TL;DR

• What it is: A deposit flight risk model scores which large-balance or rate-sensitive deposit relationships are likely to leave before the balance moves.
• The core mechanic: The model reads balance concentration, rate exposure, account tenure and quarter-end seasonality, the signals a standard concentration report does not carry.
• Who performs it: Treasury and the CFO own the liquidity risk call. iTuring.ai’s platform generates the ranked list and the retention window early enough for a call or a repricing offer.
• One caveat: No named flight-risk case study backs this capability yet. The result in this piece comes from deposit growth. It is shown here as evidence that the same underlying propensity-and-outreach mechanism works, applied to flight risk.

A deposit concentration report tells you what left last quarter. It names the relationships that shrank, the share of total deposits they held, and the dollar amount that walked out the door. None of that helps the quarter you are currently in. A flight risk model is built to answer a different question: which relationships are about to move, and is there still time to call them. This piece lays out what that difference requires operationally, what the Interagency Policy Statement on Funding and Liquidity Risk Management already expects an institution to have in place, and what a working flight risk model has to produce to be worth building.

Why Deposit Concentration Reports Catch the Flight After It Happens

A concentration report is a backward-looking record of exposure. It tells ALCO which relationships make up the largest share of deposits and how that mix shifted since the last review, after the shift has already happened. By the time the report lands on the agenda, the balance has moved, the wire has cleared, and the only decision left is how to explain the variance.

This is not a reporting failure. Concentration reports were built to document exposure for examiners and the board, and they do that well. The gap shows up when the same report gets used as an early warning tool, a job it was never designed for. A relationship that is about to leave looks identical to a stable one in last quarter’s numbers, because the signal that it is leaving has not reached the balance sheet yet. It shows up first in behavior: a large depositor calling to ask about current rates, a maturing CD that has not been rolled, a relationship manager noting a competitor’s new promotional rate in a call log. None of that reaches the concentration report until the money is already gone.

What the Interagency Policy Statement Expects You to Already Have in Place

The Interagency Policy Statement on Funding and Liquidity Risk Management already sets the baseline most institutions owe evidence for. Issued 17 March 2010 by the Federal Reserve Board, the OCC, the FDIC, the Office of Thrift Supervision, the NCUA and the Conference of State Bank Supervisors, and revised with an addendum on 1 August 2023, it remains in force today.

The policy statement calls for cash flow projections, diversified funding sources, stress testing and a cushion of liquid assets, scaled to the institution’s size, complexity and risk profile. It also requires a formal, well-developed contingency funding plan with named actions and triggers an institution can execute when funding tightens. The August 2023 addendum reinforces this directly. Institutions should maintain contingency funding plans that are actionable, reviewed and tested on a regular cycle.

A deposit flight risk model does not replace any of this. It gives the cash flow projection and the contingency funding plan a sharper input. The historical average applied across the whole deposit base becomes a named, dated list of relationships at risk of leaving. Examiners reviewing a contingency funding plan tend to ask how an institution would identify a funding shock early. A model that already does this for deposit concentration is a direct answer to that question.

Why General Attrition Models Miss Deposit-Specific Risk

A churn model built for general retail attrition reads a different clock than deposit flight risk does. Retail attrition tends to build slowly, over missed payments, reduced engagement or a competitor’s marketing push, and a model trained on that pattern optimizes for gradual signals. Deposit flight on large or rate-sensitive balances moves faster and on a schedule the institution can often see coming: a CD approaching maturity, a money market rate that has fallen behind the market, a relationship that crosses a size threshold where the depositor starts shopping rates actively.

Treating deposit flight as a subset of general churn misses the signals that are specific to deposits: balance concentration, maturity timing and rate exposure, none of which show up in a typical attrition feature set built around transaction frequency and product usage. A model built for deposits reads those signals directly. It does not need to infer them from behavior patterns designed for a different product.

The Signals That Move Before a Large Depositor Moves Their Money

Four signals carry most of the early warning in deposit flight risk. Balance concentration tracks what share of total deposits a single relationship holds, since concentrated relationships carry outsized impact when they move. Competing-rate exposure tracks how far a relationship’s current rate sits behind what is publicly available elsewhere, since large depositors who actively compare rates act on that gap. Account tenure and relationship depth track how many products and how much history a depositor has with the institution, since newer, single-product relationships move more easily than long-standing, multi-product ones. Quarter-end and seasonal timing tracks the institution’s own historical pattern of when large movements have happened before, since flight risk is rarely evenly distributed across the calendar.

None of these four signals is unusual on its own. What changes the outcome is scoring them together, continuously, against a specific relationship. A manual review of a static list once a quarter cannot keep pace with that.

What a Working Flight Risk Model Has to Produce With Enough Lead Time to Act

A flight risk model only earns its place if it produces a ranked, named list with enough lead time for someone to act on it. A score with no lead time is a postmortem with better math. The output that matters is a list of specific relationships, ranked by both the likelihood of leaving and the size of the balance at risk, delivered far enough ahead of the expected move for treasury or a relationship manager to make a call, offer a repricing conversation or escalate to ALCO.

iTuring.ai’s platform builds this on the same mechanism it applies to deposit growth and retention generally, through the Growth & Retention use case: propensity-based scoring that ranks relationships by risk, personalized outreach timed to each relationship individually, and continuous retraining on what outreach worked. The model does not stop learning once it ships. Every retention call that works or fails becomes training data for the next prediction cycle.

What Deposit Growth Prediction Looks Like in Production

The clearest evidence available for this capability comes from deposit growth. A flight-risk-specific case study is not yet available, so this section shows the same underlying mechanism applied to growth. A credit union with 500,000 members used iTuring.ai’s propensity scoring and personalized outreach to generate $45 million in net new deposits within 90 days. That is a growth result, built on the same mechanism this piece describes for flight risk: score the relationship, time the outreach, retrain on the outcome. This piece does not claim the growth result as a flight-prevention outcome. What it demonstrates is that the underlying mechanism works on deposit relationships at scale, in production, for an institution of comparable size to the community banks and credit unions this piece is written for.

iTuring.ai’s one fully cleared case study outside deposits comes from collections and recovery. A leading NBFC in India using the platform saw a 116% increase in collections recovery rate, 86% predictive accuracy and two-week deployment. It’s a different use case, built on the same propensity-and-outreach mechanism. This result does not substitute for a deposit-specific case study. It supports the claim that the underlying mechanism holds up in production.

See what this looks like for your deposit base

A conversation with our team walks through how propensity scoring and personalized outreach apply to your own deposit concentration, using your own numbers. Book a working session.

Sources

• Interagency Policy Statement on Funding and Liquidity Risk Management, issued 17 March 2010, revised with an addendum 1 August 2023. Federal Reserve Board, OCC, FDIC, Office of Thrift Supervision, NCUA, Conference of State Bank Supervisors. https://www.federalreserve.gov/boarddocs/srletters/2010/sr1006.htm
• iTuring.ai homepage growth tile: Credit Union (500K members), $45M in net new deposits in 90 days. Cleared, live site, anonymized.
• iTuring.ai Growth & Retention use case page, platform description of propensity scoring, personalized outreach and reinforcement learning on outcomes.
• iTuring.ai live-site case study, “Improve Collections and Optimize Efforts” (leading NBFC in India): 116% increase in collections recovery rate, 86% predictive accuracy, two-week deployment. Cleared, live site.