The CANRE Counterfactual
Deposit betas are not wrong so much as incomplete: in the 2022 rate cycle, the modeled numbers held well through the first 100 to 150 basis points of tightening and only broke down further out, which means deposits are convex, and any constant-beta model can be correct only locally.
The same lens that reveals this also separates a weak franchise from a merely rate-sensitive one. That is what CANRE does.
Key Highlights
- The betas were locally right and globally incomplete. The modeled deposit betas held reasonably well through the first 100 to 150 basis points of tightening, then diverged sharply. That is not a calibration error; it is convexity. A constant beta is fine near the starting point and increasingly wrong the further rates travel.
- The full-cycle miss was still large, and it has a price. Across a 16-bank peer group the median realized up-cycle beta was about 68% against a modeled 31%. On a $50B interest-bearing book, that gap is roughly $850M of unmodeled annual funding cost.
- History offers an encouraging precedent on the way down. Older cycles repriced symmetrically and low. The one recent down-cycle we can measure on a true deposit-cost basis, 2015 to 2020, gave cost back at 59% against a 47% up-beta, faster than it went on, and completed as rates reached zero. Deposits were not structurally sticky on the way down.
- What changed this cycle is the level, not the direction. Rate-sensitivity roughly doubled. The open question is whether that step-up is permanent, and whether it stays symmetric on the way down. History cannot answer this because it never saw a cycle this rate-sensitive.
- Falling NIB% is a diagnostic, not a verdict. A steep decline in non-interest-bearing share signals rate-sensitive customers. It says nothing, by itself, about whether those customers stayed. CANRE supplies the missing half.
- Two banks, same NIB story, opposite franchises. Bank A grew total deposits 21% and replaced every lost NIB dollar with $2.29 of new interest-bearing money. Bank B’s deposit base contracted about 32%. Identical surface signal; opposite conclusion.
- The lesson is process, not blame. Model beta as a curve that varies with the rate level, monitor realized behavior against a live peer benchmark, and hold an explicit view on how fast cost comes back on the way down. That is the control that keeps NIM forecasting credible.
