Making uncertainty bands more honest on gap days

Editorial guide prepared September 11, 2026, from the existing September 4 protocol and result. No new outcome evaluation is reported here.

Fit and holdout are separate

The fit dates were January 1, 2022 through June 30, 2024. The holdout dates were July 1, 2024 through August 15, 2026. Each draw sampled 400 anchors; 379 fit and 385 holdout anchors were retained. Across the draws, 35 sets were too short and one anchor lacked its own outcome. The five-session multiplier, 1.48, was fitted before the holdout was evaluated.

The five-session raw holdout coverage was 0.7013, with 90% interval [0.6597, 0.7377]. With the fitted multiplier, coverage was 0.826, with interval [0.7948, 0.8571]. All three primary coverage gates passed. This is historical holdout coverage, not directional accuracy or a guarantee for future forecasts.

Better coverage had a score cost

The reported five-session Winkler statistic, 1.48, is the median of per-subject conditioned/raw score ratios. The separate score medians were 15.181 raw and 21.59 conditioned. Dividing those medians gives about 1.422, which is a different statistic. Lower Winkler scores are better; the reported median per-subject ratio shows a score cost despite improved coverage.

For the 56 earnings-gap observations, raw coverage was 0.8214 and coverage after applying the tested gap multiplier was 0.9107. The result separately mentions an existing earnings conditioner. That mention does not attribute the raw 0.8214 measurement to that conditioner. The non-earnings-gap subset had 329 observations and coverage 0.6809 raw versus 0.8116 conditioned.

The pass does not license one multiplier everywhere

The one-session fitted multiplier was 1.615, with holdout coverage 0.7636. The ten-session multiplier was 1.365, with holdout coverage 0.7714. These are separate horizon-specific fits; 1.48 is not a universal multiplier.

The recorded rollout proposal is a flag, default off, for non-earnings gaps with absolute gap at least 3%. It calls for refitting each horizon on the full 764 retained anchors and a served-coverage receipt over at least 30 sessions before default-on. A study pass is not evidence that those deployment conditions have already been completed.

Read the original result and protocol. Their original wording and the recorded verdict remain preserved.

Reading (2026-09-04 14:20Z, written after the numbers; the bar was fixed before them). PASS at 5 d. Fit half 379 anchors (2022-01..2024-06), holdout 385 (2024-07..2026-08), the fit frozen before any holdout number.

The under-coverage is stable across time, and one multiplier fixes it. Fitted m_5d = 1.48 (fit-half raw coverage 0.662). On the holdout the raw served band covers 0.701 [0.660, 0.738] — under-covering, as required — and with ×1.48 it covers 0.826 [0.795, 0.857]: the interval contains 0.80 and the point sits inside [0.76, 0.84]. The holdout on its own would have needed ×1.35, so the fitted value over-widens by ~9 % — inside the pre-stated window, and on the safe side of the stated 80 %.

Where the under-coverage lives: non-earnings gaps, and up-gaps. Earnings gaps (n = 56 on the holdout) are ALREADY covered by the raw band (0.821) and the single multiplier over-widens them to 0.911; non-earnings gaps go 0.681 → 0.812. Up-gaps 0.644 → 0.777, down-gaps 0.756 → 0.873. So the conditioner belongs on gap days that are not earnings sessions (earnings sessions already carry the earnings conditioner, ×1.10–1.29, which on these numbers is if anything generous), and a sign split is a reasonable refinement for the implementation gate — not tuned here.

1 d and 10 d are less stable than 5 d. Fitted ×1.615 / ×1.365 lift holdout coverage from 0.605 / 0.686 to 0.764 / 0.771, short of 0.80 (the holdout needed ×1.81 / ×1.505). Better than raw, not on target; the implementation gate should refit per horizon on the full sample and carry its own served-coverage receipt.

The honest cost: the Winkler interval score is WORSE with the conditioner (ratio 1.48): the raw band's misses are mostly small, so a proper scoring rule prefers the narrow band that misses 30 % of the time. The product promises a nominal 80 % band; that promise, not the score, is what the conditioner serves. Both numbers are on the record.

What this licenses: a flagged gap_day conditioner in the calibration layer (non-earnings gap sessions with |gap| ≥ 3 %, per-horizon multipliers refit on the full 764, default OFF, gated on its own served-coverage receipt over ≥ 30 sessions before default-on), and the desk v2 input (the agent gets the gap-conditioned width). No direction measured. Runtime 3 min.

Run 2026-09-04T14:12:59.073543+00:00 (smoke=False). Fit 379 / holdout 385 anchors ({'fit_sampled': 400, 'hold_sampled': 400, 'no_embedding': 0, 'retrieve_not_ok': 0, 'short_set': 35, 'no_own_outcome': 1, 'kept': 764}). Pre-registration: research/specs/study103_gap_day_conditioner_prereg_2026_09_04.md (ledger 103).

Fitted multipliers (fit half 2022-01..2024-06)

• 1d: {'n': 379, 'm': 1.615, 'raw_coverage_fit': 0.628}
• 5d: {'n': 379, 'm': 1.48, 'raw_coverage_fit': 0.6623}
• 10d: {'n': 379, 'm': 1.365, 'raw_coverage_fit': 0.686}

Holdout (2024-07..2026-08) with the frozen multipliers

• 1d: {'n': 385, 'm_applied': 1.615, 'raw_coverage': 0.6052, 'raw_ci90': [0.5636, 0.6468], 'conditioned_coverage': 0.7636, 'conditioned_ci90': [0.7273, 0.7974], 'm_holdout_would_need': 1.81, 'winkler_raw_median': 7.757, 'winkler_conditioned_median': 10.787, 'winkler_ratio_median': 1.615}
• 5d: {'n': 385, 'm_applied': 1.48, 'raw_coverage': 0.7013, 'raw_ci90': [0.6597, 0.7377], 'conditioned_coverage': 0.826, 'conditioned_ci90': [0.7948, 0.8571], 'm_holdout_would_need': 1.35, 'winkler_raw_median': 15.181, 'winkler_conditioned_median': 21.59, 'winkler_ratio_median': 1.48, 'conditioned_coverage_earnings_gap': {'n': 56, 'raw': 0.8214, 'conditioned': 0.9107}, 'conditioned_coverage_non_earnings_gap': {'n': 329, 'raw': 0.6809, 'conditioned': 0.8116}, 'conditioned_coverage_up': {'n': 188, 'raw': 0.6436, 'conditioned': 0.7766}, 'conditioned_coverage_down': {'n': 197, 'raw': 0.7563, 'conditioned': 0.8731}}
• 10d: {'n': 385, 'm_applied': 1.365, 'raw_coverage': 0.6857, 'raw_ci90': [0.6468, 0.7247], 'conditioned_coverage': 0.7714, 'conditioned_ci90': [0.7351, 0.8053], 'm_holdout_would_need': 1.505, 'winkler_raw_median': 22.088, 'winkler_conditioned_median': 28.65, 'winkler_ratio_median': 1.365}

Descriptive multipliers on the fit half (5d; none of these is the conditioner)

• fit_m5_up: {'n': 181, 'm': 1.5, 'raw_coverage': 0.6409}
• fit_m5_down: {'n': 198, 'm': 1.45, 'raw_coverage': 0.6818}
• fit_m5_gap_3_5: {'n': 247, 'm': 1.46, 'raw_coverage': 0.6518}
• fit_m5_gap_5_10: {'n': 95, 'm': 1.4, 'raw_coverage': 0.6737}
• fit_m5_gap_gt_10: {'n': 37, 'm': 1.705, 'raw_coverage': 0.7027}
• fit_m5_earnings_gap: {'n': 19, 'm': None, 'raw_coverage': 0.7895}
• fit_m5_non_earnings_gap: {'n': 360, 'm': 1.51, 'raw_coverage': 0.6556}

PRIMARY: PASS

Registered 2026-09-04 ~13:45 UTC, before any number was pulled. Licensed by study 102: on anchors whose session gapped ≥ 3 %, the served p10–p90 band of the production analog set covered the subject's own 5 d excess 0.694 against a nominal 0.80 (up-gaps 0.656), and the state's twins WIDENED were a better band than the event's own twins. Graham: "register study 103 and run it." Question: is the under-coverage stable enough across time that ONE multiplier per horizon, fitted on earlier gap days, restores 80 % coverage on later gap days it never saw?

Ships nothing by itself. A PASS licenses a gated calibration-layer change (a gap_day conditioner, implemented like the earnings conditioner, behind a flag, with its own served receipt). No direction is measured.

Frozen design

• Anchors: liquid (dollar_vol ≥ 5e7, close ≥ 5) situation_context rows with a happening identity, a 1d V5 embedding, and |gap_pct| ≥ 3 on the anchor session. Two INDEPENDENT random draws, seed 103:
  • FIT half: 400 anchors dated 2022-01-01 … 2024-06-30.
  • HOLDOUT half: 400 anchors dated 2024-07-01 … 2026-08-15 (untouched until the multipliers are fixed).
• Served band: the production happening-then-shape analog set at k = 300 (what cohort_analyze serves), as-of cutoff = date − 15 d; the raw p10 … p90 of the members' date-matched excess at 1 / 5 / 10 sessions. Subjects with < 300 members are skipped and counted.
• Conditioner: a scalar m_h on the band's half-width about its midpoint: [mid − m·half, mid + m·half]. Fit: on the fit half, m_h = the smallest m on a 0.3 … 3.0 grid (step 0.005) at which coverage of the subject's own realized excess reaches 0.80. Fitted once per horizon; the 5 d value is the primary.
• Test: apply the fitted m_5d to the holdout half. Coverage with a 2,000-resample bootstrap CI90 over holdout subjects; the raw (unconditioned) holdout coverage the same way.

PRIMARY (one bar, pre-stated), 5 d

PASS iff all three hold on the holdout half:

1. the raw band under-covers there: raw coverage CI90 upper bound < 0.76 (the conditioner is needed);
2. the conditioned coverage's CI90 contains 0.80;
3. the conditioned coverage point estimate lies in [0.76, 0.84]. FAIL otherwise.

Reported either way: m_h at 1 / 5 / 10 d; the multiplier the holdout itself would have needed (drift); the Winkler interval score (α = 0.2) raw vs conditioned on the holdout; descriptive multipliers by gap sign, by |gap| bucket (3–5 %, 5–10 %, > 10 %) and by earnings vs non-earnings gaps (earnings_session.gap_date) — descriptive only, none of them is the conditioner; and the holdout coverage of the earnings-gap and non-earnings-gap subsets under the single multiplier.

Pre-stated readings:

• PASS: the under-coverage is a stable property of gap days → build the gap_day conditioner (flagged, gated on its own served-coverage receipt over ≥ 30 sessions before default-on); the desk v2 input inherits the width.
• FAIL by (1): the holdout no longer under-covers — study 102's finding was period-specific; nothing to fix.
• FAIL by (2)/(3): the under-coverage is real but its size drifts — a fixed multiplier is the wrong shape; the online θ layer (which already adapts per slice) is where it belongs, as a gap-day slice. That is a different registration.

No constant (3 %, k = 300, the grid, the 0.76/0.84 window) is tuned on this sample. The holdout half is not touched until m is fixed from the fit half; both halves are drawn in the same run but the fit is computed first and frozen in the JSON before any holdout statistic is computed.

Machinery

scripts/research/study103_gap_day_conditioner.py run [--smoke] → research/results/study103_*.{md,json,png}. study_ledger row 103. ~6 min at low priority.