Reflect 3 Attempt 03

Skill: Asteria Investment Office Portfolio Risk JSON Tasks

Overview

This skill covers solving tasks against the Asteria Investment Office shared environment, which exposes a REST API at a remote base URL. Tasks fall into several categories: energy-credit trade recommendations, issuer watchlist/correlation reviews, allocation view refreshes, fixed-income risk rebalances, and multi-asset committee packs.

Environment Setup

1. Base URL: Use only the provided environment base URL (e.g., http://34.46.77.124:8010). Do not use localhost or 127.0.0.1.
2. Key endpoints (discoverable from GET / and GET /api/catalog):
   • /api/catalog — lists available portfolio, policy, index, issuer, bond, and opportunity-set IDs.
   • /api/portfolios/<portfolio_id> — current holdings, constraints, market value.
   • /api/policies — policy thresholds (HY caps, duration bands, correlation thresholds, allocation mapping).
   • /api/instruments/bonds — full bond universe with candidate, energy_linked, rating_bucket, modified_duration_years, spread_bps, yield_to_maturity_pct, recommended_theme_tags.
   • /api/issuers — issuer records with watchlist, rating_bucket, sector, subsector, research_tags.
   • /api/index-levels/<index_id> — monthly index levels for correlation calculations.
   • /api/market/energy — energy market signals and pitch themes.
   • /api/allocation/opportunity-sets, /api/allocation/prior-views, /api/macro-signals — allocation taxonomy, prior-quarter views, and current macro signal scores/rationale codes.

General Workflow

1. Read the task prompt to identify the portfolio ID, task type, and required output schema.
2. Read the local payload (desk_request.json, review_request.json, allocation_request.json, committee_request.json, risk_meeting_memo.json, etc.) for context, but treat it as potentially stale.
3. Fetch current environment data for the portfolio, bonds, issuers, policies, and any relevant market signals.
4. Compute derived values (correlations, post-trade metrics, allocation views) using the current environment as the book of record.
5. Produce JSON matching the answer_template.json exactly, including key ordering, precision, and enum values.

Task Type Patterns

1. Energy-Credit Trade Recommendations (e.g., PF-EN-ALTA)

• Goal: Select exactly N BUY tickets (usually 2) totaling a specific notional (e.g., USD 8.0M split evenly).
• Constraints:
  • max_hy_allocation_pct from portfolio policy (e.g., 20%).
  • duration_band_years (e.g., 3.0–5.0 years).
  • Avoid watchlisted issuers.
  • Prefer energy_linked=true bonds aligned with current energy signals.
  • Ensure issuer and subsector diversification.
• Post-trade metrics to compute:
  • total_market_value_usd_m = current MV + new notional.
  • hy_allocation_pct = sum of HY holdings / total MV × 100.
  • weighted_modified_duration_years = weighted average duration.
  • weighted_yield_to_maturity_pct = weighted average YTM.
• Sales positioning: Match the target_segment and theme enums from the template to the portfolio strategy (e.g., multi_asset_income + lng_export_tailwind for LNG-focused trades).
• Data precedence: Always set current_environment_over_stale_payload when the local request contains stale marks.

2. International Equity Correlation Reviews (e.g., PF-INT-NEXVEN)

• Goal: Compute Pearson correlations from monthly simple returns over the requested window.
• Method:
  • Fetch /api/index-levels/<index_id> for each index in the universe.
  • Compute simple returns: (level_t - level_{t-1}) / level_{t-1}.
  • Compute Pearson correlation between each pair.
  • Round to 3 decimals.
• Outputs:
  • review_window with level_start_date, level_end_date, return_observations (N-1 months).
  • index_set sorted ascending alphabetically.
  • extreme_pairs: highest_positive and lowest pair, with pair_id sorted alphabetically.
  • concentration: Check if china_asia_dependence_flag is true (correlation between China and Asia Pac ex-JP ≥ threshold), set high_threshold_breached if any pair ≥ 0.8.
  • diversification_candidates: From allowed values, select those with low correlation to the rest of the portfolio.
  • sleeve_actions: Two actions sorted ascending by sleeve name, chosen from allowed actions/target indices.

3. Allocation View Refresh (e.g., Q2_2026 CIO Desk)

• Goal: Produce active views (UW/N/OW) for a focused set of opportunity sets.
• Inputs:
  • /api/macro-signals for the target quarter — provides score and rationale_code.
  • /api/allocation/prior-views for prior and current quarter — provides prior view and conviction.
  • /api/policies → allocation_mapping — defines score thresholds:
    • OW if score ≥ 0.35
    • UW if score ≤ -0.35
    • N otherwise
    • HIGH conviction if |score| ≥ 0.7, MEDIUM if ≥ 0.35, else LOW.
• Change determination: Compare current computed view to prior quarter view:
  • UP if view improves (e.g., N→OW or UW→N)
  • DOWN if view worsens (e.g., OW→N or N→UW)
  • UNCHANGED if same.
• Risk overlay: Choose from allowed overlay codes. Rationale codes should be ordered by business priority (highest first). Common choice: CREDIT_RISK_REDUCTION with trim_credit_beta when HY signals are negative.
• Lineage: Include as_of_date (environment date), target_quarter, prior_quarter, and policy_id (e.g., POL_ALLOCATION_MAPPING).

4. Fixed-Income Risk Rebalance (e.g., PF-FI-LUMEN)

• Goal: Reduce HY exposure and remove watchlist risk while keeping duration in band.
• Inputs:
  • Current portfolio holdings from /api/portfolios/<id>.
  • Bond details from /api/instruments/bonds.
  • Issuer watchlist status from /api/issuers.
  • Policy constraints (e.g., POL_CREDIT_RISK_REDUCTION with target_hy_reduction_pct, duration_band_years, max_hy_allocation_pct).
• Approach:
  • Sell HY bonds from the stale exception board, especially watchlisted ones.
  • Buy IG candidate bonds that are not watchlisted and fit the duration band.
  • Ensure hy_reduction_pct_points meets the minimum target.
  • Keep post_trade_duration_years within the CIO band.
• Outputs:
  • rotation.trades: SELLs first sorted by instrument_id, then BUYs sorted by instrument_id.
  • risk_metrics: post_trade_hy_allocation_pct, post_trade_duration_years, hy_reduction_pct_points, post_trade_watchlist_exposure_usd_m.
  • exception_flags: All booleans reflecting constraint passes.
  • watchlist_handling: List of sold watchlist IDs, and buys_avoid_watchlist=true.
  • risk_note_code: Choose from allowed enum (e.g., watchlist_concentration, hy_cap_pressure, duration_preservation).

5. Multi-Asset Committee Pack (e.g., PF-MA-HELIO)

• Goal: Link non-US equity correlation findings to active allocation views.
• Correlation summary: Exactly 2 items ordered [highest_concentration, best_diversifier]. Pairs sorted alphabetically. Correlation rounded to 3 decimals.
• Target sleeve actions: Ordered by opportunity set (Emerging Markets, India, Latin America, USD). Action from allowed enum.
• Allocation views: Ordered by opportunity set. Include prior_view from /api/allocation/prior-views, signal_score from /api/macro-signals, computed view/change/conviction/rationale_code.
• Rebalance trigger: Choose from allowed enum based on findings (e.g., correlation_cap_breach if high correlation exceeds threshold).
• Portfolio risk concentration flag: Boolean based on whether concentration risk is material.
• Next step: Choose from allowed enum (e.g., approve_with_monitoring when there is concentration but no hard constraint breach).

Critical Precision Rules

• Always round to the precision declared in the answer template.
• Sort lists as specified in the template (alphabetical, ascending by instrument_id, SELL before BUY then by instrument_id, etc.).
• Use only allowed enum values; any other string will fail validation.
• Include all required keys; missing keys will fail validation.

Common Pitfalls

• Stale data: The local payload may contain stale marks or outdated holdings. Always use the current environment API as the book of record.
• Portfolio not found: Some portfolio IDs in prompts may not exist in the catalog. Use the portfolio ID explicitly named in the prompt/template, but fetch its data from the environment.
• Correlation window: Use the exact level dates from the request payload. Return observations = number of levels - 1.
• Duration/YTM weighting: Weight by market value (quantity_usd_m), not equal weighting.
• HY allocation: Compute as percentage of post-trade total market value, not pre-trade.