portfolio-review

name

portfolio-review

description

Whole-portfolio periodic review that produces ONE unified holdings decision table (Buy/Hold/Trim/Sell/Avoid per name) plus watchlist ranking, a keep-what-works list, and a tax/date-aware execution sequence. Use for "full portfolio review", "go through all my holdings", "what should I do across everything", "monthly review", "rebalance my whole book". Long-term-first; not a single-name desk pass (use trading-desk for one ticker) and not the daily digest (use daily-briefing).

Portfolio Review (whole-book decision table)

Goal

One coherent, execution-ready plan across EVERY holding and watchlist name - optimized for long-term after-tax wealth. The output is a single decision table, not nine separate analyses. Periodic cadence (monthly/quarterly), not daily. Bias toward FEW high-value actions: most holdings should end "Hold".

When to invoke

• User wants the whole book reviewed at once, not one ticker
• A monthly/quarterly check-in across holdings + watchlist
• NOT for a single-name buy decision (that's trading-desk) or a morning digest (that's daily-briefing)

Decision Memory Protocol (load first, log after)

Before forming any view, load prior decisions so verdicts stay consistent across sessions:

• mcp__aifolimizer__get_cross_ticker_lessons (max_lessons=3) - portfolio-level win/loss patterns
• For any name you issue a per-ticker BUY/SELL/TRIM/HOLD/ADD on, also load mcp__aifolimizer__get_ticker_decision_history (ticker=…, max_decisions=5) and mcp__aifolimizer__get_ticker_reflection (symbol=…, n=3). If a prior decision exists and this run flips it, state explicitly WHY (new data / catalyst / price); never silently contradict a logged decision.

After output, log every actionable verdict: for each BUY/SELL/TRIM/ADD/HOLD issued, call mcp__aifolimizer__log_recommendation (skill="portfolio-review", ticker, action, conviction, rationale, target_pct, stop_pct). Skipping breaks the cross-session feedback loop and causes drift.

How to run (call in parallel where independent)

Stage 0 - context:

• get_profile (capital, accounts - never hardcode)
• get_personal_context (tax bracket, FHSA/TFSA/RRSP room, account waterfall)

Stage 1 - portfolio truth:

• get_portfolio, get_concentration_warnings, get_xray

Stage 2 - per-name signal (top holdings + watchlist):

• get_fundamentals (holdings + watchlist union)
• get_positioning_signals (crowding - gates ADD decisions)
• get_tax_loss_candidates (underwater names = harvest candidates)

Stage 3 - date awareness:

• get_earnings_calendar (symbols = holdings + watchlist) - next 14d flags
• get_dividend_calendar (symbols = holdings) - ex-div / pay timing for don't-sell-before-ex-div and tax-year placement

Stage 4 - watchlist:

• get_watchlist → fold into the fundamentals/crowding pulls above

Decision logic (per holding)

Assign ONE action:

• Sell: thesis broken OR severe concentration with weak fundamentals OR better use of capital + tax-efficient to exit.
• Trim: single-name > 10% or sector > 35% (get_concentration_warnings), OR consensus-crowded (crowding ≥ 70) and extended.
• Hold: thesis intact, sizing fine - the DEFAULT. Most names land here.
• Add (small): fundamentals strong AND not crowded (crowding < 70, ideally contrarian ≤ 30) AND room under concentration caps AND cash available.
• Conviction HIGH/MED/LOW drives any sizing.

Keep-What-Works rule (anti-overtrading - REQUIRED section)

Before any Trim/Sell, list names that should be LEFT UNCHANGED even if they look "boring", because the cost of acting outweighs the benefit:

• Large embedded gain in a taxable (Non-Reg) account → selling realizes 50%- inclusion cap gains now; holding defers it.
• Safe, growing dividend (low payout, positive FCF) → income compounding.
• Core index ETF (the user's chosen broad-market fund) doing its job → don't tinker with the base.
• Within a few days of ex-dividend → don't sell before ex-div without reason. Overtrading is a known long-term return killer. The bar to touch a working position is HIGH. State why each kept name is kept.

Output structure

1. Executive summary (3-4 sentences)

What matters most this review. The 1-3 actions that actually move the needle.

2. Holdings decision table

One row per holding:

TICKER · weight% · ACTION · conviction · horizon · tax note · risk note · reason

Sort: actionable (Sell/Trim/Add) first, Holds grouped after.

3. Keep-what-works

The leave-alone list with per-name justification (see rule above).

4. Watchlist ranking

Each name → 🔥 High-conviction buy / 👀 Monitor / ❌ Avoid (hype/overvalued/weak), with a one-line reason and ideal-entry note. Crowding-aware: a popular name with crowding ≥ 70 and no fundamental edge is ❌, not 🔥.

5. Tax + account plan

• Harvest candidates (get_tax_loss_candidates) and the Non-Reg cap-gains note
• Account placement: where new contributions go (FHSA→TFSA→RRSP→Non-Reg per get_personal_context waterfall)

6. Date plan

• Earnings within 14d (hold-through vs trim-before)
• Ex-dividend / pay dates worth timing around (get_dividend_calendar)

7. Execution sequence

Numbered, practical: do-first / wait-on / avoid. Tie each step to a table row.

Rules

• Most holdings should be Hold. If the table is mostly Trim/Sell, re-check - that's an overtrading smell, not a plan.
• Crowding gates ADD (per CLAUDE.md): crowding ≥ 70 → never default to Add.
• Every number traces to a tool call - no fabricated prices, yields, or dates.
• Long-term-first framing (user's stated goal). Don't surface intraday calls here.
• Tax note on every Sell/Trim: Non-Reg realizes 50%-inclusion cap gains; TFSA/RRSP do not.
• If get_dividend_calendar or earnings data is empty for a name (sparse for TSX), say so - don't invent dates.
• One-shot. For a single name needing the full gated workup, hand off to trading-desk.

Gotchas

• get_dividend_calendar filters out past ex-dates by design - a held dividend payer with no row just has no upcoming ex-div in the window; not an error.
• Watchlist fundamentals share the get_fundamentals cache - cheap to include.
• Concentration is whole-book; a name split across TFSA + Non-Reg still counts once.