Label calibration — what each classification has historically preceded (measured, not promised)
protocol 8a1df3604cc7 · forward ledger only · registered before computation · benchmark-relative outcomes at k=20
| Label | Direction class | n outcomes | Mean excess k=20 | Direction consistency | Badge |
|---|
| NEUTRAL | none | 1897 | 0.27% | — | DESCRIPTIVE |
| WATCH | none | 1168 | 1.94% | — | DESCRIPTIVE |
| BULLISH | positive | 658 | 3.08% | 0.63 | DESCRIPTIVE |
| WEAKENING | negative | 547 | -3.95% | 0.61 | DESCRIPTIVE |
| STRONG_BULLISH | positive | 136 | -10.39% | 0.09 | PRELIMINARY |
| NEGATIVE_EVENT | negative | 21 | -5.71% | 0.75 | UNDERPOWERED |
Registered primary: pooled direction-consistency of the directional labels at k=20 =
0.555
CI (0.467, 0.645) [DESCRIPTIVE] over n=921 outcomes /
67 clustered tickers — the interval includes 0.5, so directional meaning is not
yet demonstrated at this sample; that is the finding, printed as measured. STRONG_BULLISH reads
anti-consistent at k=20 at current n — also printed as measured; the record accrues daily.
Labels with no forward outcomes yet: BEARISH_WATCH, RISK_ALERT. Non-directional labels carry
no consistency claim by construction. The score-to-label mapping is fixed thresholds,
published in the
methodology (score-to-label thresholds) — the thresholds carry no outcome promise;
this panel measures the outcomes. Investment implication: none established — no buy, sell,
or alpha conclusion is supported by this page.
DISCLAIMER — Research / education only. Not investment advice. Past results — in-sample or forward-tracked — do not predict future performance. YUCLAW is not a registered investment adviser.
FORWARD TRACKING LEDGER — Out-of-Sample · data through 2026-08-03
date range 2026-05-20 → 2026-08-04 · total signals 4585
· directional 1390
| label | n | directional |
matured 1d/5d/20d |
hit 1d/5d/20d |
median return 5d |
median excess 5d |
| BULLISH | 662 | 662 | 660/646/539 | 58% (n=660) / 57% (n=646) / 64% (n=539) | +0.6% | +0.3% |
| NEGATIVE_EVENT | 21 | 21 | 21/19/4 | 62% (n=21) / 37% (n=19) prelim / 75% (n=4) prelim | +2.3% | +1.6% |
| NEUTRAL | 1951 | 0 | 1922/1815/1363 | — / — / — | +0.2% | +0.0% |
| STRONG_BULLISH | 136 | 136 | 136/136/128 | 43% (n=136) / 26% (n=136) / 9% (n=128) | -2.9% | -1.8% |
| WATCH | 1244 | 0 | 1207/1069/670 | — / — / — | +0.2% | +0.1% |
| WEAKENING | 571 | 571 | 560/505/301 | 45% (n=560) / 51% (n=505) / 59% (n=301) | -0.2% | -0.0% |
| OVERALL | 4585 | 1390 | 4506/4190/3005 | 51% (n=1377) / 51% (n=1306) / 55% (n=972) | +0.2% | +0.0% |
Frozen historical window (2026-02-18 → 2026-05-13) — this regime closed when forward tracking began and is never updated, by design.
IN-SAMPLE EVENT VALIDATION — Replay
date range 2026-02-18 → 2026-05-13 · total signals 1027
· directional 235
⚠ Note on the in-sample panel: market-data components C1 / C3 / C4 / C5 / C7 ran at 0.3 confidence because the upstream dashboard cache holds only the latest snapshot. Evidence components C6 (events), C8 (cascade), C9 (model trust) are point-in-time exact. The in-sample event validation therefore primarily reflects the evidence layer.
| label | n | directional |
matured 1d/5d/20d |
hit 1d/5d/20d |
median return 5d |
median excess 5d |
| BEARISH_WATCH | 1 | 1 | 1/1/1 | 0% (n=1) prelim / 0% (n=1) prelim / 100% (n=1) prelim | +0.5% | -1.1% |
| BULLISH | 50 | 50 | 50/50/50 | 40% (n=50) / 64% (n=50) / 82% (n=50) | +1.5% | +1.5% |
| NEGATIVE_EVENT | 37 | 37 | 37/37/37 | 65% (n=37) / 57% (n=37) / 51% (n=37) | -1.2% | -0.6% |
| NEUTRAL | 339 | 0 | 339/339/339 | — / — / — | +0.1% | -0.3% |
| WATCH | 453 | 0 | 453/453/453 | — / — / — | -0.1% | -0.4% |
| WEAKENING | 147 | 147 | 147/147/147 | 56% (n=147) / 50% (n=147) / 37% (n=147) | +0.1% | -0.2% |
| OVERALL | 1027 | 235 | 1027/1027/1027 | 54% (n=235) / 54% (n=235) / 49% (n=235) | +0.1% | -0.3% |
METHODOLOGY
Full methodology, data sources, limitations, and definitions are documented at
docs/methodology/backfill.md.
Key points: data window is 2026-02-18 to 2026-05-17 (~90 days, all post-Llama 3.1 70B
training cutoff); in-sample signals were reconstructed via point-in-time replay
(leak-audited, zero future-event leakage); the Forward Tracking Ledger begins
2026-05-20 and matures daily — 20-day outcomes mature in mid-June.
DISCLAIMER — Research / education only. Not investment advice. Past results — in-sample or forward-tracked — do not predict future performance. YUCLAW is not a registered investment adviser.