portfolio-decision-engine

name	portfolio-decision-engine
description	Policy-constrained Portfolio Manager (PM) decision workflow — compute investable/bucket weights, diagnose drifts and breaches, and emit three-layer recommendations with audit + chat rendering.
version	1.1.0
author	Hermes Agent
license	MIT
metadata	{"tags":["finance","portfolio","policy","decision-engine"],"sources":["docs/portfolio-policy.md","docs/pm-bot-decision-spec.md"]}

Overview

Use this skill whenever you need to transform Portfolio Truth + Policy into actionable, policy-consistent recommendations. It implements the four decision modes (Portfolio Review, Security Review, Rebalance, Exception/Override) and the three-layer output (portfolio/bucket/security) with audit trails and a concise chat renderer.

When to Use

• User asks “What should I do with the portfolio now?”, “Rebalance?”, or similar.
• A daily/weekly/monthly decision summary needs to run.
• A single security crosses a band, or the user requests an action potentially outside policy.

Prerequisites

• Portfolio Truth and Rules available (Truth, FX overrides, ringfence rules, MMF safety allocation).
• Policy thresholds defined (gold, cash, single-name caps, bands) — extend portfolio_rules.json if missing.
• Latest valuation snapshot (prices may be last-available, not RT).

Inputs

• Portfolio Truth (holdings, qty, cost, ccy, sleeves), valuation snapshot, FX.
• Policy:
  • bucket_overrides: explicit bucket mapping by isin/ticker/name
  • gold_policy: {floor_hard, floor_pref, upper}
  • cash_policy: {min_investable, pref_low, pref_high}
  • single_name_caps: {normal_cap, exceptional_cap, core_target_min, core_target_max}
• User intent (mode selection, overrides, horizon if provided).

Data Flow

1. Load portfolio truth + rules + FX → valuation snapshot.
2. Compute Safety/Obligation = ringfenced (PrivateFund/PrivateBiz) + min(120k USD, MMF). Investable = Total − Safety.
3. Classify positions into buckets (safety, cashflow, growth, hedge/insurance, tactical) via overrides → heuristics → fallback.
4. Compute bucket weights as % of Investable; compute gold% and investable cash%.
5. Diagnose portfolio stance and bucket statuses vs policy bands.
6. Generate three-layer recommendations per Decision Spec.
7. Render concise chat output; save full JSON for audit.

Modes

• Portfolio Review: overall diagnosis and 3–5 actions.
• Security Review: one security with portfolio-context view and bands.
• Rebalance: when bands/thresholds breached or new capital arrives.
• Exception/Override: user asks outside policy — must present normal rec, why outside, risk, safer alternative, and smallest-risk implementation.

Output (JSON)

• Layer 1 (Portfolio): overall_status, portfolio_stance[], primary_issue, next_best_use_of_capital, rebalance_required, override_warning_if_any.
• Layer 2 (Buckets): [{bucket_name, current_weight, target_band, status, recommended_action, rationale}].
• Layer 3 (Securities): top 3–5 items [{ticker, asset_type, current_weight, target_weight_or_range, action, priority, buy_band|trim_band zone, suggested_amount|quantity, rationale, key_risk, thesis_status, next_review_date}].

Rendering (chat message)

• DEFAULT: Detailed Chinese natural-language brief enumerating JSON fields in this fixed order:
  1. 决策时间; 2) 模式+政策版本; 3) 估值快照; 4) 总体状态; 5) 组合姿态; 6) 主要/次要问题; 7) 下一笔最优资本用途; 8) 是否触发再平衡; 9) 桶/类别诊断（当前%/目标带/状态/建议/理由）; 10) 具体标的与操作（操作、金额/数量、当前权重、目标范围/买卖区、理由、关键风险、论点、下次回顾）; 11) 方法说明; 12) 通用风险。
• Do NOT emit the old concise summary unless the user explicitly requests it. Treat this format as locked unless the user asks to change it.
• Keep separate from the 18:00 SGT valuation push (FORMAT LOCKED). Never alter the price-push format.

Auditability

• Save to ~/.hermes/portfolio/decisions/YYYYMMDD-HHMMSS_.json with: timestamp, mode, policy_version, snapshot_id, securities, payload, assumptions, override_flag, next_review_date.
• Also save a concise TXT next to the JSON for quick chat delivery.

Scheduling

• Daily 18:05 SGT: concise decision dashboard.
• Weekly (Mon 18:10 SGT): stance + drift + key actions.
• Monthly (1st trading day 18:15 SGT): allocation vs target + rebalance candidates.
• Use Hermes cronjob tool with deliver=origin so summaries arrive in-chat. Do NOT interfere with the 18:00 SGT valuation push (keep this pipeline separate).

Implementation Notes (Hermes tooling)

• read_file may prefix lines with "|". Always strip these before json.loads when using tool outputs.
• read_file can return cache messages like "Old tool output cleared…" or "File unchanged since last read". For heavy processing or repeated reads, prefer direct Python I/O (open(...)) to avoid tool cache artifacts.
• When writing decisions programmatically, persist both JSON and a compact TXT summary.
• Keep bucket/policy logic in a reusable script (e.g., ~/.hermes/portfolio/pm_engine.py) for cron reuse.

Default Bucket Overrides (examples)

• name::PrivateFund → safety; name::PrivateBiz → safety.
• MMF: allocate min(120k USD, MMF) to safety; remainder to tactical.
• name::Cash USD → tactical.
• ticker::IAU or names containing "Gold" → hedge.
• Demo Bond Fund/Demo SG Fund* fixed-income funds → cashflow.
• StructNote → growth (US) unless policy states otherwise.

Steps (Implementation)

1. Extend portfolio_rules.json with bucket_overrides, gold_policy, cash_policy, single_name_caps (include Demo Bond Fund/Demo SG Fund→cashflow; MMF remainder→tactical; safety obligations; IAU→hedge).
2. Compute Safety and Investable; derive bucket weights, gold% and investable cash%.
3. Diagnose portfolio-level stance and bucket statuses; set rebalance_required triggers (gold<3%, cash<5%, single-name>10%, etc.).
4. Select top 3–5 securities based on gaps/concentration/opportunity; generate actions (hold/add/trim/exit/watch) with zone labels (avoid false precision).
5. Enforce non-negotiables: don’t use ringfenced capital; respect caps; avoid blind averaging down; block adds if cash below minimum unless justified.
6. Emit JSON (3 layers) and render the concise chat message; persist decision with audit fields.
7. Schedule daily/weekly/monthly decision summaries via cronjob, delivering concise text + MEDIA attachments of JSON/TXT.

Verification

• Unit tests for bucket classification, Investable math, thresholds.
• Scenario checks: gold<3%, cash<5%, single-name>10%, growth>50%.
• Validate JSON schema fields present and values consistent.
• Manual review that the chat output stays within ~25 lines with 3–5 actions.

Pitfalls

• Mixing push-format report with decisions — keep them separate; never change the LOCKED 18:00 SGT push format.
• False precision in buy/trim bands — prefer zones (staged accumulation/trim-on-strength).
• Hermes tool cache artifacts (lineno prefixes, cache messages) can corrupt JSON loads — strip prefixes or use direct file I/O.
• Ignoring policy conflict when a security looks attractive — policy takes precedence; use Override Mode when user insists.

References

• docs/portfolio-policy.md
• docs/pm-bot-decision-spec.md