flightrisk

Figure 1

The shape nobody expects

Each bar is a group of banks sorted by what their deposits did this quarter. The height is how often that group lost 5% or more of its deposits in the quarter that followed. If the intuition were right, this would slope downward.

Loading FDIC Call Reports, 2015Q1 to 2026Q1

this quarter: -
banks in it: -
dropped next quarter: -
against base rate: -

Why growth predicts a fall

A large inflow is usually one depositor parking a balance that will move again: a municipality, a title company, a corporate treasury after a raise. It arrives as growth and leaves as a drawdown. The read-across is direct and uncomfortable: an account that just received a funding round is a drawdown risk, not a safe one.

It is not a small-bank artifact. Both ends stay elevated in every size band, though the right arm flattens as banks get bigger, where a single depositor is a smaller share of the book.

the mechanism: Lumpy money is transient money.It arrives as growth and leaves as a drawdown.
holds everywhere: All five size bands.The right arm flattens for the largest banks.
what it costs you: Ranking by last quarter's fall.Half the signal is on the other side.

Figure 2

Does the ranking beat the obvious answers

A ranking is only worth having if it beats what you would have done without it. Every model here is scored against the same walk-forward split, alongside four baselines and a control where the labels were shuffled before training.

walk-forward backtest

model: -
alerts that were real: -
against base rate: -
ranking within a quarter: -

The shuffled control is the number to read first. Train the same model on labels that have been shuffled and the top 1% comes back at 0.97x, with a within-quarter AUC of 0.503. The pipeline finds nothing when there is nothing to find, which is the claim every other number on this page depends on.

Ranking by last quarter's decline is worse than random, and still useful. Its within-quarter AUC is 0.406, below a coin flip, while its top 1% carries 3.49x lift. Both are correct: the U inverts the relationship across the middle of the distribution, so a rule that orders the whole list badly can still be right about the extreme.

the control: 0.97x on shuffled labels.AUC 0.503. Nothing found where nothing is.
best in backtest: Gradient boosting, 9.37x.45.6% of the top 1% were real drawdowns.
the one that broke: The linear model, out of time.5.82x to 1.59x, under naive persistence.

Figure 3

What a quarter of output looks like

The ranking is only useful if someone can act on it, so each alert carries the signals that put it there. Fifty a quarter is roughly one analyst's week.

2026Q1, top of the list

Figure 4

Where it loses

Published because a model that only reports where its author wins is an advertisement.

The linear model did not survive the holdout. It scored 5.82x in the walk-forward backtest and 1.59x out of time, below naive persistence. Gradient boosting held at 5.76x. One of these numbers would have been the headline if the holdout had not been run. reports/out_of_time.csv
The macro features add nothing. Adding regime variables moves lift from 9.37x to 9.41x, inside the fold-to-fold spread. A macro-only baseline cannot rank inside a quarter at all, because it is constant within one. reports/backtest_models.csv
Pooled AUC is inflated and is not shown. The backtest writes those columns with INFLATED in the name. A model that only knows which quarter it is scores well pooled while being unable to rank inside any single quarter, which is the only ranking an alert list needs. reports/backtest_models.csv
Bank level is not account level. The question worth answering is which customer pulls their balance. This is the public analog of it: quarterly, not daily, and one row per bank rather than per depositor. README.md
The largest banks are the thinnest slice. Above $100B there are 1,198 bank-quarters against 104,675 under $300M, so the flattening right arm in that band rests on the least data. reports/base_rate_by_size.csv
The alert list here is anonymised. The scores and the model's own reasons are real, but a public ranking of named banks by deposit-flight risk reads as a claim about those banks rather than a demonstration of a method. reports/alerts_2026Q1.csv