Agent-Based Modeling (ABM) is a powerful simulation capability for analyzing the behavior of individual agents, their interactions within a system, and the performance impacts upon the system. Unlike traditional models that rely on predicting static behavior within a system—assuming the behavior of one agent does not influence others—ABM operates at the micro-level, treating each agent—whether a person, a business, or the economic environment—as unique, with distinct goals and behaviors. This interactive approach captures the complexity and emergent system dynamics often missed in static behavioral models by regularly updating as agents influence one another and the world evolves. In this story, I share how my team at a regional bank used ABM to tackle a critical strategic challenge in auto lending. You will explore the nuances of auto lending, how ABM was applied to understand behavior and profitability, and the lessons we gained via agent interactions. Ultimately, our ABM work provided key insights that helped position the bank for success during the turbulent financial crisis of 2008 and beyond.
This auto loan example illustrates how ABM can serve as a transformative tool, not only in banking but also across related industries. Beyond its practical application, ABM fostered a cultural shift within my bank, establishing data and analysis as fundamental pillars for organizational learning and decision-making. This story showcases ABM’s versatility in navigating complex systems and guiding informed strategic decisions. Read on to discover how this innovative approach turned uncertainty into clarity.
About the author: Jeff Hulett leads Personal Finance Reimagined, a decision-making and financial education platform. He teaches personal finance at James Madison University and provides personal finance seminars. Check out his book -- Making Choices, Making Money: Your Guide to Making Confident Financial Decisions.
Jeff is a career banker, data scientist, behavioral economist, and choice architect. Jeff has held banking and consulting leadership roles at Wells Fargo, Citibank, KPMG, and IBM.
Table of Contents
Introduction: Why ABM Matters - Overview of ABM and its transformative potential.
The Auto Loan Growth Dilemma - Strategic challenges in direct and indirect auto lending.
Using Agent-Based Modeling (ABM) - Creating a virtual world of customer, dealer, and bank agents.
ABM Technical Foundations and Building the Customer Agent - Building the customer agent and modeling emergent behaviors.
Building the Dealer Agent - Understanding dealer dynamics, profitability, and competition.
Building the Bank Agent - Balancing short-term profitability with long-term credit risk.
Strategic agent interactions - Example from the three agents.
Dual Simulation Objectives - Insights from point-of-sale and life-of-loan simulations.
The Economic Environment and Capital Impact - Navigating capital needs and regulatory challenges.
Key Findings from the ABM Exercise - Credit capital, pricing strategies, dealer profitability, and risk insights.
Optimizing Value in Indirect Auto Finance with ABM - Practical applications for profitability, volume, and risk management.
Conclusion: Turning Crisis into Advantage - Strategic outcomes and lessons from the financial crisis.
Appendix: Critique of Traditional Economics – ABM’s Advantage - How ABM addresses gaps in Neoclassical Economics.
Resources for the Curious - Further readings and references.
The Auto Loan Growth Dilemma
Let us begin with some background. Auto lending is a key portfolio product for banks, typically funded by customer deposits or other bank funding sources. It provides an important avenue for generating interest income while supporting consumer purchases. Banks generally have two main approaches to consumer auto lending: Direct Lending and Indirect Lending.
Direct Lending: This involves the bank providing loans directly to borrowers, who secure pre-approval before purchasing a vehicle. It empowers borrowers to car shop confidently by disaggregating the financing from the car purchase. This could lead to lower total consumer costs - of both the car and the interest costs. However, borrowers require more effort to handle the loan approval process independently. For banks, direct lending fosters strong customer relationships but often incurs higher marketing and acquisition costs.
Indirect Lending: Here, banks partner with dealerships to offer financing at the point of sale. Borrowers benefit from convenience, as financing is integrated into the purchase process. However, total consumer costs, including the car and the interest costs, may be higher due to dealer markups or profit-sharing agreements between the dealer and the manufacturer. For banks, this approach provides access to more potential borrowers but may weaken customer relationships and reduce profit margins.
Transaction Costs in Auto Lending: Transaction costs—such as non-payment costs like consumer time, effort, and risk perception—strongly influence borrower preferences. Direct lending historically increases these perceived costs by requiring borrowers to manage financing independently from the car purchase. In contrast, indirect lending integrates financing into the purchase process, reducing transaction costs and enhancing convenience, despite potentially higher overall costs. Today, advances in digital tools plus the increase in car prices relative to those transaction costs mitigate direct lending transaction costs, making it a more viable alternative. However, indirect lending remains a popular consumer auto lending alternative.
As a former Chief Consumer Credit Officer at a regional bank, I encountered a significant challenge with our indirect auto loan portfolio, which had grown rapidly. Our direct auto business, in contrast, grew at a more muted pace. Although indirect auto lending was narrowly profitable and contributed positively to the bank's overall performance, declining margins indicated that recent growth might not be sustainable. Additionally, we were concerned about insufficient credit reserves to weather a potential downturn. Despite meeting regulatory capital requirements, we suspected more reserves were needed, especially for higher-risk loans. The pressing question became: “What should our indirect auto strategy be? Grow, maintain, or contract?”
Using Agent-Based Modeling (ABM)
To tackle this question, we employed Agent-Based Modeling. The essential part of ABM in a data-rich environment like banking is that we could create a virtual world where every customer, dealer, and bank stakeholder is represented as an independent and unique agent. While the simple motivation governance rules for each agent type were the same, each agent could express their diverse behavior in unique ways. These synthetic agents -- like digital bots -- adhered to basic motivation rules and programmed behaviors, but their outcomes were emergent. The interactions among synthetic agents over time generated patterns and dynamics that were not explicitly programmed but arose naturally from the system, offering insights into strategic guidance and predictions of essential business outcomes.
At the point of sale - when the loan is made - we defined three types of synthetic agents: customers, auto dealers, and the bank. Each agent type operated under simple motivation rules:
Customers sought vehicles, affordable loan payments, and the best available rates amid competitive banking options.
Dealers prioritized available financing options that facilitated car sales while increasing revenue.
The bank focused on generating loan revenue while managing credit risk effectively.
After the loan is made, a fourth synthetic agent is brought into the mix, representing the economy and regulatory environment. We introduce the fourth agent later in "The Economic Environment and Capital Impact" section.
Once the model was developed and operationalized, updating the data, analyzing agent interactions, and making business recommendations became relatively straightforward. Our agents moved through time in monthly increments, as most of our data was structured on a monthly basis. Each agent began the month with a set of behavioral expectations and known performance metrics, acted within the system during the month, and produced measurable outputs at the end of the month. This approach allowed for dynamic interactions between agents, as well as the incorporation of random environmental shocks, simulating the evolving and interconnected nature of real-world systems. Over time, these time-bound iterations revealed patterns critical for informing strategic decisions and understanding the compound effects of agent behaviors.
Also, this made it easier to understand agent groups tending to have more homogenous behavior. Over time, it enabled us to reduce the necessary computing resources by aggregating homogenous agents.
 | ABM Technical Foundations and Building the Customer Agent |
|---|
The technical component of the agent-based modeling (ABM) exercise required two important resources - data and expertise.
On the data side, part of my job was to work with the bank's technology organization to create analytically focused data warehouses. Similar to other banks of our size during that period, our data was centered on production systems. This meant the data was siloed within product systems for originating or servicing loan products. There was much data, but it was not organized for analysis, and the all-important metadata was often retained only in the minds of long-term employees. This foundational effort required collaboration and support from multiple organizations, including the CFO, COO, CTO, business units, and myself representing risk management. Together, we worked to liberate the data for learning and analytical application.
Once the data was ready, the next substantial effort combined expertise in coding, modeling, mathematics, finance, accounting, bank operations, and behavioral economics. At the time—in the mid-2000s—we relied on tools like Excel, Visual Basic (VB), and a few specialized business intelligence and modeling platforms. VB was the primary tool for simulating agent interactions, allowing iterative updates as behaviors evolved. Excel housed fixed inputs to the simulation models, which were updated over simulated timeframes. Today, integrated modeling and data solutions such as Python and R would significantly streamline the development process, while modern machine learning techniques, including neural networks and decision tree predictors like Random Forest, could enhance predictive accuracy and analytical depth.
The ABM simulations heavily relied on single-behavior predictive model inputs. These models helped us define each agent’s “bounded rationality,” or a unique set of behaviors that varied across our customer base. Bounded rationality means 1) each customer has a limited set of knowledge about their situation and car options and 2) that information will be utilized by the customer agent to follow their simple motivation as defined earlier. Each single-behavior model represented a specific outcome tied to a customer’s motivation related to the car or car loan.
For example, some customers prioritized securing the lowest rate, while others focused on manageable payment terms. Each customer agent acted independently of others, resulting in unique loan terms and behaviors. Over time, some agents would default, allowing us to model loss rates and mitigation strategies. Which agent would default was randomly chosen by their weighted behavioral profile. As an input to the ABM, customer agent behavior was derived from actual behavior observed in the recent past. Importantly, we let the data inform our synthetic agents as to how their motivation was expressed.
Bounded rationality and ABM represent a transformative evolution in economic theory. In the appendix, I explore how ABM addresses the shortcomings of traditional Neoclassical Economics that Behavioral Economics and Complexity Economics seek to reconcile.
By simulating external shocks, such as a recession, the customer ABM allowed us to observe how default rates spiked relative to steady-state scenarios. We added randomness and volatility to the simulation to represent a phase state-like transition—from a calmer, steadily growing economy to a more volatile, recessionary economy. In physics, this is similar to the phase state transition of water, where the application of heat causes it to change from liquid to steam. For instance, if interest rates rose sharply and housing starts declined, our synthetic customer agents with income tied to residential housing would reflect how job losses affected auto loan defaults. The simulation also allowed us to test different collections efforts and strategies, optimizing resources and policies to reduce losses while better assisting impacted customers. This insight helped the bank prepare staffing plans and refine collections policies to address economic shifts proactively.
We also modeled internally generated impact, which arises from individual agent behavior rather than external factors. An intriguing outcome of the ABM was the higher prevalence of internal factors compared to external ones. For instance, the endowment effect, as introduced by behavioral economists like Richard Thaler and Daniel Kahneman, was strongly supported by our data and emerged as a major behavioral driver. Customers were more likely to walk away from loans when their car’s value fell below a portion of the loan balance, triggering sharp, clustered volatility in our simulations. While the exact tipping point varied by the customer, the effect underscored how willingness to repay, shaped by perceived "skin in the game," could override capacity to repay—such as income or savings. This behavior emphasized the importance of collateral management in addressing the emotional and practical factors influencing loan performance.
Our ABM-modeled behavioral inputs included:
Credit behavior and frequency risk: Logistic regression models using internal bank customer data.
Endowment effect: A lagged estimate of car collateral loan-to-value served as a proxy for the customer’s willingness to pay, adjusting as their car's “skin in the game” perception evolved.
Collections behavior: Bayesian/Markov Chain predictors simulated loss timing based on actual collections data.
Loan attrition and severity risk: Logistic models paired with expected value (EV) loan balance calculations.
Production costs: Data sourced from our finance group to evaluate loan origination and servicing expenses. Origination costs are an upfront investment in creating the loan. This investment is gradually recovered over time as the customer makes loan interest payments, as shown in the 'Value Path' graphic below.
Loss timing: Vintage analysis, also known as static pool analysis, to estimate time-related impacts on agent behaviors. Loss timing was essential for adjusting cash flow discounting and predicting when and what kind of servicing effort would need to occur.
Accounting rules: Certain accounting expense treatments, like marketing expenses, enabled the bank to capitalize some expenses and not others. These treatments impacted short-term profitability and associated bank behavior. (see ASC 310–20)
For competitive insights, traditional FICO scores predicted loan sale behavior at the point of sale. Decision trees with chi-squared predictor functions added depth to our analysis. By combining these models, we developed a detailed understanding of customer behavior over the life of a loan.
This granular, “atomic-level” customer agent analysis was instrumental in capturing emergent dynamics and interactions. It laid the groundwork for adapting the ABM framework to dealer-agent behavior, enabling us to examine the profitability and operational dynamics of our dealer relationships.
 | Building the Dealer Agent |
|---|
Our next step was to apply a variant of the customer ABM to the dealers. The synthetic dealer agent model helped us understand the cost, profitability, and sales dynamics of each dealer. Dealers generally view their lender-partners as enablers to sell cars. Each loan provides dealers with revenue through mechanisms like dealer reserves. Plus, dealers see loans as tools to "move the metal" and boost car sale revenue. Essential dealer rules included the competitiveness of the customer's interest rates to facilitate sales and the balance of the competitive relationships with multiple lender options. Dealer interaction behavior was also influenced by the depth and length of the bank's relationship and whether the dealer primarily sold new or used cars.
We had less visibility into dealers' internal operations and instead relied on our relationship managers, our dealer reserve agreements, and data from the loans generated by each dealer. Dealer agents were modeled with sensitivity to the payments received from the bank for the loan and the interest rate charged to each customer. The rate was a major lever for increasing or decreasing volume, but the demand elasticity and loan mix varied depending on factors such as the dealer's car age focus (new vs. used cars) and the strength of their relationship with the bank. For new dealers or those with shallow relationships, sensitivity to rate changes was often higher. This rate sensitivity reflected the reality that dealers are customer "loan gatekeepers." Thus, dealers have a choice about how eligible loan offers will be shown to the customer at the point of sale. In recent times - 2024 - the FTC implemented the "CARS Rule" intended to encourage full disclosure of loan eligibility. It is unclear whether this new rule will significantly change gatekeeping behavior. The CARS Rule may tilt the competitive environment to reduce consumer transaction costs via better disclosure. This would make bank direct lending more viable. The good news is, whatever the new rule impact is, the synthetic agent is designed to help predict the outcomes and reflect the behavior as it emerges.
Dealers also frequently employ aspirational-level adaptation strategies to boost end-of-period sales. Aspirational-level adaptation is closely tied to bounded rationality and how agents work with limited information to make decisions. For example, a dealer might target customers aspiring to own luxury SUVs by offering limited-time financing deals, emphasizing features like premium interiors or eco-friendly technology. This approach aligns with the dealer ABM, which modeled how adjustments in pricing and incentives could influence customer demand and dealer profitability. By simulating these scenarios, the ABM revealed which dealers were effectively leveraging aspirational strategies and which required tailored support to optimize performance.
Dealer finance is highly competitive. Our bank’s loans competed with other banks’ loans and dealer reserve policies. Loans could not be sold unless the “buy rate” (the customer’s interest rate) was low enough and the dealer reserve was sufficiently high. This dealer ABM enabled us to evaluate the profitability of dealer relationships based on factors like loan volume, dealer reserve payments, spreads, and loan performance.
 | Building the Bank Agent |
|---|
Finally, the synthetic bank agent incorporated two main stakeholders—generally aligned by first-line and second-line bank governance:
The auto lending business unit, represented by the executive overseeing consumer lending, aimed to drive loan volume, ensuring happy customers and dealers.
The risk management function, represented by me as the Chief Consumer Credit Risk Officer, focused on long-term performance, credit capital, and strategy.
There were certainly other bank stakeholders with different incentives, but we kept the bank agents relatively simple and tied to essential profitability drivers like revenue and credit loss. The interaction between these two primary stakeholders often involved a “push and pull” balancing dynamic. Pricing decisions and strategic priorities served as levers where competing objectives had to align, balancing short-term volume growth with long-term risk management. Circling back to the Accounting Rules, this is a high-impact example of behavioral impact. The auto lending business unit is more likely to place a heavier weight on how indirect auto dealer reserve costs can be capitalized, increasing short-term profitability. Whereas the risk management function will be agnostic since their view is longer-term --where cashflow timing is emphasized over accounting treatments governing exchanges between the income statement and balance sheet.
A note about regulatory compliance factors. We considered adding a third regulatory stakeholder to the bank agent. Regulations like Fair Lending, the Fair Credit Reporting Act, and others impacted the bank and how we interact with the customer and the dealer. The regulatory impact has only increased since the financial crisis. To keep it simple, we resolved to maintain regulatory impact within the two stakeholders. Since complying with laws was the responsibility of the 2 main stakeholders, creating a third stakeholder would be like double counting.
Strategic agent interactions
The bank agent's interactions with the customer and dealer agents revealed significant insights. Over time and in a stable economic environment, these interactions tended to settle into predictable patterns as the customer and dealer agents adjusted to the bank’s pricing and policy strategies. However, small changes in bank agent decisions—particularly around pricing—had outsized effects on customer and dealer agent performance over time.
For example, if the bank increased interest rates for a lower-performing dealer, that dealer would sell fewer of the bank’s loans, weakening the bank-dealer relationship. This increased demand elasticity among the dealer's customers for our bank's loan products, leading to fewer competitive opportunities and attracting more price-insensitive, higher-risk borrowers. Over time, the lack of competitiveness created a downward spiral of declining loan volume and worsening credit quality, suggesting the bank might benefit from exiting relationships with some underperforming dealers. Major strategic shifts, such as tightening credit or changing dealer incentives, demonstrated clear long-term impacts, especially in volatile economic environments like recessions. These dynamics underscored the need for careful calibration of bank policies, a topic explored further in the next section.
Dual Simulation objectives
The agent's interactions were simulated at both the point of sale and over the life of the loan.
Point of Sale (POS): Our objective was to observe the interactions between the 3 main virtual agent types - customers, dealers, and the bank - to predict loan production, loan profitability, credit risk, and capital impacts. The POS interaction between the 3 agent types was essential for building the "in-the-door" population of borrowers. The interactions helped us understand sensitivities - known as demand elasticities - impacting that population. The current and anticipated future economic environment is an essential consideration at POS. The Bank agent represents the environment via their credit policies and underwriting procedures.
The Big Transition: The transition from the Point of Sale to the bank's portfolio (LoL) marks a critical shift in responsibility and control. At POS, the bank evaluates and decides whether to originate the loan, retaining full control over the loan capital. Once the loan is made, however, the responsibility shifts to the borrower, and the bank must rely on the borrower’s behavior and the influence of external economic conditions to determine the loan's performance over its life.
Life of Loan (LoL): Our objective was to observe the interactions between 3 virtual agent types - customers, the bank, and the economic environment - with various customer credit risks, endowments, attrition, and various bank collections policies and costs, specifically for customers whose loans had already been originated. We did this in order to predict the loan profitability, credit risk, and funding impacts. Notice in this objective, the dealer agent type has been removed and replaced by an environmental agent. Once a loan is booked, the uncertain future economic environment is an essential determinant of loan performance.
POS and LoL interaction: Please note that there is an interaction loop between the Life of Loan and the Point of Sale simulations, especially within the customer ABM. Ultimately, the Life of Loan simulation outcome informs policy for the Point of Sale simulation, whereas the Point of Sale sim creates the input data for the Life of Loan sim.
Value Path
Agent Interaction Outcomes: Likely economic environment scenario
Graphic note: This graphical example represents the "Likely economic environment scenario" future world. Every customer is represented by a line. The line aggregates their interactions with the other agents and their internally generated 'boundedly rational' behavioral inputs mentioned earlier. When modeling necessary UL, the ABM scenario would be stressed as a recessionary scenario. In a stressed scenario, there would be more red, falling value default customer agents; and fewer increasing value green customer agents. Also, homogenous agents, like those found in the thickest lines, would be periodically validated. As behavioral economics teaches us, diverse rationality is in constant flux. Which means homogeneity assumptions are as well.
 | The Economic Environment and Capital Impact |
|---|
In aggregate, banks exert a significant impact on the economy. Individually, however, banks behave as if they have no control over the economic environment but are very sensitive to its effects. Changes in the economic environment are why banks' capital needs change. Banks need more capital in recessionary times when defaulted loans and related stresses increase. Banks require less capital in expansionary times when loan repayment is more likely. Holding more capital is the same as reducing bank leverage. Excess leverage is how banks get into trouble. Excess leverage in a volatile, recessionary environment causes banks to struggle and potentially fail. When evaluating auto loan performance, capital is a cost. If more capital is required, higher loan interest rates are needed to produce that capital. Yet less capital-sensitive competition may siphon away those higher interest-rate car loans that would produce that capital. A natural tension exists between competition, customer motivations, and the bank's desire to stay solvent. It is the expected economic environment at the crux of resolving that tension.
Another important environmental factor is the bank regulators. The bank regulators supervise the environmental rules for how much capital banks need to hold for their auto loans. The regulators could do great service to banks and our economy by enforcing counter-cyclical capital policies, which require banks to accumulate capital in the good times when the risks are manufactured and release that capital in the bad times when those risks are realized. Under Basel III, the latest international banking accord to ensure banks hold ample capital, there is a countercyclical capital provision. Today we have a 2 tier financial services world. Banks are regulated by major bank regulators. The bank's competitive cousins are called nonbanks. Nonbanks provide similar bank loan products but are not regulated by the same regulators. This makes the practical application of capital rules inconsistent. This auto loan ABM enables the bank to implement prudent counter-cyclical capital policies. If financial services firms - both banks and nonbanks - played by the same capital rules, our economy would be subject to less severe boom-bust cycles. Economist John Kenneth Gailbraith developed a pithy saying for when banks manufacture undiscovered risks in the good times -- he called it "The Bezzle." Counter-cyclical capital policies would likely decrease the size and impact of the Bezzle in the bad times.
Key Findings
Our ABM exercise revealed critical insights:
Credit Capital: We were undercapitalized for higher-risk customers, necessitating loan loss allocation adjustments. Broadly, we determined that additional capital reserves were necessary for our auto loan product.
Loan Pricing: Pricing strategies needed revision to reflect customer behaviors and lifecycle dynamics. We were able to be more competitive with some customer agents and less with others. Changes in pricing impacted our dealer relationship. The ABM was like moving from a hacksaw to a scalpel.
Dealer Profitability: Consistent with the Pareto Principle -- also known as the 80/20 Rule -- about 80% of profitability stemmed from about 20% of the dealers. Many new or expanded relationships were unprofitable. This insight led to a strategic game plan for improving dealer profitability. The game plan included strengthening, winding down, and exiting dealer relationships.
Competitive Landscape: Competitors lacked similar sophistication in pricing adjustments, creating a competitive arbitrage opportunity. We could "sell" unprofitable loans to our competition by not making them, then "buy" loans by targeting more competitive pricing to the loans we want. The arbitrage trade was enabled, in part, because the auto dealer market had a simple "price sheet" approach to communicating loan rates and requirements.
Fixed Cost Challenges: Reduced loan volume risks triggering the “fixed cost spiral of death,” with fixed overhead spread across fewer loans. The spiral occurs when unit loan volume declines, causing fixed costs to be increased across the remaining loans, which causes higher prices, which causes lower volume... and the spiral continues. It may be very difficult for banks to cut out fixed costs. For example, since the financial crisis, the costs of regulatory compliance have become a substantial and unmoveable fixed-cost burden. This insight helped the bank understand how close we were to the fixed cost fixed cost spiral cliff.
Risk Assumptions: We knew that our loan level E(L), also known as the “expected loss” calculation, was based on data from one of the calmest, lowest loss periods in recent history – from 2000 to 2005. We developed U(L) – or additional loan loss provisions to capture unexpected loss potential – based on the tails of the E(L) distributions. We worried that the distribution kurtosis could be higher in a turbulent credit market. In other words, when the business cycle flips from expansion to recession, the forward distribution tails may become thicker than the historical normal distribution suggests. The impact from the consumer credit stakeholder added more cost and potential for the “fixed cost spiral of death.”
After collaborating with the auto business unit, we aligned on a strategy. The new auto loan focus included expanding the top-performing dealers while addressing profitability issues with the other dealers. On balance, this is the "maintain" strategy but with a focus on changing our dealer and customer mix to drive better business performance. The ABM simulations were instrumental in preparing the bank for these decisions, as they illuminated key agent interactions and suggested actionable paths forward. However, it is important to appreciate that ABM simulations, while insightful, are not perfect representations of the future due to inherent uncertainties. A best practice in ABM is rapid updating as inputs and interactions adjust and adapt. Our ABM insights did predict that competitive pressures quickly shifted, as rivals capitalized on customers by providing lower-rate loans. What we did not know was whether our lending unit could adjust the dealer and customer mix as the ABM recommended.
The answer came about a year later…
By 2006, our inability to adjust our dealer and customer mix became evident. The unprofitable loans were being swapped out, but the better loans from the profitable dealers were not being swapped in as quickly. With net volume dropping, the fixed cost spiral created an unsustainable business environment. This led us to downsize our indirect auto lending business. Bank leadership considered maintaining investment in profitable dealers and absorbing the fixed costs as a strategic investment. In time, there was reason to believe we could be successful. However, there were other investment alternatives with clearer paths to profitability and more aligned with the bank’s commercial business strategy. In a different bank with a more consumer/auto loan-focused strategy, this indirect auto business could have been more successful.
Optimizing Value in Indirect Auto Finance with ABM
For banks with significant commitments to indirect auto finance, Agent-Based Modeling (ABM) offers a robust framework to enhance profitability, increase loan volumes, and improve credit risk management. Here is how:
Increasing Profitability ABM analyzes dealer-specific performance by simulating dynamic interactions between customers, dealers, and the bank. This granular view identifies top-performing dealers, enabling tailored incentives and pricing strategies to strengthen partnerships and drive higher returns. ABM also reveals underperforming dealer relationships, offering actionable insights to improve results or reallocate resources effectively.
Driving Loan Volume ABM models borrower preferences, including sensitivity to interest rates and loan terms, helping banks design competitive offerings that attract a broader customer base. By optimizing "buy rate" policies, banks can balance borrower appeal with dealer satisfaction, significantly increasing loan origination volumes.
Enhancing Credit Risk Management ABM generates detailed customer-level risk profiles by simulating default probabilities and loss patterns under various scenarios. These insights allow banks to refine pricing models, reserve allocations, and underwriting standards to achieve sustainable, risk-adjusted returns. For example, ABM can assess the impact of economic downturns on specific customer segments, enabling proactive adjustments to credit policies.
Optimizing Dealer Relationships ABM identifies key dynamics in dealer behavior, such as pricing expectations and customer demographics, enabling banks to create financing programs that align dealer incentives with profitability goals. By understanding competitive dealer reserve practices, ABM positions the bank as a preferred financing partner while maintaining profitability.
Adapting to Market Changes ABM supports scenario testing to evaluate the effects of economic fluctuations, regulatory changes, or competitor actions. This adaptability allows banks to respond swiftly to evolving market conditions, protecting profitability and securing their competitive position.
By incorporating ABM into strategic decision-making, banks can align growth, profitability, and risk management objectives, ensuring their indirect auto finance business operates efficiently, remains competitive, and achieves sustainable success.
Conclusion: Turning Crisis into Advantage
The decision to contract our auto business proved prescient. About a year and a half later, the financial crisis struck in full force. Our unease about our U(L) assumptions and tail thickness turned out to be well-founded. The low credit risk environment of the early 2000s quickly turned into one of the worst credit environments in history -- owing to the financial crisis. With reduced exposure to higher-risk loans and robust reserves, our bank weathered the storm with minimal disruption. Our strong balance sheet allowed us to capitalize on distressed bank acquisitions under FDIC assistance, tripling the bank’s size post-crisis. By implementing counter-cyclical capital policies before the crisis, we built a war chest to navigate the challenges of the downturn. Reflecting on this journey, I am deeply proud of our credit risk team and our innovative use of ABM to guide us through uncertainty.
Also, the ABM approach, and especially the culture of respecting data as a learning asset, blossomed at the bank. The data infrastructure became a key enabler for related ABM and other useful bank applications. While I introduced and oversaw implementation, it was the other bank leaders who appreciated its value and helped align our culture.
ABM is used broadly in the field of Complexity Economics. This is a fast-growing field, helping policymakers in government and business to make better forecasts by embracing the dynamic uncertainty of the world in their models. When I used ABM in the mid-2000s, to my knowledge, we were the only bank using ABM. ABM was relatively new and avant-garde. Kudos to my bank for walking with me on this journey. While ABM is not a household name in banking yet, more progressive banks are using various forms of ABM today, including the Fed and other central banks. Also, academic leadership from Doyne Farmer, Rob Axtell, and many others are helping to bridge the gap between theory and practice.
Appendix: Critique of Traditional Economics - ABM's Advantage
Both Jeff Hulett and Doyne Farmer critique traditional economics for relying heavily on overly simplistic assumptions that fail to account for the dynamic and complex nature of real-world systems. Traditional economic models often hinge on equilibrium assumptions, presuming markets clear at a specific price and quantity where supply meets demand. These models rely on "robo-rationality," assuming agents make perfectly rational decisions with complete information. However, this overlooks the diversity of motivations, the bounded rationality of individuals and organizations, and the emergent complexities of agent interactions.
Doyne Farmer, a leader in Complexity Economics, argues that traditional models are inadequate because they ignore the non-linear interactions and adaptability of agents within evolving systems. Similarly, Jeff Hulett critiques traditional economics for assuming static, homogenous behaviors, failing to reflect how agents act dynamically and independently based on their unique circumstances.
In the banking example, ABM was superior because it did not assume a market-clearing equilibrium. Instead, each agent—whether a customer, a dealer, or the bank—was free to act based on their distinct motivations and situations. As a result, price and quantity were observed as an outcome, not enforced as an input. Customers sought affordable loans, dealers prioritized financing options to sell cars, and the bank balanced profitability and credit risk. By enabling agents to pursue their own "diverse rationality," ABM replaced the oversimplified notion of a central "robo-rationality" with a framework that reflected real-world decision-making diversity.
Diverse rationality and other challenges to traditional economics are explored in: Becoming Behavioral Economics: The Social Science Revolutionizing Decision-Making
This dynamic modeling approach provided granular insights into profitability drivers, credit risk, and competitive dynamics, while allowing agents to adapt over time and interact in non-linear ways. By capturing this complexity, ABM proved to be a transformative tool, aligning with the critiques of traditional economics and enabling the bank to navigate the uncertainties of the 2008 financial crisis with resilience and clarity.
Resources for the Curious
Jeff Hulett, 2023, “Becoming Behavioral Economics: The Social Science Revolutionizing Decision-Making,” The Curiosity Vine.
Kirman, A., 2010, Complex Economics: Individual and Collective Rationality, Routledge.
Simon, H. A. (1955). "A Behavioral Model of Rational Choice." The Quarterly Journal of Economics, 69(1), 99–118. https://doi.org/10.2307/1884852
Klein, Benjamin, Crawford, Robert G., & Alchian, Armen A. (1978). "Vertical Integration, Appropriable Rents, and the Competitive Contracting Process." The Journal of Law and Economics, 21(2), 297-326.
Farmer, J. D., and Foley, D., 2009, "The Economy Needs Agent-Based Modeling," Nature, 460(7256): 685–686.
Tesfatsion, L., 2001, "Introduction to Agent-Based Computational Economics," ISU Economics Working Paper Series.
Epstein, J. M., 2007, Generative Social Science: Studies in Agent-Based Computational Modeling, Princeton University Press.
Axtell, R., 2000, "Why Agents? On the Varied Motivations for Agent Computing in the Social Sciences," Center on Social and Economic Dynamics Working Paper No. 17.
Cyert, R. M., & March, J. G. (1963). A Behavioral Theory of the Firm. Prentice-Hall.
Brock, W. A., and Durlauf, S. N., 2001, "Discrete Choice with Social Interactions," The Review of Economic Studies, 68(2): 235–260.
Hulett, Jeff. The Subtleties of Lending Discrimination. The Curiosity Vine, 2022.
North, M. J., and Macal, C. M., 2007, Managing Business Complexity: Discovering Strategic Solutions with Agent-Based Modeling and Simulation, Oxford University Press.
Jeff Hulett, 2023, "The Business and Science of Scale - How Data, Decision Insight, and Economics Drives Success," The Curiosity Vine.
Federal Deposit Insurance Corporation (FDIC). "Bank Failures in Brief – Summary." Accessed November 26, 2024. https://www.fdic.gov/bank/historical/bank/.
Federal Reserve Board. (2020). Federal Reserve Board announces that the countercyclical capital buffer (CCyB) remains at the current level of 0 percent. Accessed November 27, 2024, from https://www.federalreserve.gov/newsevents/pressreleases/bcreg20201218c.htm.
Galbraith, John Kenneth. The Great Crash 1929. Original edition published in 1954. Mariner Books, April 28, 1997. ISBN: 9780395859995.
Financial Accounting Standards Board (FASB). Accounting Standards Codification (ASC) 310–20: Receivables – Nonrefundable Fees and Other Costs. Norwalk, CT: FASB, accessed 11/26/2024.
Taleb, Nassim Nicholas. Skin in the Game: Hidden Asymmetries in Daily Life. Allen Lane, 2018.
Richard Thaler, "Toward a positive theory of consumer choice," Journal of Economic Behavior & Organization, Volume 1, Issue 1, 980, Pages 39-60
Kahneman, Daniel, and Amos Tversky. "Prospect Theory: An Analysis of Decision under Risk." Econometrica, vol. 47, no. 2, 1979, pp. 263–291.
Nice application of Agent-based Modeling