modeling efforts in our white paper, “Macroeconomic Effects On The Modeling of Private Capital Cash Flows.”

As mentioned in a previous post, we are investigating the effects of available liquidity in the private capital market.  This leads to an obvious question: Does the Liquidity Risk Premium Still Exist in Private Equity?

It is assumed by most in the space that the answer is “Yes.”  Excess returns provided by private funds are attributable to reduced liquidity.  Lock up periods of 10+ years allow managers to find investments that would not be possible otherwise.  This premium is HIGHLY attractive in a world of low rates and cyclically high public equity valuations.  Where else can a pension or endowment find the rates of return required?

If the answer is, “No,” then Houston, we have a problem.  Money continues to flow into PE at a high rate.  A recent article in the FT (quoting data from FRG partner Preqin) show there is nearly $1.5 trillion in dry powder.  Factoring in leverage, there could be, in excess of, $5 trillion in capital waiting to be deployed.  In the case of a “No” answer, return chasing could have gone too far, too fast.

As mentioned, leverage in private capital funds is large and maybe growing larger.  If the liquidity risk premium has been bid away, what investors are left with may very well be just leveraged market risk.  What is assumed to be high alpha/low beta, might, in fact, be low alpha/high beta.  This has massive implications for asset allocation.

We are attempting to get our heads around this problem in order to help our clients understand the risk associated with their portfolios.

 

Dominic Pazzula is a Director with the Financial Risk Group specializing in asset allocation and risk management.  He has more than 15 years of experience evaluating risk at a portfolio level and managing asset allocation funds.  He is responsible for product design of FRG’s asset allocation software offerings and consults with clients helping to apply the latest technologies to solve their risk, reporting, and allocation challenges.

 

 

 

 

 

Subscribe to our blog!