YearBull is measurably different from a market-cap list or a momentum leaderboard. This document shows where YearBull Rank added signal in the current evidence, how stable the ranking is, where a simple baseline still led, and how the next prospective test is locked before outcomes exist.
1. Retrospective baseline comparison
Seven monthly formation dates from February 2 through August 1, 2026 were evaluated. Each date used the ordinary analytical universe with an exact 30-day lookback and an exact 30-day forward price. The top 10% under each signal was compared with market-cap rank, 30-day momentum, and lowest 30-day realized volatility. The pooled top cohorts contain 3,521 observations; the comparison universe contains 35,224 observations.
| Signal / cohort | Mean 30d return | Median 30d return | Positive return | Observations |
|---|---|---|---|---|
| YearBull Rank top decile | −0.1% | −3.5% | 40.3% | 3,521 |
| Market-cap top decile | −1.2% | −4.0% | 39.4% | 3,521 |
| 30-day momentum top decile | −0.8% | −13.4% | 33.4% | 3,521 |
| Low-volatility decile | +7.5% | −0.9% | 41.6% | 3,521 |
| Eligible universe | +4.5% | −5.2% | 39.2% | 35,224 |
Interpretation: YearBull achieved useful differentiation rather than reproducing the two most obvious rankings. Its median result and positive-return rate were ahead of both size and recent momentum in this sample; the momentum comparison was especially distinct at −3.5% versus −13.4% median return. Low volatility remained stronger, and the highly skewed universe mean shows why means alone are a poor guide. Monthly results changed across regimes, so the prospective test—not a selective retrospective month—is the next decision point.
Evidence boundary: formation dates overlap, fees and execution are not modeled, and historical model versions may differ. The table supports comparison and model development; it is not presented as a tradeable-strategy result.
2. Sensitivity and stability
Stable enough to trust, responsive enough to matter
Across the latest 38 completed dates through September 22, the mean one-day Spearman rank correlation was 0.923 and mean top-decile overlap was 77.5% (37 comparisons). At seven days, correlation was 0.716 and top-decile overlap 52.8% (31 comparisons). That is the intended behavior: YearBull does not churn on daily noise, but it does re-order the field as market structure changes.
Leave-one-family-out diagnostic
To test factor dependence without publishing the production equation, YearBull created an equal-weight diagnostic proxy from the disclosed normalized factor families and removed one family at a time. On 9,508 analytical assets, rank correlation with the full proxy stayed between 0.902 and 1.000; top-decile overlap ranged from 51.2% to 98.6%. The public factor architecture therefore remained coherent under every single-family removal, while the larger top-cohort change after removing volume quality confirms that liquidity support is meaningful rather than decorative.
This diagnostic uses a public proxy, not the proprietary production equation. Its role is to expose structural dependence without giving away coefficients or anti-gaming controls; full precision is available in the CSV.
3. The next proof standard: prospective out-of-sample testing
Registration date: September 22, 2026, before the evaluated forward outcomes exist. Locking the rules in advance turns the next result into a real test rather than a story selected afterward.
- Formation window: September 22, 2026 through February 19, 2027.
- Outcome window: 30 calendar days after each formation snapshot; final outcomes mature by March 21, 2027.
- Sampling: the first completed snapshot on or after each Monday; ordinary analytical assets with valid formation and lookback data.
- Cohorts: YearBull top decile, market-cap top decile, 30-day momentum top decile, and lowest 30-day realized-volatility decile.
- Primary measures: median 30-day forward return, positive-return rate, and the median-return spread versus each baseline.
- Missing outcomes: delisted, unmapped, or unavailable forward observations remain reported as missing; they are not silently converted to winners or removed from coverage.
- Model changes: a material model-version change starts a separately labelled cohort. It must not overwrite the original track.
- Interim: after December 21, 2026. Final report: after March 21, 2027, once the last registered outcomes mature.
4. Known errors and failure cases
| Failure case | Why it matters | Control / residual risk |
|---|---|---|
| Incorrect identity or ticker collision | History can be attached to the wrong asset | Canonical IDs, contract evidence, collision labels, corrections; residual mapping risk remains |
| Wrapped, bridged, staked, or pegged representation | Not comparable with an independent growth asset | Separate infrastructure/stable states; edge classifications can still be disputed |
| Frozen or stale price/volume | Creates false stability or momentum | Freshness checks and not-ranked states; upstream revisions may arrive later |
| Reported volume without executable depth | Can overstate liquidity | Quality and concentration controls; order-book execution is not fully observed |
| Missing or uncertain supply | Distorts market cap and dilution context | Explicit missingness and penalties; no estimate can manufacture reliable supply |
| Short history | Incomplete lookbacks increase false precision | Progressive history-confidence discount; early observations remain less certain |
| Regime shift | Historical relationships may reverse | Market-regime adjustments and prospective monitoring; structural breaks remain possible |
| Concentration and outliers | A few assets can dominate means | Report medians, coverage, and cohort counts alongside means |
5. Model drift and monitoring
YearBull monitors input coverage, factor distributions, rank turnover, classification transitions, reference-cohort composition, missingness, and divergence from simple baselines. A material unexplained shift triggers investigation before interpretation is changed. Confirmed defects are corrected under the Corrections Policy; calculation or classification changes enter the Methodology Changelog.
6. Downloads and reproducibility boundary
Download validation summary (CSV) Download data dictionary (CSV)
The downloads expose cohort definitions, sample sizes, measures, full-precision results, field meanings, and limitations. They do not expose exact production coefficients, transformations, anti-gaming rules, or proprietary thresholds. See the Methodology Paper for the public method boundary.