What was going to be forecast
Each year, the spread of branch deposits inside each US state (FDIC Summary of Deposits): for each state and year, the equal-weighted distribution of ln(branch deposits) minus ln(the national median branch deposit), over branches other than main offices with positive deposits, in the 50 states and DC. Two optimal-transport methods were pre-registered against simple baselines: a pooled Wasserstein autoregression around the yearly barycenter, and a Wasserstein regression on state deposit growth. The plain-language explanation is on the learning page.
The kill rule and the numbers
| Burn-in years (next-year targets) | 1994–2003 (1995–2003) |
|---|---|
| Median next-year W₂² (99-point grid, 459 pairs) | 0.0078208250 |
| Median half-sample W₂² (510 state-years, seed 20261009) | 0.0195766804 |
| Ratio, next-year / half-sample | 0.3994969939 |
| Burn-in target years with next-year below the floor | 9 of 9 |
| K1: kill if next-year ≤ half-sample | fired |
| K3: shape term, next-year 0.0036937468 vs half-sample 0.0093507007 | fired |
| K2: shape share of the move 0.4200 (kill if < 0.15) | did not fire |
| Planning power, primary test (5% gain, Holm 2.5%, T = 22) | 0.0705 |
| Out-of-sample years scored (2004–2025) | 0 |
The half-sample distance is the noise floor: split a state's branches in one year at random into two halves and measure how far apart their quantile functions are. If consecutive years are closer than that, the year-to-year signal is below the noise and there is nothing for a forecast to beat. Any one of the pre-stated kill rules was enough; K1 and K3 both fired. Power is a second, independent reason not to proceed: with about 22 test years and a common year shock across states, a 5% gain would be detected about 7% of the time.
Pre-registration record
| File (tracks/fdic-sod-ot/prereg/) | sha256 |
|---|---|
PREREG_fdic_sod_ot.md | 2f7f7917e484663ecfc7260345de24e947094b886bdfc7f9ef1ade857aaf988c |
fdic_sod_ot_learning.html | d1aced8016107f88a81742239fd334ffc0375fb6753aa63c7abe543483fc2ba7 |
test_design_fdic_sod_ot_v1.json | b9c1aef3086b9dc70e00f74fd88890674d2497bd7a58330c36d92be672627db3 |
GATE_THRESHOLDS_fdic_sod_ot.md | 14018b0c4f44de33f30e276134ddfa42334f64044333db19d5a924ae4f9af4a8 |
PIN.txt | ab4efcd9d6032be439c4585591664fa128db2a1281db9339be1725479c406cc4 |
The design JSON keeps oos_authorized: false, and that flag is not the switch. Because the burn-in
kill fired, an out-of-sample run of this object would not start even if an approval file were written later.
A different object would need its own pre-registration.