Asteria Investment Office

Asteria Investment Office — CIO / Credit Desk Skill

This skill solves tasks that return a single strict JSON object describing an investment decision for an Asteria portfolio. The shared environment is the authoritative book of record, exposed as a read-only HTTP/JSON API. Each task ships an answer_template.json (in input/payloads/) that defines the exact output schema: required keys, controlled enums, list ordering, and numeric precision. Always conform to that template; it overrides any default below.

0. Universal workflow (do this every time)

1. Read the prompt and every file in input/payloads/. The payload is intake context only — it may carry stale marks, holdings, watchlist flags, or preferences. Note the portfolio_id, the requested universe/quarter/window, and the exact output contract in answer_template.json.
2. Pull live data from the API (base http://127.0.0.1:8036, all GET): start with /api/policies (thresholds), /api/catalog (valid ids), and the portfolio/data endpoints relevant to the task type (see §1).
3. Apply source precedence: when the payload disagrees with the environment on marks, ratings, holdings, watchlist status, quantities, or policy data, use the environment. The payload's *_as_of_date/snapshot_date is usually older than the environment as_of_date (currently 2026-05-29).
4. Use the environment as_of_date for every as_of_date output field (read it from /api/policies or any portfolio record — they agree). Do NOT echo the payload's request/memo date.
5. Compute deterministically using the formulas in §3, then map to enums (§4). Round each field to its template precision (2 dp → e.g. 5.80 is 5.8 numerically; emit the number, JSON drops trailing zeros).
6. Honor ordering rules in the template (sort lists exactly as specified).
7. Return only the JSON object — no prose, no markdown fences.

1. API endpoints and what to call them for

Most list endpoints accept simple equality filters on field names (?rating_bucket=HY, ?candidate=true, ?quarter=Q3_2026, ?issuer_id=...).

2. Key policy constants (read live; values as of POLICY_SET_2026_05)

From /api/policies:

• credit_default / credit_risk_reduction: max_hy_allocation_pct = 20.0, duration_band_years = [3.0, 5.0], issuer_concentration_limit_pct = 12.0, subsector_min_count_for_diversified = 2. target_hy_reduction_pct = 0.0 for credit_default, 4.0 for credit_risk_reduction (the risk-reduction policy used by rebalance portfolios such as PF-FI-LUMEN).
• correlation: correlation_high_threshold = 0.8, correlation_low_threshold = 0.2, window 2025-05-30..2026-04-30.
• allocation_mapping:
  • view_score_thresholds: OW_min = 0.35, UW_max = -0.35, neutral in (-0.35, 0.35).
  • conviction_thresholds: HIGH_abs_min = 0.7, MEDIUM_abs_min = 0.35, LOW_abs_below = 0.35 (i.e. |score| ≥0.7 HIGH; ≥0.35 MEDIUM; else LOW).
  • view_rank: UW=-1, N=0, OW=1.
• multi_asset_risk: committee_escalation_threshold = two_or_more_material_exceptions.
• Top-level policy_id = POLICY_SET_2026_05 and as_of_date = 2026-05-29.

Always re-read these — never hardcode if the live values differ. A portfolio's own constraint_policy_id (from /api/portfolios) tells you which constraint block applies (e.g. PF-FI-LUMEN → POL_CREDIT_RISK_REDUCTION).

3. Calculation formulas and rounding

3.1 Portfolio metrics (credit tasks)

Let post-trade positions be {instrument_id: quantity_usd_m} (apply BUY +qty, SELL −qty to current holdings; drop zero/negatives).

• total_market_value_usd_m = Σ quantity. For a cash-neutral rotation (sells = buys) MV is unchanged; for a new-sleeve BUY package MV = current MV + Σ buys. (Precision 2.)
• hy_allocation_pct = 100 × (Σ qty where rating_bucket=="HY") / MV. (Prec 2.)
• weighted_modified_duration_years = Σ(qty × modified_duration_years) / MV. (Prec 2.)
• weighted_yield_to_maturity_pct = Σ(qty × yield_to_maturity_pct) / MV. (Prec 2.)
• hy_reduction_pct_points = pre-trade HY% − post-trade HY%. (Prec 2.)
• post_trade_watchlist_exposure_usd_m = Σ qty where the holding's issuer has watchlist == true. (Prec 1.)
• Issuer concentration (per issuer) = 100 × (Σ qty for that issuer) / MV; compare to issuer_concentration_limit_pct.

Rounding: round each reported field independently to its declared precision (standard half-up rounding). Trade notionals are in USD millions at precision 1.

3.2 Correlations (equity review / committee)

For each index, fetch monthly levels over the requested window (inclusive of level_start_date..level_end_date).

• Monthly simple returns: r_t = level_t / level_{t-1} − 1. N monthly levels ⇒ N−1 return observations (12 levels ⇒ 11). return_observations = N−1.
• Pearson correlation between two return series x, y: corr = Σ(x−x̄)(y−ȳ) / sqrt(Σ(x−x̄)² · Σ(y−ȳ)²). (Population vs sample std is irrelevant — it cancels.) Round to 3 decimals.
• highest_positive (a.k.a. highest_concentration) = the pair with the maximum correlation; lowest / best_diversifier = the pair with the minimum (most negative) correlation, over all pairs in the requested index set.
• Pair ids must be sorted alphabetically within each pair, and pair lists sorted as the template says.

4. Mapping decisions to enums (allocation engine)

This logic drives allocation views in tasks 003 and 005 and the sleeve actions.

For each requested opportunity_set at the target quarter:

1. signal_score = macro-signals score for that opportunity_set+quarter (report at precision 3 when the template asks for it).
2. view from score via view_score_thresholds: score ≥ OW_min → OW; score ≤ UW_max → UW; else N.
3. prior_view = prior-views record where quarter == target_quarter (its previous_quarter is the prior quarter; this record is the standing view going into the target quarter). Use its view (and its conviction if you need the prior conviction).
4. change = compare new view rank vs prior_view rank (UW=-1,N=0,OW=1): higher → UP, lower → DOWN, equal → UNCHANGED.
5. conviction from |score| via conviction_thresholds: ≥0.7 → HIGH, ≥0.35 → MEDIUM, else LOW.
6. rationale_code = the rationale_code field on the macro-signal record for that opportunity_set+quarter (do NOT invent one). Common values: EUROPE_RECOVERY, JAPAN_POLICY_RISK, CHINA_DEPENDENCE, INDIA_OFFSET, LATAM_DIVERSIFIER, DURATION_SUPPORT, HY_VALUATION_RISK, DOLLAR_DEFENSIVE, RATE_CUT_SUPPORT, CREDIT_SPREAD_RISK, GROWTH_IMPROVES, NEUTRAL_BALANCE.
7. asset_class (when required) = from /api/allocation/opportunity-sets (Equities / Duration / Credit / Currency).

This rule set reproduces every allocation row deterministically — trust it.

4.1 Risk overlay (allocation-refresh task)

Choose overlay_code + primary_action from the dominant cross-asset tilt in the computed views, and rationale_codes as the priority-ordered drivers:

• OW Duration (e.g. U.S. Treasuries OW) + UW Credit (HY UW) → DURATION_QUALITY_TILT / tilt_to_duration_quality. rationale_codes (highest priority first): the supportive duration driver (DURATION_SUPPORT), then the credit-risk driver (HY_VALUATION_RISK), then the next strongest negative-risk driver (e.g. CHINA_DEPENDENCE).
• Predominantly UW credit / spread risk → CREDIT_RISK_REDUCTION / trim_credit_beta.
• Predominantly OW cyclical equities, risk-on → EQUITY_BETA_EXTENSION / add_cyclical_equity_beta.
• Defensive currency tilt dominant → CURRENCY_DEFENSIVE_HEDGE / add_currency_hedge.
• No material tilt → NO_OVERLAY / hold_policy_weights. Order rationale_codes by business priority (strongest-conviction / largest |score| risk driver first).

4.2 Sleeve actions (correlation review & committee)

Map each sleeve/opportunity_set to an action enum (trim, add, hold, hedge, monitor, rotate):

• The sleeve driving the highest correlation / concentration (e.g. the China/EM core) and/or carrying a UW view → trim.
• A low/negatively-correlated diversifier with an OW view (e.g. Latin America, EM-ex-China) → add.
• A high-conviction OW offset (e.g. India) → add.
• A currency sleeve used as a defensive offset (e.g. USD) → hedge.
• Neutral/no-change → hold or monitor.

5. Per-task SOPs

Task A — Energy-credit trade package (e.g. PF-EN-ALTA)

Template keys: trade_package (2 items, BUY), post_trade_metrics, constraint_checks, sales_positioning, data_precedence.

1. Read the desk request (ticket_count, total notional, allowed actions). Split the total notional evenly across the tickets (e.g. 2 × USD 4.0m for an 8.0m package).
2. Build the eligible buy universe: bonds with candidate == true and (for energy mandates) energy_linked == true, whose issuer is NOT watchlisted (/api/issuers watchlist==false). This enforces watchlist_avoidance_pass.
3. Select two bonds that (a) are different issuers and different subsectors (satisfies selected_issuer_diversification_pass and selected_subsector_diversification_pass), (b) keep post-trade hy_allocation_pct ≤ max_hy_allocation_pct and post-trade weighted_modified_duration_years inside duration_band_years, and (c) improve carry while fitting the dominant energy theme and client pitch (do not blindly max carry — favor the theme with the highest energy signal score, typically LNG/gas, and avoid stretching HY to the cap). A robust, suitable package pairs an IG thematic anchor (e.g. LNG exporter) with one non-watchlist HY carry name in a different subsector.
4. Compute post_trade_metrics per §3.1 (MV = current MV + buys).
5. constraint_checks: each boolean is the pass/fail of its rule on the chosen package (HY cap, duration band, selected-issuer diversification, selected- subsector diversification, watchlist avoidance).
6. sales_positioning:
   • target_segment from the payload's client_context → multi_asset_income ("multi-asset income"), private_bank_income, insurance_general_account, pension_liability_matching, endowment_opportunistic.
   • theme from the dominant energy signal: high LNG signal → lng_export_tailwind; oil oversupply/negative → oil_oversupply_caution; midstream defensive → midstream_stability; renewables selective → transition_bond_selectivity; refining/watchlist caution → avoid_watchlist_yield_trap.
7. data_precedence: if the stale snapshot conflicts with the environment (MV, HY%, duration, holdings, ratings) → current_environment_over_stale_payload; if no conflict → no_conflict_found.
8. Order trade_package ascending by instrument_id.

Task B — International equity correlation review (e.g. PF-INT-NEXVEN)

Template keys: review_window, index_set, extreme_pairs, concentration, diversification_candidates, sleeve_actions.

1. Use the payload's review_window (level_start/end) and index_universe. Fetch levels for each index in the window; compute returns and return_observations = N−1.
2. Compute Pearson correlations for all index pairs (§3.2). Report extreme_pairs.highest_positive (max) and .lowest (min), pair ids sorted alphabetically, correlation to 3 dp.
3. index_set = the requested universe, sorted ascending alphabetically.
4. concentration:
   • high_threshold_breached = (max correlation ≥ correlation_high_threshold, i.e. ≥0.8).
   • china_asia_dependence_flag = true when the concentration is driven by China/EM/Asia overlap (the highest pairs are global-developed or EM/Asia/China names and the memo flags China/Asia dependence).
   • primary_code: CHINA_ASIA_DEPENDENCE when China/Asia/EM overlap drives it; GLOBAL_DEVELOPED_OVERLAP when developed-world indices dominate; NO_MATERIAL_CONCENTRATION when nothing breaches the high threshold.
5. diversification_candidates (from the allowed set, ascending): pick the indices that structurally reduce the identified concentration — for a China/Asia dependence, IDX_EM_EX_CHINA (de-China EM) and IDX_LATAM (lowest, often negative, correlation to China). Exclude an index that is itself part of the concentrated Asia/EM beta (e.g. IDX_INDIA is highly correlated to China, so it is usually NOT a China-dependence diversifier).
6. sleeve_actions (length 2, ordered by sleeve): trim the concentrated sleeve (e.g. China → IDX_CHINA) and add the best diversifier (e.g. Latin America → IDX_LATAM). target_index_id from the allowed set.

Task C — Active allocation view refresh (CIO desk)

Template keys: task_id, as_of_date, target_quarter, prior_quarter, policy_id, allocation_views, risk_overlay.

1. task_id = template required_value (e.g. train_003); target_quarter/ prior_quarter from payload; as_of_date = env date; policy_id = top-level policy_id from /api/policies (e.g. POLICY_SET_2026_05).
2. allocation_views: one row per opportunity_set in the payload's focus_opportunity_sets, in that exact order. For each row compute asset_class (opportunity-sets), view, change, conviction, rationale_code per §4.
3. risk_overlay per §4.1 (overlay_code, primary_action, priority-ordered rationale_codes).

Task D — Fixed-income risk rebalance (e.g. PF-FI-LUMEN)

Template keys: rotation.trades, risk_metrics, exception_flags, watchlist_handling, risk_note_code.

1. Goal: reduce HY and clear watchlist exposure while keeping duration in band; prefer a cash-neutral rotation (Σ SELL notional = Σ BUY notional, MV unchanged). Use the portfolio's POL_CREDIT_RISK_REDUCTION block — target_hy_reduction_pct = 4.0.
2. SELLs: sell every watchlisted holding in full (clears watchlist_exposure), then sell additional HY holdings as needed to hit the HY-reduction target while keeping post-trade duration in [3.0, 5.0]. Use current quantities from the API (override stale memo quantities).
3. BUYs: only candidate == true, non-watchlist, typically IG bonds from the memo shortlist / eligible universe (enforces buys_avoid_watchlist == true). Size buys so total buys = total sells and post-trade duration stays in band (longer-duration IG buys offset the duration lost by selling).
4. Compute risk_metrics per §3.1: post_trade_hy_allocation_pct, post_trade_duration_years, hy_reduction_pct_points, post_trade_watchlist_exposure_usd_m (target 0.0).
5. exception_flags: hy_cap_pass (post HY ≤ cap), duration_band_pass (post dur in band), target_hy_reduction_met (reduction ≥ target_hy_reduction_pct), watchlist_exposure_cleared (post watchlist exposure == 0).
6. watchlist_handling: watchlist_sell_ids = sold watchlist instruments (ascending), buys_avoid_watchlist = true if no buy is watchlisted.
7. risk_note_code: watchlist_concentration when the driving issue was watchlist exposure; hy_cap_pressure when HY cap was the binding concern; duration_preservation when duration was the key trade-off; carry_tradeoff when carry was sacrificed; no_action if no trade.
8. Order trades: all SELL before all BUY, then ascending instrument_id within each action.

Task E — Multi-asset risk-board reconciliation (e.g. PF-MA-CYGNUS)

Mixed credit + equity sleeves under POL_MULTI_ASSET_RISK. Identify and rank material exceptions: HY over/near cap, duration out of band, watchlist concentration, and equity correlation ≥ correlation_high_threshold. Escalate to committee when two or more material exceptions exist (committee_escalation_threshold). Follow the specific answer_template's enums and ordering; reuse §3 (credit metrics + correlations) and §4 (mappings).

Task F — Committee combined decision (e.g. PF-MA-HELIO / PF-MA-VEGA)

Template keys (HELIO example): correlation_summary, target_sleeve_actions, allocation_views, rebalance_trigger, portfolio_risk_concentration_flag, next_step.

1. correlation_summary (order: highest_concentration, then best_diversifier): from the requested equity index set compute max and min Pearson correlation (§3.2); pair ids sorted alphabetically, corr 3 dp.
2. target_sleeve_actions (fixed item order per template) per §4.2.
3. allocation_views (fixed item order per template): one row per opportunity_set with prior_view, signal_score (3 dp), view, change, conviction, rationale_code per §4.
4. rebalance_trigger: correlation_cap_breach when max equity correlation ≥ high threshold; else hy_cap_pressure / duration_drift / watchlist_concentration per the binding credit issue; committee_review if none binding but committee is reviewing.
5. portfolio_risk_concentration_flag = true when a high-threshold concentration (correlation or otherwise) is breached.
6. next_step: approve_with_monitoring when actionable but with a flagged risk; approve_rotation when clean; defer_pending_risk_review when exceptions need escalation; reject_constraint_breach when a hard constraint is violated.

6. Common pitfalls and exclusion rules

• Source precedence: always recompute from live API positions/marks; never trust stale payload quantities, watchlist flags, or HY% (e.g. a memo's stale_quantity_usd_m may differ from the current holding).
• as_of_date = environment date (2026-05-29), not the payload date.
• Watchlist comes from /api/issuers (watchlist bool) joined via issuer_id; never buy a watchlisted issuer; sell all watchlisted holdings in risk-reduction tasks.
• candidate == false bonds are not buyable — they are existing-holding or off-the-run instruments. Only buy candidate == true.
• HY cap is a ceiling (≤20%), not a target. Do not stretch HY to the cap for carry; suitable income packages stay comfortably inside it.
• Issuer & subsector diversification in trade packages is checked on the selected bonds (different issuers, different subsectors), not the whole portfolio (the existing book may already exceed the issuer-concentration limit on a legacy line, which is not your responsibility to fix in a 2-ticket add).
• Return observations = levels − 1. Off-by-one here breaks correlations.
• Pair ordering: index ids alphabetical within each pair; then order the pair list per the template.
• Allocation change is measured against the standing prior view for the target quarter (the prior-views record whose quarter == target_quarter), not against a different quarter.
• rationale_code and signal_score are read from macro-signals — do not derive a rationale from your own narrative.
• Rounding precision is per field (notionals 1 dp; %, duration, YTM, MV 2 dp; correlations & signal scores 3 dp). Round at the end, each field independently.
• Emit only the JSON object, with every required key present and only the allowed enum values.

7. Quick reference: enum cheat-sheet

• View: UW / N / OW from score thresholds (−0.35 / +0.35).
• Change: UP / DOWN / UNCHANGED (new view rank vs prior view rank).
• Conviction: LOW / MEDIUM / HIGH from |score| (0.35 / 0.7).
• Trade actions: BUY / SELL / HOLD / NO_TRADE (credit package); BUY / SELL (rotation).
• Sleeve actions: trim / add / hold / hedge / monitor / rotate.
• data_precedence: current_environment_over_stale_payload / local_payload_over_current_environment / no_conflict_found.
• Concentration primary_code: CHINA_ASIA_DEPENDENCE / GLOBAL_DEVELOPED_OVERLAP / NO_MATERIAL_CONCENTRATION.
• Overlay: DURATION_QUALITY_TILT / CREDIT_RISK_REDUCTION / EQUITY_BETA_EXTENSION / CURRENCY_DEFENSIVE_HEDGE / NO_OVERLAY paired with tilt_to_duration_quality / trim_credit_beta / add_cyclical_equity_beta / add_currency_hedge / hold_policy_weights.
• Committee next_step: approve_rotation / approve_with_monitoring / defer_pending_risk_review / reject_constraint_breach.