PREREG: Fed CP issuance maturity mix (applied lab; OT is one candidate)

Passed burn-in, pinned, awaiting Jared's approval
No out-of-sample (held-out) run has been made, and none starts without a separate approval file. The design keeps oos_authorized: false, which is a record, not the switch. Rendered from the pinned source tracks/fed-cp-ot/prereg/PREREG_fed_cp_ot.md, sha256 e3da87af2da1002b0635959d751dad4b885c59c985edc3b3a23612fa5bee0ecf · learning page · CP funding lab plan.

Pinned files

Status: BURN-IN GATE PASSED. Out-of-sample (OOS) scoring is NOT authorized (oos_authorized: false).

Verdicts (from compute_cp_ot_burnin.stdout.txt):

• (i) OT_CLAIM = CLEARED. K4 cleared only narrowly; see §4.
• (ii) DATASET_USABLE = yes.
• (iii) FORECASTABLE_SIGNAL (any method) = yes.

Prepared 2026-10-09 (ET). Draft only; not sent anywhere.

1. Files

2. Data

• Source: the XML VOL dataset (see DATA_FACTS.md).
• NONFIN duplication repaired by reading NAA and NA2 from the XML. The CSV pairs equal {NAA, NA2} in 0.999845 of keys. The CSV is used only as a check.
• Primary types: AAA, FAA, NAA, NA2.
• MKT is excluded from the object because it is not the sum of the types (exact match 0.001109). It is used only as a stress comparator.
• Caveats:
  • The volumes cover rate-eligible issues only.
  • The rating criteria changed on 2007-06-18 (Fitch dropped), inside the burn-in.

3. Object and windows

• Object: the weekly (Mon–Fri, Friday label) dollar mix over 6 maturity buckets, for each type.
• Ground metric: log geometric bucket midpoints in days (0.693, 1.903, 2.649, 3.367, 4.048, 4.996). Log-days were chosen over rank because a location shift then reads as a proportional change in maturity. The step between buckets is nearly uniform, so rank is close and is kept as a sensitivity check.
• Weekly, not daily, aggregation. Daily data have day-of-week effects and thin NAA days (a burn-in daily minimum of 0 issues).
• Distance: exact W2² between step quantile functions.
• Windows:
  • Burn-in: weeks ending 2001-01-05..2008-12-26, 417 weeks. It includes the 2007–08 stress so the warm-up sees a stressed regime.
  • OOS: weeks ending 2009-01-02..2026-10-02, 927 calendar weeks. CPFF was active 2008–2010, so the first OOS years are a policy regime.

4. Gate (labelled burn-in values; pooled over 4 types)

How to read this:

• The weekly mix moves about 4× more than the issue-level sampling floor, and about two-thirds of that movement is shape, not location or scale.
• Caveat: the day-split floor is only 22% below the week-to-week change. Much of the week-to-week movement is within-week, day-level variation. That limits how well any model can forecast it. The issue-level floor is a lower bound on noise (it ignores the spread of issue sizes within a bucket).

Forecastable signal (any method; burn-in only)

• EWMA alpha selected on 2002–2004: 0.2 (losses: 0.1: 0.1433, 0.2: 0.1399, 0.3: 0.1425, 0.5: 0.1534).
• Evaluated on 2005–2008 (208 weeks), mean W2²: P 0.2694, EWMA0.2 0.1894, MA4 0.2079, EXP 0.4493.
• Ratios:
  • P/EXP 0.600.
  • EWMA/EXP 0.422.
  • EWMA/P 0.703.
• SIGNAL = yes (threshold 0.90).
• Calm 2005–2006 (non-gating): P 0.2158, EWMA 0.1408, EXP 0.2614.
• Smoothing beats persistence: last week alone is noisy.

Planning power (K4; PLANNING ONLY, not an OOS result)

• Proxy differential d = L(P) − L(MA4), averaged over types:
  • n 413, mean 0.05725, sd 0.15779;
  • AR(1) phi −0.155, residual sd 0.15620;
  • gain scale 0.17873 (the MA4 mean loss).
• Test: T = 927, Bartlett L = 30, fixed-b with an AR(1)-simulated critical value.

• K4: min power at gain 0.05 and 2.5% = 0.5015 ≥ 0.50 → not fired.
  • This is knife-edge. The Monte Carlo SE is about 0.011 (2000 reps), so another seed could fall below 0.50. The pinned rule and seed decide, and the result is CLEARED.
  • Read it this way: the OOS test can detect about a 10% gain reliably, and a 5% gain only about half the time.
• Non-gating contrast (a short-series test at T = 50, Gaussian DGP): power 0.062 at gain 5% and 0.135 at gain 10%. The 927-week length is what makes the test useful.
• The bootstrap null is conservative (size 0.0135 at nominal 0.025).

5. Methods (specified, not fit)

These are as in GATE_THRESHOLDS §5:

• P, EWMA, EXP;
• LS (location-scale on log-duration);
• ALR-AR, ALR-VAR(1), MNL-AR (fractional multinomial logit), RIDGE on lagged ALR;
• WAR(1) (JTSA 2022, DOI 10.1111/jtsa.12590);
• TMAR (transport-map AR, arXiv:2105.05439).

SEL rule (trailing 156 weeks, re-selection every 13 weeks, 1% tie band, then complexity order) decides the shipped app model. If Family B fails, the app headline is persistence.

Background: Fréchet regression (AoS 2019, DOI 10.1214/17-AOS1624; background only) and Wasserstein regression (JASA 2021, DOI 10.1080/01621459.2021.1956937).

6. Trials (TRIAL_COUNT)

• Confirmatory OOS tests: 3.
  • Family A: WAR vs B*, TMAR vs B*, Holm 0.05.
  • Family B: SEL vs P, at 0.05.
• Burn-in gate runs: 1 (no reruns).
• Facts runs: 2. The second added missing-date detail and fixed a mislabel. No threshold depended on the first.
• Outputs printed beyond the thresholds file, all non-gating:
  • per-type medians;
  • the d ACF;
  • the T = 50 contrast;
  • MA4 and alpha-grid losses.
• Stress study: descriptive only, 4 signals × 3 events (+E4). No p-values. Not computed.

7. Leakage and exposure log

1. FS structural facts over all dates: series ids, valid and ND counts, zero-day counts per series for 2001–2026 (e.g. FIN.10_20 has 1764 zero days), and the missing dates.
2. FS consistency checks over all dates: CSV vs XML value-match fractions, and MKT vs sum-of-types fractions. These involve value comparisons over OOS dates, but no distributional statistics.
3. FS burn-in summaries included 2008-12-29..31, 3 days of the first OOS week.
4. BS: OOS valid-day counts per type-week (structural; no amounts).
5. The 2026-10-08 inventory memo mentioned FIN volume detail for 2026-10-01.
6. Web: the Fed CP About page, release and DDP URLs, the FRASER crisis timeline, the Fed 2020-03-17 press release, FDIC PR-16-2023, DFPI, LII 15 U.S.C. 77c.
7. No RATES values have been parsed; only RATES series ids were listed.
8. Drive: read-only metadata search.

8. Authorization

OOS requires a separate APPROVAL_fed_cp_ot.txt, written by the user, that references PIN.txt. The oos_authorized flag is a record, not the switch.

9. Constraints

• Classical statistics only.
• NumPy and pandas, no SciPy.
• Citations come from the verified list; papers older than 2020 are labelled background.