Sigma Vector generates, backtests, grades, and continuously re-validates equity strategies under pinned, reproducible configurations. An AI agent layer turns the survivors into a daily workflow over your own accounts. Analysis and decision support — not a robo-trader.
6 / 53
Strategies clear the full gate
1 / 4
Universes support a defensible ranking
0 / 9
Popular external signals survived the last sweep
10.4 yr
Of walk‑forward history, 2016–2026
01
Strategy generation, backtesting, and re-validation run under pinned configurations. Identical inputs produce identical outputs, byte for byte. No hand-tuning between runs.
Pinned configs · Versioned data tapes02
Every strategy faces an overfitting gate, a deflated performance bar, and an out-of-sample rank-persistence test. Survivors carry an explicit grade — A or provisional — and deployment is gated, not assumed.
PBO · DSR · Rank persistence03
The AI agent layer turns the engine into a daily workflow: graded picks, portfolio review against your actual holdings over a read-only connection, defensive-exit planning. It drafts; you decide.
MCP integration · Read-only syncThat is the design. A pipeline that rejects nothing validates nothing — what ships is what passed the bars, not what looked best in a sweep. Five rules make the rejection mean something.
Hypotheses, scope, and pass–fail bars are written down and frozen before anything is scored. Variants discovered mid-study are deferred to named follow-ups — never quietly promoted.
Go / no-go bars frozen verbatimSignals that reach NO-GO, or fail to reject the null, are archived with their evidence. Nothing on the list gets re-pitched. The list is below.
Standing research ledgerEvery production run pins a canonical configuration and a versioned data tape. The same run, requested twice, is the same run — byte for byte.
Pinned config · Versioned tapesGrades come from walk-forward performance and out-of-sample rank persistence, re-reviewed on a schedule. A strategy that stops earning its grade loses it.
Grade A · Provisional · OutPeriodic red-team passes review the methodology itself — selection bias, leakage, multiplicity. Findings land in the same ledger as everything else.
Last pass 2026-07These are not trophies; they are calibration. The bars that closed these studies are the same bars the deployed strategies passed.
Earnings-surprise drift (PEAD)
No-go
2026-06-20
Below its pre-registered significance bar; the remainder was size, sector, and momentum exposure.
13F institutional accumulation
No-go
2026-06-20
Negative net of costs. Wrong sign, not just weak.
Congressional trading tracker
No-go
2026-06-20
No effect after controls.
Ensemble of all nine swept signals
No-go
2026-06-20
Residual return indistinguishable from zero once factor exposure was removed.
Consensus entry timing
No-go
2026-06-10
Waiting for strategies to agree meant entering after the move. Beat its baseline on 0 of 15 splits.
Defensive-stop overlays, every shape tried
Null held
2026-07-05
Drawdown benefit robust in 1 of 9 configurations; return cost robust in 6 of 9. Not deployed.
An excerpt. Access includes the full ledger — every pre-registration, readout, and closure, with the run IDs to reproduce them. 62 threads on file.
The engine's standing recommendations, drawn only from strategies that hold a current grade.
Trim-or-strengthen analysis of your actual holdings, synced read-only from your broker.
Automated stops failed our own gate — see the list above — so exits are drafted as advisories, not rules. It flags overextended positions; you decide.
Conversational analysis over the engine's evidence and your accounts, through an MCP integration. Answers cite the run they came from.
Leave an email to be contacted when access opens. One message; nothing else.
Recorded. We will write when access opens.