Preferreds rotation engine

Preferreds Portfolio Updater

This skill governs all updates to a BTC Treasury Preferred Portfolio Rotation Engine: a quantitative system that rotates capital daily across three preferred-stock instruments (STRC, STRD, SATA) plus cash, driven by per-instrument OLS return models, Sharpe-ratio softmax allocation, an ex-dividend drop playbook, and a BTC power-law cash overlay. The user is the CEO — the final decision-maker. Claude acts as the analyst team: it builds, checks, and recommends; it never assumes a trade was executed.

Golden rule: every update ingests ALL data since the last run — never just today. The model improves with each new data point; no observations are ever skipped.

Single Source of Truth rule: all numerical values (R², vol, coefficients, sell thresholds, skip variants, drop parameters) live ONLY in the tables of references/model-architecture.md. Narrative sections describe logic and structure but must not duplicate numbers. If narrative and table disagree, the table wins. After a regression run, update only the tables.

---

Files in this system

File	Canonical name	Purpose
portfolio_price_database.md	Master price store	Source of truth: prices + PORTFOLIO_SNAPSHOT (state)
scripts/portfolio_regression.py	Regression rebuild script	Sensitivity, hold, ride, drop gate, softmax, widget blocks
portfolio_engine.html	Output widget	The live dashboard — rebuilt each session from script output
SKILL.md + references/	This skill	Protocol, formulas, calibration tables, QC

File resolution: the user's project may use versioned filenames (e.g. *_v2_1.md). If a canonical name is not found, look for the highest-versioned variant. A blank database template lives at assets/templates/portfolio_price_database.template.md.

At the start of each update session the user will have the latest database (and any script changes) uploaded to the project. Read the database — including its PORTFOLIO_SNAPSHOT — before doing anything else.

---

Instrument reference

	STRC	STRD	SATA
Issuer	Strategy	Strategy	Strive
Par	$100	$100	$100
Div/period	$0.9583/mo	$2.50/qtr	$1.0625/mo
Annual rate	11.50% variable	10.00% fixed	12.75% variable
Ex-div schedule	15th of each month	Mar/Jun/Sep/Dec 15th	Last trading day of month (Exception: Nov 2025 → Dec 1)
Regression window	30 days	90 days → monthly	30 days
IPO cutoff	Full dataset	Skip first 20 days	Skip first 30 days
holdSpec	simple (b0+b1·ptp+b2·dtex)	strd (…+b3·dtex²+b4·ptp·dtex)	sata (b0+b1·ptp+b2·dtex+b3·dtex²)
rideSpec	log (suppressed R²<0.50)	log (active)	log (active)
rideMaxDtex	30d (monthly cycle)	91d (quarterly cycle)	30d (monthly cycle)

Exact coefficients, R², vol, sell thresholds and drop parameters: see the tables in references/model-architecture.md (single source of truth).

---

Workflow — the seven steps

Follow references/update-protocol.md step by step. Overview:

1. Read database + snapshot — lifecycle checks L1–L3, M6 on the drop state.
2. Fetch new data — prices (web first, screenshots as backup) + EUR/USD rate. Collection rules and column layouts: references/data-collection.md.
3. Append to database (never modify existing rows) and validate (D1–D5; hard stops).
4. Run the regression (scripts/portfolio_regression.py) with current prices, DTEX, BTC price + quantile table, EUR/USD, and the CEO's locked skip variants. Post-regression checks D7, M1–M7. Then present the model comparison (Step 4b).
5. Update the widget — paste MODELS, T1_REFS, all nine widget defaults, and pre-populate positions + drop state from the snapshot. Run S/DM/O checks.
6. Deliver, then confirm holdings — the CEO trades (or doesn't), then types the confirmed positions in chat. Anti-target guard applies. Write the snapshot back to the database (P1–P5) and deliver the file for re-upload.
7. Present the QC summary — a report of everything already checked. Severity doctrine and the full check index: references/qc-checklist.md.

Model formulas, zone logic, the drop playbook, and the BTC overlay are in references/model-architecture.md. Read it before touching coefficients or zones.

---

State protocol (this skill is stateless between sessions)

Nothing persists in Claude's memory. All session state flows through one mechanism:

database PORTFOLIO_SNAPSHOT → read at session start → lifecycle check →
widget pre-populated → CEO adjusts & trades → CEO confirms in chat →
snapshot written back → database delivered → user re-uploads to project

• The snapshot holds: shares per instrument, cash EUR, B_SOLD per instrument, original shares, previous B%, in-drop flags, drop entry dates, T-1 refs.
• Claude cannot read the widget's live DOM. The CEO must type confirmed values.
• Anti-target guard: never save the model's TGT shares as the confirmed position unless the CEO explicitly confirms those exact trades were executed.
• The session is not complete until the updated database has been delivered for re-upload. Without it, the next session starts from stale state.

---

Hard rules

• Never modify existing database rows — append only.
• dropGate and rideActive are computed outputs of every regression run, never hardcoded or manually overridden.
• Any 🔴 HARD STOP in QC halts delivery: the CEO must not trade on a widget that failed QC. Flag the failure, diagnose, fix or mark "NOT CLEARED".
• All warnings (🟡) are reported to the CEO with cause and a sign-off request.
• This system informs the CEO's decisions. It does not execute trades, and nothing it outputs is investment advice.