Track record
Walk-forward performance record
Run ID 6a9d09710236e6e7. US equities, 1,928-ticker universe, May 2025 – May 2026. All signals generated by the production engine with no parameter changes after seeing results. Exit at lifecycle close, 2% fixed sizing, cash otherwise.
Per-stage performance
Stage is determined at signal emission time from the live theme activation state — igniting, building, broadening, or mature. It does not change retroactively.
| Stage | Signals | % of total | Hit rate | Avg return | SQ gate | Notes |
|---|---|---|---|---|---|---|
| Broadening | 70 | 22% | 56% | +10.3% | ≥ 7 | Confirmed propagation across the cluster — leader plus multiple real confirmers. This is the highest-return stage because the move has already passed the hard test of breadth. |
| Building | 76 | 24% | 56% | +6.0% | ≥ 7 | Leader plus at least one confirmed follower. The structure has begun propagating but hasn't yet reached full breadth. |
| Mature | 120 | 38% | 55% | +4.7% | ≥ 7 | Broad participation established. Largest population by count. Returns are lower because some of the move is already priced in by the time the theme is fully confirmed. |
| Igniting | 50 | 16% | 48% | +1.2% | ≥ 8 | Leader has fired but no confirmed followers yet. The highest-risk stage — more than half are false starts. The SQ gate is tightened to ≥ 8 to partially compensate. |
| All stages | 316 | 100% | 56% | +6.4% | Sharpe 3.65 · max DD 3.3% · CAGR 52.1% |
Signal funnel
~316 trades from a 1,928-ticker universe over 12 months is intentionally low frequency. The engine emits broadly and filters hard. Each stage below represents a meaningful quality gate, not an arbitrary cutoff.
Signal quality score — full breakdown
SQ is computed at the point of signal emission from four sub-scores, each 0 – 5. The only empirically negative cluster is SQ ≤ 6. All other thresholds are based on observed return distribution across the walk-forward.
Theme probability, cluster cohesion, reaction breadth. Penalised 25% if theme class is noise or local move. This sub-score captures whether the structural map agrees the move is real.
Confirmed follower count, reaction breadth across the cluster, simultaneous impulses. Directly measures whether the move has actually spread beyond the leader.
Correlation jump and breadth expansion pace. Capped when breadth is already very high — entry late into a mature move scores down, correctly penalising chasing.
Theme probability, cohesion, news alignment, and structural edge score. Penalised for low confidence or generic theme labels. Lower weight deliberately — narrative can be wrong even when structure is right.
| SQ range | Verdict | Observed characteristics |
|---|---|---|
| 17 – 20 | High conviction | Strong structure, breadth evidence, and narrative alignment across all four sub-scores. |
| 14 – 16 | High conviction | All sub-scores above midpoint. Typically broadening or building stage with confirmed propagation. |
| 10 – 13 | Tradeable | Core of the tradeable population. Quality gate passed with margin. |
| 7 – 9 | Marginal | Passes the gate but expect lower average returns. Igniting signals require SQ ≥ 8 within this band. |
| ≤ 6 | Rejected | 48% hit rate, −1.1% average return. The only cluster in the dataset with negative expected value. Rejected. |
Exit mechanics
The walk-forward uses lifecycle exits, not fixed time stops. Understanding what triggers an exit matters for interpreting the return figures.
The primary exit mechanism. When a within-leader theme goes inactive — the leader loses structural control, confirmers fade, or breadth collapses — Jodie signals the exit. The trader closes the position at that point.
If the same ticker fires a new signal during an active hold, the exit is not triggered by that re-signal. The hold continues until the theme itself goes inactive. This prevents whipsawing out of a continuing move.
No fixed maximum hold. The walk-forward showed median holds of 2–4 trading days for building and broadening stages. Mature themes can hold longer. Igniting signals tend to resolve quickly — either confirmed and promoting to building, or failing back to noise.
Stop-losses, partial sizing, pyramiding, and intraday exit timing are not modelled in the walk-forward. The sizing is 2% fixed per signal with cash for everything else. Real trading would layer these on top.
Known limitations of this record
These are not legal disclaimers. They are real gaps that affect how much weight to put on these numbers.
1,928-ticker universe. International and crypto signals are emitted but not walk-forward validated yet. Do not apply these statistics to non-US signals.
Intraday lagged propagation edges cannot be reconstructed for historical dates, so the SQ score in the walk-forward understates what the live system would have scored. Live SQ values are slightly higher than their replay equivalents for the same formation.
The walk-forward does not model bid-ask spread, market impact, or commissions. At 2% sizing on liquid US equities the impact is small but non-zero.
The intraday_theme_confirmed signal type (member count ≥ 5, rank score ≥ 100) has 6 weeks of live data — 65–66% directional hit rate on that sub-population, but n=29–34. Not included in the main scorecard. Not yet paper-traded.
48% hit rate is below 50%. It is included because it is net positive at 2% fixed sizing across the full sample, and raising the SQ gate to ≥ 8 partially compensates. A more conservative configuration would exclude igniting entirely.
See the methodology behind these numbers
How the structural graph, live signal engine, and quality gate work — and what the engine is and isn't designed to do.