Fewshot Attempt 02

Asteria Investment Office — Institutional Portfolio-Risk Solver Skill

Reusable execution knowledge for the Asteria Investment Office remote environment. Test tasks draw from three workflows: (A) energy / fixed-income trade strategy, (B) international correlation review, (C) cross-asset active allocation view updates (sometimes combined in one committee JSON). Re-derive every value from the live environment per task — do NOT memorize specific answers.

0. Data-precedence rule (applies to every task)

The remote environment at <remote-env-url> is the current book of record. Files in input/payloads/ are intake context and MAY BE STALE: they often carry an earlier snapshot date, stale holding quantities/exception boards, or stale prior preferences. When a local payload conflicts with the environment, prefer the environment unless the task prompt explicitly instructs otherwise.

Staleness signals to ignore in favor of the environment:

• stale_holding_snapshot, stale_exception_board, stale_local_note, candidate_shortlist_from_prior_week, snapshot_date keys.
• Any payload date earlier than the environment as_of_date (currently 2026-05-29).
• A stale note saying "USD overweight as a defensive offset" or "old worksheet highlighted…" — these are prior preferences, not current views; refresh from macro signals / index levels.

The as_of_date you put in every answer = the environment's as_of_date (2026-05-29), NOT the payload's request/memo date.

1. Environment endpoints & field shapes

Base URL: <remote-env-url>. Call with curl or python urllib.

• GET /api/catalog — available portfolio ids, policy ids, index ids, issuer ids, bond ids, opportunity sets. Good first call to confirm IDs.
• GET /api/policies — all policy thresholds (see §2).
• GET /api/portfolios — portfolio summaries.
• GET /api/portfolios/<id> — {portfolio_id, name, objective, strategy, base_currency, as_of_date, market_value_usd_m, constraints{policy_id, duration_band_years, max_hy_allocation_pct, ...}, holdings[{instrument_id, quantity_usd_m, asset_class, sleeve, notes}]}. The constraints.policy_id tells you which credit policy applies (and thus the target HY reduction).
• GET /api/instruments/bonds — supports filters ?candidate=true, ?rating_bucket=HY. Each bond: {instrument_id, issuer_id, issuer_name, candidate, energy_linked, coupon_pct, maturity, modified_duration_years, rating, rating_bucket (IG|HY), sector, subsector, spread_bps, yield_to_maturity_pct, recommended_theme_tags}.
• GET /api/issuers — {issuer_id, issuer_name, sector, subsector, rating_bucket, watchlist (bool), credit_outlook, research_tags}. Key the watchlist check by issuer_id.
• GET /api/market/energy — oil/gas/LNG/refining/renewables signals (theme context for §A; not directly required for the arithmetic).
• GET /api/indices — index metadata.
• GET /api/index-levels — a dict mapping each index_id to a list of {date, level} monthly observations for ALL indices. GET /api/index-levels/<id> returns one index. No correlations are precomputed — compute them yourself (§5).
• GET /api/allocation/opportunity-sets — 25 sets, each {opportunity_set, asset_class (Equities|Duration|Credit|Currency), sub_asset_class, display_order}. Use this to map an opportunity-set name to its asset_class.
• GET /api/allocation/prior-views — list of records for multiple quarters; each {opportunity_set, quarter, previous_quarter, view (UW|N|OW), conviction}. You must filter quarter == target_quarter; the view on that record is the prior view held entering the target quarter.
• GET /api/macro-signals — list of {opportunity_set, quarter, score, rationale_code, drivers}. Filter quarter == target_quarter.

Roster of list endpoints also accept simple equality filters matching field names, e.g. ?rating_bucket=HY, ?candidate=true, ?quarter=Q2_2026.

2. Policy conventions (GET /api/policies)

Top-level: policy_id: POLICY_SET_2026_05, as_of_date: 2026-05-29.

• credit_default (policy_id POL_CREDIT_DEFAULT): duration_band_years [3.0, 5.0], max_hy_allocation_pct 20.0, issuer_concentration_limit_pct 12.0, subsector_min_count_for_diversified 2, target_hy_reduction_pct 0.0.
• credit_risk_reduction (POL_CREDIT_RISK_REDUCTION): same as above but target_hy_reduction_pct 4.0. Used by risk-rebalance portfolios (rotation tasks). The per-portfolio constraints names which one applies.
• correlation (POL_CORRELATION_DEFAULT): correlation_high_threshold 0.8, correlation_low_threshold 0.2, review_window_start 2025-05-30, review_window_end 2026-04-30.
• allocation_mapping (POL_ALLOCATION_MAPPING):
  • view_score_thresholds: OW_min 0.35, UW_max −0.35, neutral [−0.35, 0.35].
  • conviction_thresholds: HIGH_abs_min 0.7, MEDIUM_abs_min 0.35, LOW_abs_below 0.35.
  • view_rank: {UW: −1, N: 0, OW: 1}.
• multi_asset uses allocation_mapping + correlation_default + credit_default.
• multi_asset_risk uses credit_risk_reduction + correlation_default; committee_escalation_threshold: "two_or_more_material_exceptions".

For credit tasks, read the portfolio's constraints to pick target_hy_reduction_pct (0 vs 4) and the duration band.

3. Precision & ordering conventions (always honor the template)

Always follow the per-field precision declared in input/payloads/answer_template.json. Verified defaults:

• trade_package/rotation notional_usd_m and quantity_usd_m: 1 decimal.
• post_trade_metrics: total_market_value_usd_m 2 dp; hy_allocation_pct 2 dp; weighted_modified_duration_years 2 dp; weighted_yield_to_maturity_pct 2 dp.
• correlation: 3 decimals.
• signal_score: 3 decimals.
• hy_reduction_pct_points: 2 dp; post_trade_watchlist_exposure_usd_m: 1 dp.

Ordering rules (verified):

• Energy BUY-only trade_package: sort ascending by instrument_id.
• Rotation .trades: SELL before BUY, then instrument_id ascending within each action.
• index_set, diversification_candidates, pair ids within a pair: ascending alphabetical.
• allocation_views / committee lists: follow the request payload's focus/opportunity_sets order (NOT alphabetical) unless the template says otherwise — read the template's ordering field.
• rationale_codes (overlay): business priority order, highest first.
• watchlist_sell_ids: ascending instrument_id.

4. Post-trade metric recipe (all credit workflows)

Given current holdings (env quantity_usd_m) + trades (BUY adds, SELL subtracts), build the post-trade holdings map, then:

1. total_market_value_usd_m = Σ post-trade quantities (round 2 dp). For a funded rotation (Σ sells = Σ buys) this equals the current portfolio MV.
2. hy_allocation_pct = (Σ q where bond rating_bucket=='HY') / total × 100.
3. weighted_modified_duration_years = Σ(q × modified_duration_years) / total.
4. weighted_yield_to_maturity_pct = Σ(q × yield_to_maturity_pct) / total.
5. hy_reduction_pct_points = pre_trade_hy_pct − post_trade_hy_pct, where pre_trade_hy_pct = (current HY MV) / (current total MV) × 100.
6. post_trade_watchlist_exposure_usd_m = Σ post-trade q where the bond's issuer (via issuer_id → /api/issuers) has watchlist == true.

Get rating_bucket, modified_duration_years, yield_to_maturity_pct from /api/instruments/bonds; get the watchlist flag from /api/issuers keyed by issuer_id. Round each metric per §3. This recipe reproduces the env-verified metric set exactly (market-value-weighted).

Constraint booleans

• hy_cap_pass: post_trade_hy_allocation_pct <= max_hy_allocation_pct (20.0).
• duration_band_pass: post_trade weighted_modified_duration_years within [band[0], band[1]] inclusive (default [3.0, 5.0]).
• target_hy_reduction_met: hy_reduction_pct_points >= target_hy_reduction_pct (0 for credit_default, 4.0 for credit_risk_reduction).
• watchlist_exposure_cleared: post_trade_watchlist_exposure_usd_m == 0 (and buys_avoid_watchlist == true).
• selected_issuer_diversification_pass: selected buys come from distinct issuers (and no issuer exceeds issuer_concentration_limit_pct).
• selected_subsector_diversification_pass: selected buys span ≥ subsector_min_count_for_diversified (2) distinct subsectors.
• watchlist_avoidance_pass: no bought bond's issuer is on the watchlist.

5. Workflow A — Energy / fixed-income trade strategy (BUY tickets)

Inputs: portfolio_id, ticket count, total notional (+ split rule, e.g. evenly), allowed actions (usually BUY), income period, preferred exposures (e.g. "LNG exporters", "gas demand", "non-watchlist carry"), client segment, watchlist-yield sensitivity.

SOP:

1. GET /api/portfolios/<id> → as_of_date, market_value_usd_m, holdings, constraints (policy_id, duration band, HY cap).
2. GET /api/instruments/bonds?candidate=true → candidate universe. Filter to energy_linked==true (or per preferred exposures). Build a lookup of bond → {rating_bucket, modified_duration_years, yield_to_maturity_pct, issuer_id, subsector}.
3. GET /api/issuers → mark watchlist issuers. Exclude every candidate whose issuer watchlist==true (watchlist_avoidance).
4. Select exactly the requested ticket count, summing to the total notional with the required split. Prefer bonds that: (a) improve carry (higher YTM) while keeping post-trade HY % ≤ cap and post-trade duration within the band; (b) diversify — ≥ 2 distinct subsectors among selected buys and distinct issuers; (c) match the desk's preferred exposures/themes / recommended_theme_tags.
5. Compute post-trade metrics (§4) on env holdings + selected buys.
6. Derive constraint booleans (§4).
7. sales_positioning.target_segment + theme: pick from the template's enums driven by desk context (e.g. multi_asset_income + lng_export_tailwind when LNG exporters are selected; use avoid_watchlist_yield_trap when a watchlisted high-carry bond was the near-miss that you correctly excluded).
8. data_precedence = current_environment_over_stale_payload whenever the payload snapshot differs from the env (the usual case). Only use no_conflict_found if values truly match.
9. Sort trade_package ascending by instrument_id; round notional_usd_m to 1 dp.

Rotation variant (sell existing pressure, fund IG buys):

• SELL in full every watchlisted holding (clear avoidable watchlist risk).
• SELL additional HY holdings as needed to (i) fund the IG buys and (ii) meet/exceed target_hy_reduction_pct. Size so Σ sells ≈ Σ buys (funded rotation → total MV preserved). Use env quantities, NOT the stale_exception_board quantities (they differ).
• BUY only IG, non-watchlist candidates whose durations, blended in, keep portfolio duration inside the band. Avoid any desk-shortlisted candidate flagged for issuer-status concern.
• watchlist_sell_ids = watchlisted holdings sold (ascending instrument_id); buys_avoid_watchlist = true.
• risk_note_code enum: watchlist_concentration when watchlist clearing dominated; hy_cap_pressure when capping HY was the lever; duration_preservation when duration ballast drove buys; carry_tradeoff; no_action.

6. Workflow B — International correlation review

Inputs: portfolio_id, as_of, review_window{level_start_date, level_end_date}, index_universe[], CIO concern codes.

SOP:

1. GET /api/index-levels. For each index in index_universe, filter observations to [level_start_date, level_end_date] inclusive and sort ascending by date.
2. Compute monthly simple returns per index: r[i] = level[i+1] / level[i] − 1. return_observations = #levels − 1 (e.g. 12 monthly levels → 11 returns).
3. For every unordered pair, compute Pearson correlation of the two return series (equal length, aligned by date). Round to 3 decimals.
4. extreme_pairs.highest_positive = max-corr pair; lowest = min-corr pair. Each pair_id sorted alphabetical; correlation to 3 dp.
5. index_set = index_universe sorted ascending alphabetical.
6. concentration:
   • high_threshold_breached = any pair corr ≥ correlation_high_threshold (0.8).
   • china_asia_dependence_flag = true if any pair ≥ 0.8 involves a China/Asia-related index (IDX_CHINA, IDX_AC_ASIA_PAC_EX_JP, IDX_EM, IDX_EM_EX_CHINA, IDX_INDIA).
   • primary_code = CHINA_ASIA_DEPENDENCE if that flag is true; else GLOBAL_DEVELOPED_OVERLAP if the high cluster is only among developed/global indices (IDX_WORLD, IDX_EAFE, IDX_ACWI_IMI); else NO_MATERIAL_CONCENTRATION. The memo's concern codes (ASIA_BETA_OVERLAP, CHINA_DEDICATED_SLEEVE) hint toward CHINA_ASIA_DEPENDENCE.
7. diversification_candidates: indices with the lowest (most negative / lowest-abs) correlation to the concentration cluster — typically IDX_EM_EX_CHINA and IDX_LATAM. Sort ascending alphabetical; restrict to the template's allowed_values.
8. sleeve_actions: trim the concentration-driving sleeve (e.g. China → trim, IDX_CHINA) and add a diversifier (e.g. Latin America → add, IDX_LATAM). Respect the template's length/ordering and target_index_id allowed values; action ∈ {trim, add, hold, hedge, monitor, rotate}.

7. Workflow C — Cross-asset active allocation view updates

For each requested opportunity_set (in the request's focus order):

1. asset_class = opportunity-sets[os]['asset_class'] (Equities / Duration / Credit / Currency). Note mappings: U.S. Treasuries → Duration; Corporate High Yield → Credit; EUR/USD/currency names → Currency.
2. prior_view = prior-views filtered to quarter == target_quarter AND opportunity_set == os → view. (Filter by quarter — the endpoint holds multiple quarters; the wrong record gives the next quarter's view.)
3. macro = macro-signals filtered to quarter == target_quarter AND opportunity_set == os → score, rationale_code.
4. signal_score = round(score, 3).
5. view: OW if score ≥ 0.35; UW if score ≤ −0.35; else N.
6. change: compare view_rank[new] vs view_rank[prior_view]: UP if new>prior, DOWN if <, UNCHANGED if ==. Rank {UW:−1, N:0, OW:1}. (A score moving 0.6→0.4 is still OW→OW = UNCHANGED.)
7. conviction: HIGH if |score| ≥ 0.7; MEDIUM if 0.35 ≤ |score| < 0.7; LOW if |score| < 0.35.
8. rationale_code = macro['rationale_code'] (the enum from the signal record — GROWTH_IMPROVES, RATE_CUT_SUPPORT, CREDIT_SPREAD_RISK, DOLLAR_DEFENSIVE, CHINA_DEPENDENCE, LATAM_DIVERSIFIER, INDIA_OFFSET, DURATION_SUPPORT, HY_VALUATION_RISK, EUROPE_RECOVERY, JAPAN_POLICY_RISK, NEUTRAL_BALANCE).
9. Order rows per the template (usually the request payload's focus order).
10. policy_id (top-level, when required) = POLICY_SET_2026_05. as_of_date = env as_of (2026-05-29); target_quarter/prior_quarter from the request (e.g. Q2_2026 / Q1_2026).

Risk overlay (standalone allocation memo)

• overlay_code + primary_action: pick the dominant theme from the derived views:
  • Duration OW + HY UW → DURATION_QUALITY_TILT / tilt_to_duration_quality.
  • Credit/HY stress dominant → CREDIT_RISK_REDUCTION / trim_credit_beta.
  • Equity growth tilt → EQUITY_BETA_EXTENSION / add_cyclical_equity_beta.
  • Currency defensive → CURRENCY_DEFENSIVE_HEDGE / add_currency_hedge.
  • No strong tilt → NO_OVERLAY / hold_policy_weights.
• rationale_codes: the rationale codes of the most consequential views, in business priority order (highest priority first).

Combined committee JSON (correlation + allocation, e.g. multi-asset sleeve)

• correlation_summary: 2 items — highest_concentration (max-corr pair) and best_diversifier (min-corr / most-negative pair). Each pair sorted alphabetical; correlation to 3 dp (Pearson of monthly simple returns from /api/index-levels over the window).
• target_sleeve_actions: per sleeve, action ∈ {trim, add, hold, hedge, monitor, rotate} (e.g. trim the concentration sleeve, add the diversifier, hedge a defensive currency).
• allocation_views: per opportunity_set include prior_view, signal_score, view, change, conviction, rationale_code (all derived per §7 steps 2–8). Use the prior from /api/allocation/prior-views (filtered to target quarter), not the stale_local_note.
• rebalance_trigger enum: correlation_cap_breach if any relevant pair ≥ 0.8; else hy_cap_pressure / duration_drift / watchlist_concentration / committee_review per the dominant risk.
• portfolio_risk_concentration_flag (bool): true when the correlation high-threshold (0.8) is breached.
• next_step enum: approve_rotation / defer_pending_risk_review / approve_with_monitoring (typical when concentration flagged but overlay addresses it) / reject_constraint_breach.

8. Common misjudgments & exclusion rules

1. Stale-payload values: don't use stale_holding_snapshot MV/HY%/ quantity or stale_exception_board quantities — they differ from the env (e.g. payload 58.5 MV vs env 60.0; payload 10.0 held vs env 12.0). Always recompute from /api/portfolios/<id>.
2. Watchlist yield trap: before selecting any candidate, check its issuer_id in /api/issuers and exclude if watchlist==true. A high-carry watchlisted bond is the classic distractor.
3. Duration-ineligible distractor: a long-dated or HY bond whose inclusion would push post-trade duration outside [3.0, 5.0] or breach the HY cap. Always compute post-trade metrics before finalizing.
4. Wrong prior view: /api/allocation/prior-views holds multiple quarters. Filter quarter == target_quarter. Taking the first/last record without filtering yields the next quarter's view and corrupts change.
5. Mis-deriving change: it is new view vs prior view by view_rank, NOT score direction or conviction change. Equal views ⇒ UNCHANGED.
6. Correlation on levels: must be Pearson of monthly simple returns (level[i+1]/level[i] − 1), not of price levels.
7. Window/observation count: filter levels to [start, end] inclusive; return_observations = #levels − 1 (12 levels → 11 returns).
8. Trade ordering: energy BUY-only ⇒ ascending instrument_id; rotation ⇒ SELL-before-BUY then instrument_id ascending. Swapping these fails the sort check.
9. Funded-rotation base: Σ sells = Σ buys so total MV is preserved; compute HY reduction against the env pre-trade HY MV / env total MV, not the stale board.
10. Stale desk preference override: ignore old "USD overweight as defensive offset" notes when refreshing — re-derive view/change from current macro-signals + April index levels.
11. Missing data_precedence: when a payload snapshot conflicts with env, set current_environment_over_stale_payload. It is almost always set (the payloads are deliberately stale).
12. Ignoring the answer template's exact field set / enums: re-read answer_template.json each task — required keys, allowed enum values, ordering instructions, and precision per field vary slightly across tasks. Return only the JSON object, no narrative.

9. End-to-end checklist (every task)

1. Read prompt.txt + every input/payloads/*.json (request + answer template). Identify the workflow(s) and the required top-level keys.
2. Confirm IDs via /api/catalog if unsure; then pull the per-task endpoints.
3. Use env as_of_date as the answer's as_of_date.
4. Compute from the environment (holdings/bonds/issuers/index-levels/prior- views/macro-signals/policies) — never trust stale payload snapshots.
5. Round per the template; sort per the template; use only allowed enum values.
6. Verify pass/fail booleans against §4 thresholds.
7. Return only the JSON object.