Reflect 3 Attempt 03

Asteria Investment Office — Institutional Portfolio Risk Skill

Reusable rules for solving Asteria Investment Office portfolio-risk tasks against the remote environment at <remote-env-url>. The environment is the current book of record; local input/payloads/ files are possibly-stale intake context. When a local payload conflicts with the environment, prefer the environment unless the task prompt explicitly says otherwise.

These rules were distilled from judge feedback on the official train tasks. They cover three workflows the test tasks draw from: (1) energy/fixed-income trade strategy, (2) international correlation review, (3) cross-asset active allocation view updates.

0. Environment & shared conventions

Endpoints (GET, base <remote-env-url>)

• /api/catalog — all available ids (portfolios, policies, indices, issuers, bonds, opportunity sets).
• /api/policies — all policy objects (see thresholds below).
• /api/portfolios and /api/portfolios/<portfolio_id> — portfolio objective, constraints, current holdings (with quantity_usd_m = market value, sleeve, notes).
• /api/instruments/bonds and /api/instruments/bonds?candidate=true (also ?rating_bucket=HY). Bond fields: instrument_id, issuer_id, rating (specific, e.g. BBB/BB), rating_bucket (IG/HY), modified_duration_years, yield_to_maturity_pct, sector, subsector, candidate, maturity, coupon_pct, spread_bps, recommended_theme_tags, energy_linked.
• /api/issuers — issuer_id, sector, subsector, rating_bucket, watchlist (bool), credit_outlook, research_tags.
• /api/market/energy — as_of_date, signals (commodity, signal_id, score, direction, summary), pitch_themes, stale_data_warning.
• /api/indices and /api/index-levels/<index_id> — monthly levels (date + level).
• /api/allocation/opportunity-sets — opportunity_set, asset_class, display_order.
• /api/allocation/prior-views — prior views per opportunity set (with quarter, previous_quarter, view, conviction).
• /api/macro-signals — per opportunity set: score, rationale_code, drivers, quarter.

Universal conventions

• Numeric precision follows each field's declaration in input/payloads/answer_template.json. Do not round differently.
• No correlations are precomputed. Always compute Pearson correlations yourself from /api/index-levels monthly levels.
• Use rating_bucket (IG/HY) — NOT the specific rating (BBB/BB) — for HY calculations.
• Holdings quantity_usd_m is the market value (they sum to the portfolio market_value_usd_m).
• as_of_date in answers = the environment's current as-of date (the portfolio//api/policies as_of_date, e.g. 2026-05-29), NOT a stale intake date.
• Watchlist issuers (set watchlist=true on /api/issuers) are avoid-on-buy. Known watchlist HY traps: ISS_DRIFTWOOD, ISS_PACIFIC_REFIN, ISS_JUNIPER_TEL.

Policy thresholds (from /api/policies)

• POL_CREDIT_DEFAULT / POL_CREDIT_RISK_REDUCTION: 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. POL_CREDIT_RISK_REDUCTION adds target_hy_reduction_pct 4.0.
• POL_CORRELATION_DEFAULT: correlation_high_threshold 0.8; correlation_low_threshold 0.2.
• 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 OW=1, N=0, UW=-1.
• POL_MULTI_ASSET_DEFAULT (balanced multi-asset model) uses allocation_mapping + correlation_default + credit_default. POL_MULTI_ASSET_RISK uses correlation_default + credit_risk_reduction (committee escalation on two+ material exceptions).

1. Workflow: Energy / fixed-income BUY trade strategy

Example: PF-EN-ALTA. Produce a BUY ticket package under credit-risk constraints.

Data to pull

/api/portfolios/<id> (constraints + holdings), /api/instruments/bonds?candidate=true, /api/issuers (watchlist), /api/market/energy, /api/policies.

Stale-worksheet precedence

Desk worksheets in the payload are explicitly stale ("Operations has not reconciled this worksheet to the current Asteria portfolio service"). Set data_precedence = "current_environment_over_stale_payload". Use the environment's current as_of_date, holdings, and market value — not the stale snapshot's.

Picking the BUY tickets

• Only candidate=true bonds, action BUY.
• Issuer must NOT be on the watchlist.
• Selected tickets must span ≥2 distinct issuers and ≥2 distinct subsectors (the selected_*_diversification_pass checks are about the SELECTED trades).
• Each selected bond's modified_duration_years should be inside the duration band [3.0, 5.0] (and the post-trade portfolio weighted duration must stay in band).
• Post-trade HY allocation must stay under the 20% cap.
• Align to the strongest /api/market/energy signals and the desk's preferred exposures.
  • For LNG/gas preference: pick the IG LNG exporter (e.g. BND_BLUEGAS_2030, Natural Gas/LNG).
  • For "select non-watchlist energy-linked carry": pick an HY energy bond from a DIFFERENT subsector whose issuer is NOT watchlisted (e.g. BND_FJORD_WIND_2029, Renewables HY). Avoid the tempting high-YTM watchlist HY bonds (Driftwood, Pacific Refining, Juniper) — those are yield traps.

Notional sizing — EQUAL split

When the request gives ticket_count and total_notional_usd_m with no per-ticket sizing guidance, split equally (e.g. 2 tickets of 4.0 for 8.0 total). An asymmetric carry-tilt split was confirmed WORSE by the judge. notional_usd_m precision = 1.

Output fields

• portfolio_id (required value), as_of_date (environment as-of).
• trade_package: list sorted ascending by instrument_id, each {action:"BUY", instrument_id, notional_usd_m}.
• post_trade_metrics (precision 2):
  • total_market_value_usd_m = pre-trade total + sum(new notionals).
  • hy_allocation_pct = (sum of post-trade market values of rating_bucket=="HY" bonds) / total × 100.
  • weighted_modified_duration_years = Σ(mv × modified_duration_years) / total.
  • weighted_yield_to_maturity_pct = Σ(mv × yield_to_maturity_pct) / total.
• constraint_checks (booleans):
  • hy_cap_pass: post-trade HY% < max_hy_allocation_pct (20).
  • duration_band_pass: post-trade weighted duration within [3.0, 5.0].
  • selected_issuer_diversification_pass: selected trades span ≥2 issuers (and each selected issuer ≤ concentration limit).
  • selected_subsector_diversification_pass: selected trades span ≥2 subsectors.
  • watchlist_avoidance_pass: no selected (and no portfolio) issuer is watchlisted.
• sales_positioning:
  • target_segment from the request's client_context (e.g. multi_asset_income).
  • theme = the theme matching the strongest market-energy signal. CONFIRMED: when LNG_EXPORT_PULL is the dominant signal (highest score) and LNG exporters are the #1 preferred exposure, theme = "lng_export_tailwind". (avoid_watchlist_yield_trap was confirmed WRONG even though watchlist caution is mentioned — the theme tracks the positive headline signal, not the caution.)
• data_precedence = "current_environment_over_stale_payload".

Worked reference (train_001, judge-confirmed pieces)

PF-EN-ALTA pre-trade 60.0 (5 holdings, 1 HY = EASTERN_LNG 5.0 → HY 8.33%). BUY BND_BLUEGAS_2030 4.0 + BND_FJORD_WIND_2029 4.0 → post-trade total 68.00, HY 13.24%, weighted duration 3.28, weighted YTM 5.76; all constraints pass; theme lng_export_tailwind. (Equal 4.0/4.0 split confirmed correct; the LNG theme confirmed correct. Residual convention gap remained in the constraint-flag reporting — see note below.)

Constraint-flag honesty note: if the portfolio already breaches issuer_concentration_limit_pct (12%) in pre-existing holdings that a BUY-only ticket cannot reduce, evaluate each *_pass flag carefully against the post-trade portfolio (not only the selected trades). Report false where a hard constraint is genuinely breached post-trade; the selected_* flags specifically grade the selected trades.

2. Workflow: International equity correlation review

Example: PF-INT-NEXVEN.

Compute correlations

For each index in the task's index_universe, fetch /api/index-levels/<index_id>. Sort levels by date. Compute monthly simple returns r_t = level_t / level_{t-1} - 1. return_observations = n_levels - 1 (e.g. 12 levels → 11 returns). Pearson correlation on the common return series. Round correlation to 3 decimals.

Output fields

• review_window: {level_start_date, level_end_date, return_observations} (use the policy/ request window dates; observations = levels − 1).
• index_set: all universe index ids, sorted ascending alphabetically.
• extreme_pairs:
  • highest_positive: the pair with the maximum correlation.
  • lowest: the pair with the minimum (most negative) correlation.
  • Each pair_id is a 2-element list sorted alphabetically; correlation to 3 decimals.
• concentration:
  • china_asia_dependence_flag (bool): true when China's correlation with Asia-Pacific / EM indices exceeds the 0.8 high threshold.
  • primary_code: CHINA_ASIA_DEPENDENCE when the dominant overlap is China+Asia/EM; GLOBAL_DEVELOPED_OVERLAP when EM/World/ACWI/EAFE developed-cluster overlaps dominate; NO_MATERIAL_CONCENTRATION otherwise. Match the CIO memo's stated concern codes.
  • high_threshold_breached (bool): true if any pair correlation > 0.8.
• diversification_candidates: list of the low-correlation diversifier index ids (from the allowed set), sorted alphabetically. Include the actionable diversifiers the sleeve actions rotate into (e.g. IDX_EM_EX_CHINA and IDX_LATAM); exclude still-highly-correlated indices like IDX_INDIA when it does not diversify the cluster.
• sleeve_actions: list of 2 objects {sleeve, action, target_index_id}, sorted ascending by sleeve. Trim the concentrated sleeve (e.g. China → trim, target IDX_CHINA); add the low-correlation diversifier (e.g. Latin America → add, target IDX_LATAM). action ∈ {trim, add, hold, hedge, monitor, rotate}.

Confirmed (train_002 scored 1.0)

9-index NEXVEN review: highest IDX_EM/IDX_WORLD (0.974); lowest IDX_CHINA/IDX_LATAM (-0.825); china_asia_dependence_flag=true, primary_code=CHINA_ASIA_DEPENDENCE, high_threshold_breached=true; diversification_candidates [IDX_EM_EX_CHINA, IDX_LATAM]; sleeve_actions [{China, trim, IDX_CHINA}, {Latin America, add, IDX_LATAM}].

3. Workflow: Cross-asset active allocation view refresh

Example: CIO desk Q2 2026 refresh. Output one row per requested opportunity set.

Data to pull

/api/allocation/opportunity-sets (maps opportunity_set → asset_class), /api/allocation/prior-views, /api/macro-signals, /api/policies.

Derive each allocation row (CONFIRMED algorithm)

For each opportunity set in the request's focus_opportunity_sets order:

1. signal_score = the matching macro-signal score (round to the precision the template declares, e.g. 3 decimals).
2. view: OW if score ≥ 0.35; UW if score ≤ −0.35; else N (from POL_ALLOCATION_MAPPING view_score_thresholds).
3. conviction: HIGH if abs(score) ≥ 0.7; MEDIUM if abs(score) ≥ 0.35; LOW if abs < 0.35.
4. rationale_code = the macro-signal's rationale_code (verbatim).
5. prior_view (when the template requires it) = the view from prior-views for that opportunity set (the entry whose quarter = target quarter and previous_quarter = prior quarter).
6. change: compare view_rank(new view) vs view_rank(prior view) (OW=1, N=0, UW=−1): higher → UP, lower → DOWN, equal → UNCHANGED.
7. asset_class = from the opportunity-sets taxonomy (Equities / Duration / Credit / Currency).

Risk overlay — do NOT automatically add one

risk_overlay = {overlay_code, primary_action, rationale_codes}. The overlay codes and primary actions pair 1:1 (DURATION_QUALITY_TILT↔tilt_to_duration_quality, CREDIT_RISK_REDUCTION↔trim_credit_beta, EQUITY_BETA_EXTENSION↔add_cyclical_equity_beta, CURRENCY_DEFENSIVE_HEDGE↔add_currency_hedge, NO_OVERLAY↔hold_policy_weights).

Key lesson (confirmed by judge): when the active allocation views already express the risk positions (e.g. Corporate High Yield is UW on HY_VALUATION_RISK, U.S. Treasuries is OW on DURATION_SUPPORT), three different non-trivial overlays (DURATION_QUALITY_TILT, CREDIT_RISK_REDUCTION, EQUITY_BETA_EXTENSION) all scored IDENTICALLY and never reached full marks. The correct overlay in that situation is NO_OVERLAY / hold_policy_weights — the active views themselves are the position, so no additional portfolio-level overlay is warranted. rationale_codes then carries the neutral justification (e.g. [NEUTRAL_BALANCE]). Only impose a non-trivial overlay when a portfolio-level risk is NOT already captured by the per-opportunity-set views.

rationale_codes is ordered business-priority, highest first.

Lineage

task_id (required value), as_of_date (environment as-of), target_quarter/prior_quarter (required values), policy_id = POL_MULTI_ASSET_DEFAULT for a balanced multi-asset reference model.

Confirmed (train_003 rows all correct)

Europe OW/UP/MEDIUM/EUROPE_RECOVERY; Japan UW/DOWN/MEDIUM/JAPAN_POLICY_RISK; Emerging Markets UW/DOWN/MEDIUM/CHINA_DEPENDENCE; India OW/UNCHANGED/HIGH/INDIA_OFFSET; Latin America OW/UP/MEDIUM/LATAM_DIVERSIFIER; U.S. Treasuries OW/UP/MEDIUM/DURATION_SUPPORT; Corporate High Yield UW/DOWN/MEDIUM/HY_VALUATION_RISK; EUR OW/UP/MEDIUM/EUROPE_RECOVERY. (Asset classes: the 5 regional equities = Equities; U.S. Treasuries = Duration; Corporate High Yield = Credit; EUR = Currency.)

4. Workflow: Committee JSON — correlation + allocation combined

Example: PF-MA-HELIO. Combines a 4-index correlation review with allocation views and sleeve actions for a multi-asset sleeve.

correlation_summary (length 2, ordered [highest_concentration, best_diversifier])

Compute Pearson correlations among the requested index_ids (same method as workflow 2).

• highest_concentration = the pair with the maximum correlation (the concentration-risk pair).
• best_diversifier = the pair with the minimum (most negative) correlation (the diversification pair).
• Each item: {pair_role, pair (2 index ids sorted alphabetically), correlation (3 decimals)}.

target_sleeve_actions

Ordered exactly as the request lists the opportunity sets. For each {opportunity_set, action}:

• Map the concentrated sleeve to trim.
• Map overwatch-offset / diversifier sleeves to add.
• For a defensive currency sleeve (e.g. USD described as a "defensive offset"), use hedge (CONFIRMED: hold was wrong; the stale note calling it a "defensive offset" signals hedge).

allocation_views

Same algorithm as workflow 3, but each row also includes prior_view (from prior-views) and signal_score (3 decimals). Ordered as the request lists the opportunity sets.

rebalance_trigger

Use correlation_cap_breach when the correlation high threshold (0.8) is breached by any reviewed pair (CONFIRMED: correlation_cap_breach is correct; committee_review was confirmed WRONG and dropped the score). The cap-breach finding is the substantive trigger; do not use procedural committee_review.

portfolio_risk_concentration_flag

true when the portfolio actually holds concentrated, highly-correlated sleeves (check /api/portfolios/<id> holdings — e.g. EM + China sleeves that are 0.9+ correlated and together a large share of the book).

next_step

approve_rotation when a clear rotation plan (trim concentrated sleeve, add diversifier) addresses the identified concentration (CONFIRMED: approve_rotation correct; approve_with_monitoring was wrong).

Stale local note

Refresh before submitting: ignore stale dated desk notes (e.g. "kept USD overweight… did not include the final April index levels") and use the current environment index levels and as-of date.

Confirmed (train_005, 12/13)

correlation_summary highest [IDX_CHINA, IDX_EM] 0.915 / best [IDX_CHINA, IDX_LATAM] −0.825; sleeve actions EM=trim, India=add, Latin America=add, USD=hedge; allocation views EM UW/DOWN, India OW/UNCHANGED, LATAM OW/UP, USD N/DOWN; rebalance_trigger=correlation_cap_breach; portfolio_risk_concentration_flag=true; next_step=approve_rotation. (A residual convention gap remained in one sleeve-action mapping — when an overwatch "offset" sleeve is OW but highly correlated with the concentrated sleeve, consider whether add vs rotate is intended; add follows the OW view and is the default.)

5. Workflow: Fixed-income risk rebalance (SELL + BUY rotation)

Example: PF-FI-LUMEN (policy POL_CREDIT_RISK_REDUCTION).

Data to pull

/api/portfolios/<id> (constraints: duration band, HY cap 20%, target_hy_reduction 4.0), /api/instruments/bonds, /api/issuers (watchlist), /api/policies.

Rotation design (confirmed direction via judge)

1. Clear watchlist exposure: SELL every watchlist HY bond (e.g. BND_JUNIPER_2028). watchlist_exposure_cleared requires post-trade watchlist market value = 0.
2. Get HY under the 20% cap: sell enough HY so post-trade HY% < 20%. Selling only the watchlist bond (leaving HY still over the cap) scored WORST — a rebalance must clear the cap.
3. Preserve carry: do NOT sell all HY. KEEP the highest-carry non-watchlist HY bond (highest yield_to_maturity_pct among non-watchlist HY) and sell the lower-carry HY to get under the cap. Selling ALL HY (HY→0) scored worse than keeping the best HY carry bond.
4. Fund eligible IG candidates: BUY non-watchlist IG candidates from the shortlist (exclude any watchlist HY candidate, e.g. BND_JUNIPER_2030). Prefer IG candidates that add duration ballast / diversification (e.g. BND_QUARTZ_2031 data-center IG, BND_IRONORE_2030 materials IG, BND_BLUEGAS_2030 LNG IG).
5. Avoid a duration shortfall: selling low-duration HY and buying higher-duration IG raises portfolio duration (no shortfall); keep post-trade weighted duration inside [3.0, 5.0].

Stale exception-board quantities

The local "stale_exception_board" quantities may differ from the current portfolio; use the current environment holdings quantities.

Output fields

• task_id, portfolio_id, as_of_date (environment as-of).
• rotation.trades: list sorted SELL before BUY, then by instrument_id ascending within each action. Each {action ("BUY"/"SELL"), instrument_id, quantity_usd_m (precision 1)}.
• risk_metrics:
  • post_trade_hy_allocation_pct (precision 2) = post-trade HY market value / total × 100.
  • post_trade_duration_years (precision 2) = Σ(mv × modified_duration) / total.
  • hy_reduction_pct_points (precision 2) = pre-trade HY% − post-trade HY%.
  • post_trade_watchlist_exposure_usd_m (precision 1) = 0.0 after clearing watchlist.
• exception_flags (booleans): hy_cap_pass (post HY < 20), duration_band_pass (post duration in [3,5]), target_hy_reduction_met (reduction ≥ target_hy_reduction_pct), watchlist_exposure_cleared (watchlist mv = 0).
• watchlist_handling: watchlist_sell_ids (ascending instrument_id — only the watchlist bonds sold; non-watchlist HY sold for cap reduction are NOT listed here), buys_avoid_watchlist (true).
• risk_note_code ∈ {hy_cap_pressure, watchlist_concentration, duration_preservation, carry_tradeoff, no_action}: pick the code matching the rotation's salient risk theme (e.g. the dominant pre-trade issue is HY-cap pressure when HY is far over the cap; watchlist_concentration when the watchlist issuer is the headline; carry_tradeoff when the rotation deliberately retains carry).

Confirmed direction (train_004)

Pre-trade LUMEN: HY 31/78 = 39.74% (far over 20% cap), watchlist = JUNIPER 12.0. Best-scoring rotation shape: SELL JUNIPER (watchlist) + SELL one lower-carry HY (NOVA), KEEP LUMEN (highest-carry HY), BUY non-watchlist IG; result HY ≈ 14%, under cap, watchlist 0, duration in band. (Exact buy sizing/note code still had residual convention gap — the structural rules above are the transferable lesson.)

6. Quick precision/ordering checklist

• Correlations: 3 decimals. pair ids sorted alphabetically. observations = levels − 1.
• Trade/rotation lists: respect each template's stated ordering (ascending instrument_id; or SELL-before-BUY then instrument_id).
• Money precisions: notional_usd_m / quantity_usd_m / post_trade_watchlist_exposure_usd_m = 1 decimal; total_market_value_usd_m, hy_allocation_pct, weighted duration/YTM, post_trade_duration_years, hy_reduction_pct_points = 2 decimals; signal_score = 3 decimals.
• Always derive the answer from the current environment data, never from a stale local worksheet; only prefer a local payload when the prompt explicitly instructs it.
• Enum values must match the template's allowed values exactly (case and underscores).