Every systematic equity team eventually runs into the same structural catch-22. Textbook quantitative research demands seven to ten years of point-in-time, survivorship-free history across multiple macro regimes before a signal touches live capital. That rule works well for primary regulatory disclosures such as SEC EDGAR XBRL filings, Form 4 insider transactions, Form 8-K item notices, and USPTO patent grants, where server acceptance timestamps are immutable down to the second.
The catch is that big multi-manager platforms have spent decades mining those exact same long-history archives. Residual inefficiencies still exist in the statutory fine print, especially when you isolate subtle cross-filing mechanics and retain delisted issuers, but the margins are earned basis point by basis point.
Meanwhile, fast-moving alternative data streams (specialized federal registry deltas, procurement dockets, clinical trial status changes, or niche operational exhaust) often carry far larger uncrowded dislocations precisely because large pods cannot buy a neat ten-year Parquet backfill from a data broker. Either the data source did not exist five years ago, or its historical archive overwrites status fields in place without preserving what a subscriber actually saw on that Tuesday morning. Backtesting an overwritten archive is worse than having no backtest at all because it manufactures fake certainty out of lookahead bias.
Two Scout Layers Under One Risk Engine
Rather than forcing short-history feeds into bogus backtests or ignoring uncrowded data until its alpha decays, Starling Quantitative splits signal generation into two distinct layers feeding a single deterministic combiner:
| Architecture Layer | Primary Data & History | Capital & Validation Rule |
|---|---|---|
| Layer 1: Proven Foundation | Immutable primary filings (SEC XBRL, Form 4, 8-K, Schedule 13D, 10-K supply chains, FDA clearances, USPTO patents) across 3,607 Russell 3000 CIKs (555 delisted retained). | Validated via 5-fold Purged Walk-Forward CV (12-day embargo). Compounded at +61.76% unlevered OOS (2022 to 2026) vs. SPY +54.33% without margin debt. |
| Layer 2: Ephemeral Sandbox | Differentiated, short-history, or mutable-watermark alternative data feeds captured via live real-time delta polling with immutable arrival timestamps. | Quarantined from historical backtest promotion. Starts at $0 in live shadow behind a One-Way Glass Wall; earns or loses canary weight on forward 6-factor residual alpha. |
The One-Way Shadow and Canary Glass Wall
The fastest way to ruin a profitable core book is to average twenty unvalidated "promising" signals into four proven ones. In linear factor models, noisy experimental features dilute the conviction of graduated signals and trigger unnecessary turnover costs.
To prevent contamination, every ungraduated scout and sub-arm in Layer 2 sits behind a strict One-Way Glass Wall. While a scout is in SHADOW ($0 capital) or micro-canary burn-in, its bullish or bearish feature values are recorded on every live arrival event, demeaned against point-in-time GICS Sector ETFs, and orthogonalized against five OHLCV style factors (Market, 20-day Dollar Volume Size, 12-1 month Momentum, 5-day Short-Term Reversal, and 14-day ATR Volatility). However, probationary scouts are mathematically blocked from inflating or diluting the composite conviction score of the graduated Layer 1 core.
The only exception across that glass wall is asymmetric downside protection: verified statutory distress hazards (such as an auditor pulling reliance on financials in an 8-K Item 4.02 or a late-filing Form NT notice) always retain one-way veto power to block a long entry.
How Fast Should an Ephemeral Signal Graduate or Die?
Managing ephemeral feeds creates a sharp statistical trade-off. If you demand 250 independent live events over three years before allocating a dollar, a short-lived market inefficiency will be arbitraged flat before you ever trade it. If you promote a shadow scout after six green trades in a single tech rally, you end up funding lucky sector beta.
We govern promotion and demotion through three deterministic brakes, with zero human overrides and zero LLM discretion:
1. Date-Clustered Effective Sample Size (Kish N_eff): Fifty signals that all fire during the same two-week earnings or macro wave are two bets, not fifty. Graduation gates count date-clustered effective cohorts rather than raw ticker counts so a single sector wave cannot fake statistical significance.
2. The 50/50 Drop-1 Jackknife Guard: Sparse event distributions are fat-tailed; one +45% biotech buyout can make a losing scout look profitable on sample mean alone. Before any sub-arm can scale above baseline canary sizing, we recompute its residual return after deleting its single best historical trade. If removing that single outlier turns the mean negative, multi-unit upsizing is blocked automatically.
3. Smooth Empirical Bayes Decay: Graduation is a revocable lease, not a permanent tenure badge. Every scout's Grinold Information Coefficient is shrunk toward a zero-skill prior (IC_0 = 0.0) weighted by effective sample size. When institutions eventually crowd into an alternative dataset and its live residual markouts flatten toward zero, its portfolio weight shrinks smoothly back into shadow without waiting for a committee post-mortem.
Read the Underlying Technical Monographs
For the complete equations governing the One-Way Canary/Shadow Glass Wall, Empirical Bayes Grinold shrinkage, and Dual-Maturity Drop-1 Jackknife governance, see Research Notes RN-02 and RN-03: