Decision Methodology

What the system is looking for, how it scores candidates, and the academic + market-regime rationale behind every weight. Every agent prompt operationalizes this document — change it and the model behavior changes accordingly.

This document defines what the system is looking for and how it scores candidates. Every agent prompt operationalizes this document. If you change the methodology, update both this file and the agent prompts together.

Goal recap: Find SanDisk-style asymmetry (1400% in a year) — not steady-Eddie picks. Default behavior is silence. Publish ~1-3 times per quarter at most, only when something genuinely clears a high bar.

1. What works (research-backed signal universe)

The signal weights below are calibrated against academic literature plus 2025-2026 market regime observations (AI capex cycle, hyperscaler concentration, US-China decoupling, memory/optical/power infra tightness). Weights are deliberately conservative — many "anomalies" don't survive transaction costs at retail scale, so we lean hard on the most-replicated effects.

Tier 1 — strong evidence, high weight

Insider open-market buying clusters (Lakonishok-Lee 2001, Cohen-Malloy-Pomorski 2012, Akbas et al. 2014)

Multiple insiders, open-market purchases (NOT options exercises), within ~90 days. Documented alpha 4-8% per year, concentrated in small/mid caps.
CEO/CFO buys carry the strongest signal; clusters of 3+ unrelated insiders are highest-quality.
Form 4 transaction codes matter: P (open-market purchase) is what we want. A (grant), G (gift), F (tax withholding) are noise.
2026 caveat: Insider buying is rare in this regime (most insiders are net-sellers due to RSU vesting); when it happens it's an unusually strong signal.

Earnings surprise drift (PEAD) (Bernard-Thomas 1989, robust through 2024)

Positive earnings surprise → 60-90 day price drift. Effect ~3-7% on the surprise quartile.
Particularly potent when paired with guidance raises and analyst-estimate revisions.
Best operationalized as: catalyst in window + recent positive surprise pattern + improving margins.

Fundamental quality + cheap valuation (Asness/Frazzini/Pedersen "Quality minus Junk" 2018)

High ROIC + low debt + earnings stability = quality factor. Long quality, short junk pays ~5% per year.
Combined with cheap (FCF yield, EV/EBITDA), filters most fads. The real prize: cheap + improving quality.
2026 application: In a market where hyperscaler-adjacent names trade at 30-50x forward P/E, finding a structurally improving business at <15x is the value-style edge.

Tier 2 — moderate evidence, medium weight

Unusual options activity (UOA) / whale flow (Pan-Poteshman 2006, Cao-Chen-Griffin 2005, Hu 2014)

Large institutional positioning often shows up first in options markets before stock moves. Documented effect: option volume imbalances predict next-day returns (Pan-Poteshman); pre-takeover call buying surges weeks ahead (Cao-Chen-Griffin).
Specific signals: volume/OI ratio > 1 (new positioning, not just rebalance); whale blocks (>1000 contracts) at OTM strikes; sudden put/call skew flips; sweep orders on the offer.
Implementation gap: Real UOA requires a feed (Cheddar Flow, Unusual Whales, FlowAlgo, Quiver). Free Yahoo data gives us volume+OI on the chain, which lets us detect crude UOA — strikes where today's volume materially exceeds open interest. Better detail is a v2 paid-feed integration.
Treat as Tier 2 confirmation — strong corroborator of an insider/13F thesis, weak as a standalone driver because retail can mis-read large institutional hedge flow as bullish positioning.

Concentrated 13F initiations from skilled managers (Cohen-Polk-Vuolteenaho 2010 "Best Ideas")

Top 1-3 positions of historically successful managers outperform their diversified benchmark.
Notable trackers: Berkshire, Pershing Square, Appaloosa, Greenlight, Scion (Burry), Pabrai, Polen.
2026 caveat: 13F lag (45 days) limits use as a leading signal but useful for confirmation.

News-derived sentiment (Tetlock 2007, Loughran-McDonald 2011)

Pessimism in financial media predicts forward returns (mean-reversion). Modern: FinBERT, LM dictionaries.
Most useful as a contrarian signal at extremes, OR as confirmation of catalyst momentum.

Options skew & put/call ratio (Cremers-Weinbaum 2010)

Call IV minus put IV predicts cross-sectional returns. Bullish skew → outperformance.
Useful for confirmation and structure selection (high IV → CSP attractive).

Politician trades — STOCK Act disclosures (Ziobrowski et al. 2004; mixed since)

Ziobrowski 2004 showed Senate portfolios outperformed S&P by 12% annualized. Later studies (Eggers-Hainmueller 2013) attribute most of that to sector concentration, not stock-picking skill.
Modern context (Pelosi tracker, Capitol Trades, Quiver Quant): popular but low-quality as a standalone signal. Disclosures lag 30-45 days; many are spouse trades.
Treat as a low-weight confirmation signal, not a primary driver. Disclose to readers when an idea is influenced by it.

Tier 3 — weak evidence, low weight (use as confirmation only)

Pure technical indicators (MACD, RSI, MAs) (Lo-Mamaysky-Wang 2000; Park-Irwin 2007)

Most effects are small after transaction costs at retail scale. For 3-12 month holds, technicals matter mostly for timing the entry, not generating the thesis.
Used as: don't enter when RSI is 80+ on a parabolic chart; prefer MACD cross confirming trend.

Short interest (Asquith-Pathak-Ritter 2005, Diether-Lee-Werner 2009)

High SI alone is ambiguous. High SI + improving fundamentals = squeeze setup (rare but spectacular). High SI + deteriorating = avoid.
2026 application: Squeeze setups are loud and noisy; the true SanDisk-class trade rarely had high SI at entry. Use SI as a risk indicator, not a primary signal.

Anti-signals (gating / red flags)

Going-concern audit qualifications
Pending material litigation (DOJ, SEC enforcement, class action with merit)
Customer concentration > 30% (single customer can break thesis overnight)
Convertible/warrant overhang creating dilution risk > 10% of float
Accounting irregularities (restatements, auditor changes, SOX issues)
Liquidity below $5M average daily volume (can't size meaningfully)
Penny-stock or micro-cap pump signals (sudden volume + chat-room mentions, unverified press releases)

2. The Catalyst-Edge Framework

Every publishable idea satisfies all three:

Mispricing

The market price is meaningfully below a defensible intrinsic value estimate. "Defensible" means: estimate built from filings + fundamentals + sector comps, not from extrapolating recent growth.

Operationalized:

Forward P/E below sector median by ≥20%, OR
EV/EBITDA below sector median by ≥30%, OR
FCF yield ≥ 7% with stable/growing FCF, OR
Sum-of-the-parts narrative where reported earnings understate intrinsic value (e.g., R&D-heavy capitalize-vs-expense distortions, hidden segments)

Catalyst

Something specific within the time horizon (3-12 months) that forces a re-rating. Without a catalyst, mispricing can persist forever.

Operationalized — must name AT LEAST ONE:

Earnings within 90 days where recent surprises are positive
Product launch / FDA decision / regulatory milestone with date
Contract / partnership announcement that materially changes revenue trajectory
Spin-off, reorganization, or strategic review
Sector inflection (e.g., HBM4 ramp, hyperscaler capex acceleration)
M&A activity in the comp set (re-rating by association)

Edge in data

We saw something the market missed. This is the hardest and the most important. If the dossier doesn't articulate this clearly, the trade isn't real edge — it's just a factor exposure.

Examples of real edge:

Read the actual 10-K and found segment data showing one division printing 60% gross margins inside a 25%-overall company
Cross-referenced Form 4s and saw the CFO buying shortly before guidance was raised
Listened to 4 earnings calls and noticed the same incremental positive language that hasn't shown up in analyst notes yet
Sector capacity model shows the Street's revenue estimates are based on FY24 capacity assumptions when the company has already announced FY26 capex doubling

Examples of NOT edge (skip):

"It's cheap on P/E" — too well-known to be edge
"Buybacks are coming" — well-known to be edge
"AI tailwinds" — definitely not edge
"Technicals look great" — at best timing, not thesis

3. Scoring rubric

Each idea gets a composite score 0-100. Components:

Category	Max Pts	Sources
Smart-money cluster	25	insider Form 4 cluster + 13F deltas + politician trades
Options flow / UOA	10	options chain V/OI, whale blocks, IV skew
Catalyst	25	earnings calendar + filings + news + sector context
Mispricing	15	fundamentals + peer comparison + SOTP analysis
Quality	15	ROIC, margins, balance sheet
Momentum/technical	10	price history + MACD + RSI + MAs

Smart-money cluster (max 25)

Insider open-market purchases by 3+ unrelated officers/directors in last 90 days totaling ≥$500K: 10 pts

Add 3 pts if CEO is among the buyers
Add 3 pts if CFO is among the buyers
Add 2 pts if total purchases > $5M

Concentrated 13F initiation or material add by a top-10 known manager (Berkshire, Pershing Square, Appaloosa, Greenlight, Scion, Polen, Pabrai): 5 pts

Add 3 pts if the manager has put it in their top-3 positions

Politician STOCK Act disclosure showing recent buy: 2 pts (low weight by design — disclosure is delayed and noisy)

Add 1 pt if multiple unrelated members within 30 days

Penalty: −10 pts if material insider selling exceeds material insider buying

Options flow / UOA (max 10)

Whale call buying within 30 days at a relevant expiry, OTM strikes, V/OI > 1: 4 pts

"Whale" = blocks ≥ 1000 contracts at a strike OR sweep orders ≥ $250K notional
V/OI > 1 indicates new positioning, not just turnover

Bullish IV skew vs sector (call IV > put IV) over rolling 5 days: 3 pts
Sudden change in put/call ratio (latest day vs trailing 30-day avg) consistent with thesis direction: 2 pts
Large open-interest building at a specific strike implying defended price level: 1 pt
Penalty: −5 pts if recent flow is meaningfully bearish (whale put buying, skew flipped to puts) without a hedging explanation
Implementation note: Without a UOA feed, we infer flow from yfinance chain V/OI ratios and large OI clusters. This is crude. Score conservatively. A paid UOA feed (v2) would let us be more confident.

Catalyst (max 25)

Earnings within 90 days AND last 4 quarters' EPS surprise > 0%: 10 pts
Pending product launch / FDA / contract milestone within horizon: 8 pts
Recent guidance raise (last 90 days): 5 pts
Sector tailwind quantifiable in $ (e.g., named hyperscaler capex guide that flows to this name): 5 pts but capped at total 25

Mispricing (max 15)

Forward P/E ≥ 20% below sector median (with non-deteriorating earnings): 5 pts
EV/EBITDA ≥ 30% below sector median: 4 pts
FCF yield ≥ 7% with stable FCF: 4 pts
SOTP gap ≥ 25% (segment-level analysis from 10-K): 5 pts (replace one of the above if used)

Quality (max 15)

ROIC ≥ 15% (or trending there with cash flow visibility): 5 pts
Gross margin expanding ≥ 3 percentage points YoY: 5 pts
Net debt / EBITDA ≤ 2x (or net cash): 5 pts

Momentum / technical (max 10)

Stock above both 50DMA and 200DMA: 3 pts
12-1 month price return positive AND beating sector index: 4 pts
RSI in 40-65 range (not extreme): 2 pts
MACD signal-line cross to upside in last 30 days (confirmation): 1 pt

Composite thresholds (CORE risk class)

Score	Decision	Confidence	Position size cap
0–44	Skip	—	—
45–59	Paper-track if specific/measurable and no hard gate; otherwise skip	1–2	0.0%
60–69	Publish	3	1.5–2.0%
70–79	Publish	4	2.5–3.0%
80–89	Publish	5	4.0–5.0%
90+	Publish; unusually rare	5	5.0%

Composite thresholds (ASYMMETRIC risk class — small-cap moonshots)

The ASYMMETRIC bucket is a different game. Sub-$5B mcap names where the edge is asymmetric upside (10x+ possible) but base-rate of losses is brutal (90%+ go to zero or stay flat). A diversified portfolio of 30-50 asymmetric bets sized at 0.5-1.5% each is how this plays profitably; a 3% bet on one is how you blow up. Compliance enforces these caps in code.

Score	Decision	Confidence	Position size cap
0–44	Skip	—	—
45–59	Paper-track only with a concrete dated catalyst/event path; otherwise skip	1–2	0.0%
60–69	Publish	3	0.5%
70–79	Publish	4	1.0%
80+	Publish	5	1.5%

Note: NO asymmetric idea exceeds 1.5%, regardless of conviction. The expected-value math says size for the variance, not the median outcome.

Asymmetric-specific signal weighting

Within the existing 0-100 scoring, asymmetric setups warrant:

Smart-money cluster: insider buying clusters in small caps are

multiple times more informative than in mega caps (Lakonishok-Lee 2001 effect strongest sub-$2B). Weight as-is but be more skeptical of mega caps where insider activity is mostly RSU vest noise.

Anti-signal gates relaxed slightly: customer concentration > 30% and

some balance-sheet weakness are expected in small caps. Compliance doesn't auto-kill on these for risk_class=asymmetric, but they DO get surfaced in the bear case.

Liquidity gate relaxed: $500K average daily volume floor (vs $5M for

core). Anything below $500K is genuinely illiquid even at 1% sizing.

Catalyst is even more important: without a catalyst, small caps

drift down or sideways for years. A score-70 asymmetric without a clear catalyst is a skip.

Default structure: long_stock. Options structures (CSP, CC, strangle)

are wrong shape for asymmetric upside. Naked is reckless. LEAPS would be the correct leverage but are deferred to v2.

Every asymmetric idea is rendered with a prominent ASYMMETRIC badge on the site so the risk profile is unmissable.

4. Structure selection

Once an idea is publishable, pick the structure:

Long stock — when to choose

Asymmetric multi-quarter thesis with named catalyst path
IV is not particularly elevated (covered-call premium would be too thin)
Trader prefers full upside participation
Default for anything with score 80+

Cash-Secured Put (CSP) — when to choose

Fundamentally healthy name where you'd be happy to own at strike
IV is elevated (recent earnings, regulatory uncertainty, news catalyst pulled forward)
Want income while defining a "buy zone"
Structure parameters:

Strike at ~10–15% below current price
Expiry 45–90 DTE (delta-decay zone)
Annualized yield ≥ 15% to be worth the capital lockup
If-assigned cost basis is the actual risk metric — calculate it

Covered Call (CC) — when to choose

Already-owned position (or new long-stock entry simultaneously)
Range-bound thesis or anticipating a fade after a recent run-up
IV elevated
Strike at 10–20% above entry, 45–90 DTE
If-called return ≥ 15% annualized

Short Strangle — when to choose

Range-bound thesis with elevated IV (high implied volatility in the chain)
Comfortable owning at put_strike AND no parabolic-upside scenario through call_strike
Annualized yield ≥ 12% to be worth the dual-side premium
Structure parameters:

Both strikes ~10–20% OTM
Same expiry, 30–60 DTE (gamma decays fastest in this window)
Compute breakevens: call_strike + total_premium = upper, put_strike - total_premium = lower
"Total premium" is the income; max profit is the entire premium if stock between strikes at expiry

Risk: uncapped beyond breakevens. Treat as a defined-conviction premium-collection trade, not a yield-chasing default.

Naked Put — when to choose

Same setup as CSP but you're using margin instead of locking cash
Annualized yield ~2x the CSP equivalent, but margin calls if underlying drops sharply
Strike at ~10–15% below current price; 45–90 DTE
Recommend ONLY for accounts with sufficient margin headroom — note this in the trade

Naked Call — when to choose (rarely)

Bearish thesis with elevated IV
Theoretically unlimited downside if underlying rallies. Reserved for extreme cases:

high IV crush setup, clear bear catalyst, ability to defend (close or convert) on adverse move

Confidence ceiling: 4 (never publish a naked call at conf 5 — uncapped risk

doesn't deserve full sizing)

Strike at ~10–15% above current; 30–45 DTE
All publishes must include explicit upside-shock thesis + defense plan in conditions

Skipped in v1

Vertical spreads (debit/credit) — defer to v2 for defined-risk variants
Calendar spreads, diagonals — too many parameters to operationalize cleanly
Iron condors / iron butterflies — combinations of the above; defer

5. Risk management rules

Position sizing

The score determines the cap; the dossier-specific risk profile picks the actual size within the cap. Reduce position size if:

Customer concentration > 25%
Single regulatory or geopolitical risk dominates the thesis
Liquidity is at the lower end (avg volume < $20M)
Catalyst is binary (FDA, pivotal trial, single-customer contract)

Stop conditions

Every idea defines a stop signal, not just a price. Examples of valid stops:

"Close if Q3 GM guide is below 33% (signals ASP roll-over has begun)"
"Close if the rumored M&A is publicly denied by the target board"
"Close if 200DMA is breached on weekly close"

A simple "% stop" alone is not enough — the stop should reference the thesis.

Time stops

Every idea has a time horizon (3–12 months). At horizon expiry, close regardless of price unless the catalyst has clearly extended.

Return target

Long-stock ideas must have a source-backed path to at least 20% upside from the idea entry price. Lower-upside setups are research notes, not trade ideas. Paper-tracked entries can still be logged if the event path is measurable, but the upside/return target must be explicit so the monitor can judge it.

Concentration

Soft cap: no more than 3 open ideas in the same sub-sector at once (e.g., 3 HBM names, 3 optical names). Sector rotation correlates them.

6. The agent operationalization

Each agent uses this methodology differently:

Hunter

Scans for catalyst events that warrant investigation:

Insider buying clusters (high-priority — Form 4 filings filtered for code P)
Recent earnings beats with guidance raises
M&A rumors / SEC 8-Ks of strategic actions
Sector inflection news (capex guides from hyperscalers, supply-shortage narratives)
13F season filings showing concentrated initiations
Politician disclosures (STOCK Act feeds)

Scout

Investigates one symbol deeply. Builds the dossier with explicit grading per category:

Reads filings, transcripts, news
Pulls fundamentals, options, technicals, insider activity
Cross-references hedge fund holdings (if data available)
Articulates the edge in data — the specific thing the market is missing
Returns dossier with raw scoring inputs (not the score itself — that's Analyst's job)

Analyst

Consumes dossier
Computes the composite score against the rubric in section 3
Selects structure per section 4
Sizes per section 5
Either drafts a normal idea, drafts a zero-position paper_track idea for

measurable 45-59 setups, or returns {skip, reason}

Devil's Advocate

Attacks the scoring inputs:

Are the insider buys really open-market (code P) or just option exercises?
Is the catalyst date confirmed in filings, or just paraphrased from a news headline?
Is the mispricing driven by structurally lower margins (a value trap) rather than temporary?
Is the data current, or stale by 6+ months?

Either PASSES the scored draft or KILLS with specific issues
Can downgrade confidence/size

Compliance

Verifies scoring claims trace to sources
Sizes are within caps for the assigned confidence
Disclosure language is present
No personalized advice, no guarantees, no promotional tone

7. Self-improvement loop

Every closed idea is a learning opportunity. The flow:

monitor.js detects close conditions (target hit, stop, time, thesis broken) and moves the idea to content/closed/.
reviewer.js is auto-dispatched. It re-fetches current state, compares to the original thesis, scores which inputs were predictive, and extracts specific lessons.
Lessons append to content/lessons.json with category, pattern, evidence, applicability, and confidence.
Scout and Analyst load the lessons database before forming new theses. Lessons that contradict a likely thesis raise the bar; lessons that confirm one lower it.
The /lessons.html page on the site is the public auditable record of what the system has learned.

What makes a good lesson:

Specific — names sector, signal, condition (not "always do thorough research")
Actionable — changes how a future scoring decision would be made
Evidence-backed — cites the specific closed idea that produced it

The Reviewer is explicitly trained to reject generic platitudes. Lessons that don't change scoring behavior aren't real lessons.

The track record page shows the outcomes (price-based wins/losses). The lessons page shows the meta-outcomes (what the system learned from those outcomes). Together they form the closed feedback loop.

8. What this system is NOT

Not a backtest. The methodology is forward-looking. There is no historical out-of-sample validation built in (yet). Track-record page is the validation, accumulating in real time.
Not factor-investing. Factors compound to small premia; we want concentrated asymmetric trades.
Not high-frequency. Time horizon is 3–12 months. Daily noise is irrelevant.
Not infallible. Every signal in section 1 has been wrong many times. Default-silence + Devil's Advocate + Compliance gating are the safeguards.

9. Future extensions (v2+)

Add tools/insider.js parsing transaction codes from Form 4 XBRL (currently only metadata)
Add tools/hedge_fund.js for 13F deltas (whalewisdom or similar)
Add tools/politician.js for STOCK Act feeds (Capitol Trades / Quiver)
Add tools/sentiment.js running FinBERT or local-model scoring on news bodies
Add quantitative backtest harness: feed historical catalysts through the system, measure hit rate
Schema versioning + closed-idea audit trail for retroactive system tuning

This methodology is a living document. Update as the regime shifts and as new evidence accrues.