Why separate Signal Analyst, Bear Auditor, and Portfolio Risk Officer roles instead of prompting a single model?

When a single model proposes a trade thesis and evaluates its own downside in one forward pass, autoregressive conditioning biases subsequent tokens toward confirming the initial hypothesis. Isolating an adversarial Bear Auditor with separate evaluation criteria (accrual quality, liquidity depth, and factor crowding) followed by a deterministic non-LLM Risk Officer substantially reduces false-positive capital commitments.

Why anchor multi-agent deliberation in classical factor sleeves rather than unconstrained narrative search?

Unconstrained narrative search lacks a stationary statistical base rate, making Quarter-Kelly position sizing ill-posed. Anchoring proposals in documented empirical anomalies such as cross-sectional momentum (Jegadeesh and Titman, 1993), Post-Earnings Announcement Drift (Bernard and Thomas, 1989), and short-horizon liquidity mean-reversion provides calibrated prior odds before qualitative auditing begins.

Why are live open positions masked on the Starling Quantitative public scorecard?

Starling Quantitative is a proprietary desk rather than an investment newsletter or promoter. Masking ticker symbols and entry prices while positions are open protects proprietary execution intellectual property under 18 U.S.C. 1836 and prevents copycat crowding or individual security promotion, while full ticker symbols and fills are disclosed upon position settlement and audited independently via Collective2.

Adversarial Multi-Agent Quorums & Systematic Factor Consensus

SECTION 01

Single-Pass Confirmation Bias & Error Correlation

A naive architecture in modern algorithmic research asks a single language or statistical model to ingest corporate disclosures, technical price history, and macro headlines inside one prompt, output a directional probability $\hat{p}$, and immediately dispatch an order. In live equity and event markets, this monolithic pattern fails for two mathematical reasons: autoregressive self-confirmation and unanchored base rates.

Once an autoregressive model generates an affirmative thesis token sequence $T^{+}$, subsequent reasoning tokens are conditioned on $T^{+}$, inflating subjective confidence while underweighting balance-sheet accruals, liquidity spread costs, or sector beta crowding. Formally, if $M$ homogeneous prompts share pairwise error correlation $\rho$, ensemble variance fails to vanish as $M \to \infty$:

$$\operatorname{Var}\!\left(\frac{1}{M}\sum_{m=1}^{M} \varepsilon_m\right) = \sigma^2 \left(\rho + \frac{1 - \rho}{M}\right) \xrightarrow{M \to \infty} \rho\,\sigma^2$$

Reducing the asymptotic error floor $\rho\,\sigma^2$ requires structural role orthogonality: separating candidate discovery from adversarial falsification and removing position sizing from probabilistic text generation altogether.

SECTION 02

Empirical Cross-Sectional Factor Priors

Multi-agent deliberation is only as reliable as the statistical base rate feeding Stage 01. Rather than scanning unstructured news narratives in isolation, Starling Quantitative anchors candidate generation in four peer-reviewed empirical factor sleeves with multi-decade out-of-sample literature:

1. Cross-Sectional Momentum & Volatility Contraction

Following Jegadeesh and Titman (1993), intermediate-horizon relative strength persists across 3-to-12-month windows when accompanied by constructive trend alignment and contracting realized return dispersion prior to range expansion. Candidates are screened on trend structure $P_{i,t} > \text{SMA}_{i,50} > \text{SMA}_{i,200}$ paired with the 20-to-60-day Volatility Contraction Ratio:

$$\text{VCR}_{i,t} = \frac{\hat{\sigma}_{i,20\text{d}}}{\hat{\sigma}_{i,60\text{d}}} < \tau_{\text{vcp}}, \qquad R_{i,12-1} = \prod_{s=t-12}^{t-1}(1 + r_{i,s}) - 1$$

2. Post-Earnings Announcement Drift (PEAD)

Following Bernard and Thomas (1989), institutional portfolio committees adjust benchmark weights gradually after material quarterly surprises. Candidate shocks are ranked by Standardized Unexpected Earnings ($\text{SUE}$) and operating cash-flow conversion rather than headline sentiment adjectives:

$$\text{SUE}_{i,q} = \frac{E_{i,q} - \mathbb{E}_{q-1}[E_{i,q}]}{\hat{\sigma}(E_{i,q} - \mathbb{E}_{q-1}[E_{i,q}])}$$

3. Short-Horizon Oversold Mean-Reversion

At 3-to-5-day horizons, non-fundamental liquidity shocks temporarily depress liquid large-cap equities below short-term equilibrium value (Lehmann, 1990). Candidate dislocations combine Wilder’s 14-period Relative Strength Index ($\text{RSI}_{14} < 30$) with standardized 20-day Bollinger excursion $z_{i,20} = (P_{i,t} - \mu_{i,20}) / \sigma_{i,20} < -2.0$, strictly conditioned on absence of credit distress or accounting restatement.

4. Calibrated Event & Prediction Market Pricing

For binary and multi-outcome event contracts, raw dealer and order-book asks are first normalized via Power ($\sum_i \pi_i^k = 1$) and Shin ($z$) devigging as derived in Research Note RN-01 before comparing internal ensemble forecasts against market consensus.

SECTION 03

The Three-Stage Adversarial Quorum Protocol

Every candidate generated by an empirical factor sleeve traverses three sequential stages before an order can reach execution routing:

STAGE	COMMITTEE ROLE	PRIMARY MANDATE & OUTPUT
01	Signal Analyst	Quantifies factor z-score, catalyst horizon, historical base-rate prior $p_0$, and explicit thesis invalidation conditions.
02	Bear Auditor	Adversarial counter-review: audits working-capital accruals, upcoming binary events, bid-ask spread drag, and sector crowding; emits likelihood penalty $\Lambda_{\text{bear}} \le 1$.
03	Portfolio Risk Officer	Deterministic non-LLM engine: applies Bayesian shrinkage, Quarter-Kelly sizing ($c = 0.25$), sector covariance caps, and predefined stop-loss boundaries.

In log-odds space $\ell(p) = \ln\!\big(p / (1 - p)\big)$, the Risk Officer updates the empirical sleeve base rate $\ell(p_0)$ with shrinkage weight $w \in (0, 1)$ and the Bear Auditor’s downside penalty $\ln \Lambda_{\text{bear}} \le 0$:

$$\ell_{\text{post}} = \ell(p_0) + w\,\Delta\ell_{\text{signal}} + \ln \Lambda_{\text{bear}}, \qquad \hat{p} = \frac{1}{1 + e^{-\ell_{\text{post}}}}$$

Whenever $\ln \Lambda_{\text{bear}}$ pushes net expected edge below transaction and slippage hurdles, the candidate is vetoed deterministically before reaching order routing.

SECTION 04

Rebalance Cadence & Predefined Risk-Stop Governance

Two operational failure modes erode systematic factor returns in retail implementations: excessive intraday churn and unbounded loss retention when a fundamental thesis breaks.

In standard institutional factor construction, cross-sectional ranks are evaluated on disciplined periodic schedules (such as end-of-day universe snapshots queued for scheduled T+1 open rebalance windows or fixed periodic intervals) rather than reacting impulsively to every tick fluctuation. Periodic evaluation ensures full cross-sectional breadth across the investable universe and allows the three-stage Bull/Bear/Risk committee sufficient deliberation compute to audit filings thoroughly.

Simultaneously, every approved position carries a hard, predefined maximum adverse excursion boundary from entry (for example, classical institutional risk discipline capping single-position price loss at 7% to 8% from entry cost) paired with Quarter-Kelly portfolio sizing (Research Note RN-02). If $s \in (0, 1)$ denotes the predefined stop distance and $r_{\max}$ denotes the maximum portfolio equity risked on a single thesis, notional portfolio weight $w_i$ satisfies:

$$w_i = \min\!\left(\,c \cdot f_i^*, \;\; \frac{r_{\max}}{s_i}, \;\; w_{\text{sector,max}}\,\right), \qquad c = 0.25$$

Capital not actively deployed across approved factor positions is swept systematically into liquid short-duration U.S. Treasury Bill ETFs (SGOV), earning risk-free carry while preserving immediate dry powder during elevated market volatility regimes.

SECTION 05

Adaptive Sleeve Weighting via Discounted Thompson Sampling

Factor efficacy rotates across macroeconomic regimes: momentum breakouts thrive in low-dispersion trends, whereas oversold mean-reversion dominates range-bound liquidity shocks. Rather than hardcoding static sleeve allocations, Starling Quantitative models each sleeve’s out-of-sample payoff precision with an exponentially discounted posterior ($\lambda = 0.98$):

$$\alpha_{k,t} = 1 + \lambda\,(\alpha_{k,t-1} - 1) + \mathbb{I}(R_{k,t} > 0), \qquad \beta_{k,t} = 1 + \lambda\,(\beta_{k,t-1} - 1) + \mathbb{I}(R_{k,t} \le 0)$$

With forgetting factor $\lambda = 0.98$ (effective memory horizon $1 / (1 - \lambda) = 50$ settled trades), sleeves experiencing calibration decay automatically cede risk budget toward sleeves demonstrating positive out-of-sample resolution without abrupt regime-switch overfitting.

SECTION 06

Scope of Public Foundations vs. Confidential Trade Secrets

Starling Quantitative practices selective institutional transparency: we publish our statistical factor foundations, probability calibration mathematics, multi-agent governance structure, and third-party audited equity curve openly, while retaining our proprietary execution timing, order-routing microstructure, and internal risk thresholds as confidential trade secrets of Biwada LLC under 18 U.S.C. § 1836.

For the same legal and market-impact reasons, active OPEN positions on the Live Progress Scorecard mask ticker symbols and entry prices until positions are closed in full. Once closed, settled trades are disclosed in the historical audit log and verified independently via Collective2 Strategy #158114255.

SECTION 07

Technical FAQ

Why separate Signal Analyst, Bear Auditor, and Risk Officer roles instead of prompting one model?

Single-pass inference conditions risk evaluation on tokens that already endorsed the trade, inflating conviction. Isolating an adversarial Bear Auditor tasked strictly with falsifying the thesis, followed by a deterministic mathematical Risk Officer, decorrelates specification errors before capital is committed.

Why anchor multi-agent deliberation in classical factor sleeves?

Peer-reviewed anomalies such as cross-sectional momentum, Post-Earnings Announcement Drift, and short-horizon liquidity mean-reversion provide stationary empirical base rates $p_0$, ensuring Quarter-Kelly sizing operates on calibrated priors rather than unanchored narrative probabilities.

Why are live open positions masked on the public scorecard?

Masking active tickers prevents individual security promotion and protects proprietary execution intellectual property while positions are live in the market. Full ticker identities and realized returns are revealed upon settlement and tracked independently on Collective2 (#158114255).

AUDITED FORWARD LEDGER // COLLECTIVE2 #158114255

Inspect Realized Equity Curve, Benchmark Comparison & Settled Trade Audit

VIEW LIVE SCORECARD →