AI-Human Experiment: Building a Smarter Deposit Model for a Volatile World

Let's be honest: the 2022-2025 rate cycle broke a lot of old models. For those of us in banking, relying on static beta assumptions to predict how deposit rates would react to the Fed's aggressive tightening felt like navigating a hurricane with a paper map. It was clear we needed a fundamentally smarter approach.

This led me to an experiment: What if we paired expert human quants with advanced AI agents to build a next-generation deposit repricing model from the ground up? The goal wasn't just to improve statistical fit, but to build something robust, economically intuitive, and truly fit for purpose in modern asset-liability management (ALM).

The results were fascinating. The AI developed a volatility-adjusted dynamic beta model that not only achieved 97.5% explanatory power and cut forecast errors by 35% during the recent turbulent period but also gave us a powerful new playbook for human-AI collaboration in quantitative finance.

The AI-Powered Workflow: A New Kind of Quant Team

My experiment wasn't about letting an AI run wild. It was a structured collaboration.

1. Stage 1: Human-led Hypothesis. I started with a core economic thesis: a logistic dynamic beta framework was likely the right functional form, but it was missing something during periods when policy rates were pivoting. I gave the AI this initial hypothesis and a curated universe of potential driver variables: Fed funds rates, liquidity premiums, and various term spreads.
2. Stage 2: AI-driven Challenger Development. With the initial parameters set, I tasked independent AI agents to act as a challenger team. Their mandate was to explore the variable space and build competing models. They rapidly iterated through dozens of specifications, including enhanced logistic models with different macro drivers and quadratic formulations designed to capture convexity.
3. Stage 3: Independent AI Validation. A third AI agent, acting as an internal validation function, stress-tested all candidate models. It ran a full suite of statistical diagnostics, out-of-sample performance tests, and parameter stability analyses—automating what would typically take a human team weeks to complete.

The Breakthrough Model: A Look Under the Hood for Practitioners

For banking professionals looking to leverage this work, here is the final functional form of our recommended model. It combines a long-run equilibrium relationship with a short-term error correction mechanism, making it robust for both forecasting and risk management.

1. The Long-Run Cointegrating Relationship: This defines the target equilibrium deposit rate based on market conditions.

ILMDHYLD_t = α + β_t × FEDL01_t + γ₁ × FHLK3MSPRD_t + γ₂ × 1Y_3M_SPRD_t + ε_t        

2. The Dynamic Beta Function (with Volatility Adjustment): This is the core innovation, defining how the sensitivity (beta) of the deposit rate changes.

# First, calculate the beta based on the rate level β_t_level = β_min + (β_max - β_min) / (1 + exp(-k × (FEDL01_t - m))) 

# Then, adjust it for market volatility β_t = β_t_level × (1 - λ × (σ_t / σ*))        

Key Variables:

• σ_t: 24-month rolling volatility of the monthly change in the Fed Funds rate.
• σ*: The long-run average volatility over the entire sample period.
• λ: The volatility dampening parameter, which our model calibrated to ≈0.34.
• β_min / β_max: The floor and ceiling for the beta, which we bounded between 35.2% and 57.8%.
• m: The inflection point where competition rapidly accelerates, calibrated at a 2.73% Fed Funds rate.

3. The Short-Run Error Correction Component: This governs how quickly the deposit rate adjusts back to its long-run equilibrium after a shock.

ΔILMDHYLD_t = φ₀ + φ₁ × ΔFEDL01_t + φ₂ × ε̂_{t-1} + ν_t        

Key Variables:

• Δ: Represents the one-month change in the variable.
• ε̂_{t-1}: The error term from the long-run equation in the previous month. The coefficient φ₂ dictates the speed of adjustment back to equilibrium.

This complete structure ensures that the model is not only statistically robust but also grounded in economic theory, capturing both long-run relationships and short-term dynamics.

Where Human Expertise Was Irreplaceable

The AI agents were phenomenal at computational tasks, but the experiment's success hinged on my ability to guide and interpret their work.

• Economic Sense-Checking: Early on, an AI agent proposed a model with a negative coefficient on CPI inflation. Statistically, it fit well. Economically, it seemed backward. The insight was recognizing this wasn't a flaw but was capturing the Fed's aggressive anti-inflationary policy response, a nuanced, second-order effect an AI couldn't intuit on its own.
• Defining "Best": The AI optimized for pure statistical fit (R²). I had to balance that against other critical factors: regulatory acceptance (SR11-7 and Basel III), implementation complexity, the plausibility of the model under extreme stress, and the ability to explain it to a less technical audience.
• Assessing Real-World Impact: An AI can't tell you how a 10-basis-point change in predicted deposit duration will alter the bank's hedging strategy or its reported Economic Value of Equity (EVE). That requires a human risk manager who understands the intricate connections between the model and the bank's balance sheet.

The Surprise Finding: Why Volatility Can Dampen Competition

One of the most counterintuitive yet powerful findings was the role of the volatility dampening parameter (λ). I believe, as many would, that volatility should increase repricing betas as depositors shop for yield and liquidity becomes scarce.

However, the model showed that high volatility in the path of interest rates, distinct from high levels, temporarily slows repricing. Why?

1. Strategic Hesitation: When the rate path is uncertain, banks delay aggressive pricing moves to see what competitors do first. This information lag slows the market’s collective response.
2. Risk Management Overrides: During volatile periods, internal risk committees often impose conservative pricing caps to protect margins, overriding purely competitive instincts.
3. Depositor Psychology: While some depositors chase yield, many prioritize stability during uncertain times, creating a "flight to safety" that reduces their rate sensitivity.

The 24-month rolling volatility window captured this perfectly. It created an asymmetric effect: a lag in repricing on the way up during volatile periods and a slightly faster adjustment on the way down as volatility subsided. It’s a sophisticated dynamic that a simpler model would miss entirely.

AI's Next Job: Revolutionizing Model Governance & Maintenance

Developing a great model is only half the battle. The real work, and often the most tedious for quant teams, is the lifecycle management. This is where AI agents are poised to deliver their next big productivity win.

• Automated Performance Monitoring: Imagine an AI agent that, every month, automatically back-tests the model against actuals, recalculates performance metrics, and flags any degradation. It could generate a complete model performance dashboard for the risk committee before a human even logs in.
• Intelligent Recalibration & Refresh: Instead of a rigid annual review, an AI can continuously assess whether a model refresh (re-estimating coefficients) or a full recalibration (re-evaluating the functional form) is needed.
• "Living" Documentation: Model documentation is often a static, painful process. AI can create a "living document" that updates in real-time as data sources change, parameters are refreshed, or performance metrics evolve. It can generate pristine, regulator-ready reports on demand, freeing up quants to focus on what they do best: developing better models.

The New Playbook for Quantitative Modeling

My takeaway from this experiment is that the future isn't about replacing human experts with AI; it's about creating a powerful collaborative framework where each plays to its strengths.

Thus, a playbook for moving forward is to build on this synergy:

1. Human-led Strategy: We start with an economic hypothesis grounded in our deep domain expertise.
2. AI-driven Exploration: We use AI agents for the heavy lifting of rapid iteration, challenger model creation, and comprehensive statistical testing.
3. Human-centric Validation: The final model selection is a human judgment call, balancing statistical power with economic intuition, risk implications, and business strategy.
4. AI-powered Governance: We deploy AI agents to handle the routine but critical tasks of model monitoring, maintenance, and documentation, ensuring our models remain robust and compliant throughout their lifecycle.

The institutions that master this collaborative approach will not only build better models but will also build them faster and maintain them more effectively. They will gain a sustainable advantage in risk management, regulatory relations, and strategic agility. The future of quantitative finance belongs to teams that perfect this powerful partnership.

What's your take on using AI for model development and governance?