Fewshot Attempt 03

Asteria Investment Office — Operational Skill

Environment

All tasks use the shared Asteria API at http://34.46.77.124:8010. Treat the API as the current book of record. Local payload files (in input/payloads/) are intake context — they may contain stale marks, outdated worksheets, or pre-refresh preferences. When the API and a local payload disagree, the API wins.

The current environment as_of_date is 2026-05-29. Always set as_of_date to this value in outputs. Set data_precedence to "current_environment_over_stale_payload" whenever the local payload contains stale/older data that the API supersedes.

Key endpoints

Workflow: Always start by reading the API

1. Fetch /api/catalog to orient on available IDs.
2. Read the answer template (input/payloads/answer_template.json) for the required output shape, key names, enum values, ordering rules, and numeric precision.
3. Read the payload (desk request, review request, allocation request, risk memo, committee request) for the task's scope — which portfolio, which instruments/indices/opportunity-sets are in play, and any stale data warnings.
4. Fetch the portfolio (/api/portfolios/<id>) for current holdings and constraint policy.
5. Fetch the relevant policy (/api/policies) — the portfolio's constraint_policy_id tells you which thresholds apply.
6. Fetch supporting data (bonds, issuers, indices, index-levels, macro-signals, prior-views, opportunity-sets) relevant to the task.

Task archetypes and computation rules

A. Credit trade / rebalance (bond portfolios)

Selecting bonds:

• Prefer bonds where candidate: true.
• For energy-linked mandates, prefer energy_linked: true.
• Avoid issuers where watchlist: true (from /api/issuers). If the task requires avoiding watchlist buys, never propose a BUY for a watchlisted issuer's bonds.
• If the task requires selling watchlist positions, identify holdings whose issuer has watchlist: true.
• For income/carry objectives, prefer higher yield_to_maturity_pct while respecting constraint policy.

Post-trade metrics — key formulas:

total_market_value_usd_m (precision 1 or 2, follow template):

= sum of all holding market values after proposed trades
= current portfolio MV + net new buy notional (for buy-only)
= current portfolio MV (for rotation/rebalance — MV is preserved unless funded)

For a rotation (sell to fund buy), total MV stays at the pre-trade level. For a buy package funded from new sleeve allocation, total MV = pre-trade MV + total buy notional.

HY allocation pct (precision 2):

hy_allocation_pct = (sum of HY-bucket holding quantities) / total MV * 100

Where HY bucket = bonds whose rating_bucket is "HY" in the bond master.

weighted_modified_duration_years (precision 2):

weighted_duration = sum(quantity_i / total_MV * duration_i) for all holdings

weighted_yield_to_maturity_pct (precision 2):

weighted_YTM = sum(quantity_i / total_MV * YTM_i) for all holdings

hy_reduction_pct_points (precision 2):

hy_reduction = pre_trade_hy_pct - post_trade_hy_pct

post_trade_watchlist_exposure_usd_m (precision 1):

= sum of quantities of all holdings whose issuer has watchlist: true after trades

Constraint checks (booleans):

• hy_cap_pass: post_trade_hy_allocation_pct <= max_hy_allocation_pct (from policy)
• duration_band_pass: duration_band_years[0] <= post_trade_duration <= duration_band_years[1]
• selected_issuer_diversification_pass: no single issuer among the BUY selections exceeds issuer_concentration_limit_pct of total MV (check if any issuer appears on more than one selected ticket)
• selected_subsector_diversification_pass: at least subsector_min_count_for_diversified distinct subsectors among BUY selections
• watchlist_avoidance_pass: no BUY ticket is for a watchlisted issuer
• target_hy_reduction_met: hy_reduction_pct_points >= target_hy_reduction_pct from policy
• watchlist_exposure_cleared: post_trade_watchlist_exposure == 0

Trade ordering in output:

• Template specifies the rule. Common patterns: SELL before BUY, then alphabetical by instrument_id within each action group. Or: sort ascending by instrument_id.

B. Equity correlation review

Computing Pearson correlation from index levels:

1. Filter index levels to the review window (from level_start_date through level_end_date). Both endpoints are included — the first date is the base for the first return.
2. Compute monthly simple returns for each index:

   r_t = (level_t / level_{t-1}) - 1
   

   where t and t-1 are consecutive monthly observation dates.
3. The number of return observations = number of level dates minus 1.
4. Compute Pearson correlation between each pair of return series. Round to 3 decimal places.
5. Include all requested indices in index_set, sorted alphabetically.

Extreme pairs:

• highest_positive: the pair with the largest positive correlation (closest to +1.0). If ties, pick alphabetically.
• lowest: the pair with the most negative correlation (closest to -1.0). This is the lowest numeric value, not the lowest absolute value.
• Each pair's index IDs must be sorted alphabetically within the pair.

Concentration / dependence check:

• Compare the highest correlation against correlation_high_threshold (default 0.8 from POL_CORRELATION_DEFAULT).
• china_asia_dependence_flag: true if the China-related pair exceeds the high threshold AND involves IDX_CHINA + an Asia/Emerging index.
• high_threshold_breached: true if any pair exceeds correlation_high_threshold.
• primary_code: "CHINA_ASIA_DEPENDENCE" if China concentration is the primary concern, "GLOBAL_DEVELOPED_OVERLAP" for developed-market overlaps, "NO_MATERIAL_CONCENTRATION" if no threshold breached.

Diversification candidates:

• Indices that appear in the lowest-correlation pair (the best diversifier) OR that have pairwise correlations below correlation_low_threshold (default 0.2). Sorted alphabetically.

Sleeve actions:

• For concentrated sleeves: "trim".
• For diversification candidates: "add".
• Sorted alphabetically by sleeve name (or by the order in the template).

C. Allocation view refresh

Determining active view (UW / N / OW):

From /api/policies → allocation_mapping.view_score_thresholds:

• score >= OW_min (0.35) → "OW"
• score <= UW_max (-0.35) → "UW"
• Otherwise → "N"

Determining conviction (LOW / MEDIUM / HIGH):

From /api/policies → allocation_mapping.conviction_thresholds:

• |score| >= HIGH_abs_min (0.7) → "HIGH"
• |score| >= MEDIUM_abs_min (0.35) → "MEDIUM"
• Otherwise → "LOW"

Determining change (UP / DOWN / UNCHANGED):

Compare current view to prior-quarter view (from /api/allocation/prior-views):

• Map views to ranks: UW = -1, N = 0, OW = +1 (from view_rank in policy).
• If current rank > prior rank → "UP"
• If current rank < prior rank → "DOWN"
• If equal → "UNCHANGED"

Rationale code:

• Use the rationale_code from /api/macro-signals for the given opportunity_set and target quarter.
• The signal score column is also the signal_score field (precision 3).

Prior view:

• From /api/allocation/prior-views: look up the view for the opportunity_set where previous_quarter matches the task's prior quarter.
• Important: The prior-views endpoint returns entries where quarter is the target quarter and previous_quarter is the prior quarter. Match on both opportunity_set and previous_quarter.

Risk overlay:

• Select the overlay whose primary_action best matches the dominant tilt across views:
  • "DURATION_QUALITY_TILT" / "tilt_to_duration_quality" — when OW duration/IG, UW HY
  • "CREDIT_RISK_REDUCTION" / "trim_credit_beta" — when reducing credit risk
  • "EQUITY_BETA_EXTENSION" / "add_cyclical_equity_beta" — when adding equity exposure
  • "CURRENCY_DEFENSIVE_HEDGE" / "add_currency_hedge" — when hedging currency
  • "NO_OVERLAY" / "hold_policy_weights" — when no significant tilt
• rationale_codes: List the rationale codes of the views that drive the overlay decision, in business priority order (strongest/clearest signal first).

Ordering:

• allocation_views list: follow the order of focus_opportunity_sets from the request payload (not alphabetical).
• rationale_codes list: business priority order, highest priority first.

D. Multi-asset committee (hybrid: correlation + allocation)

This archetype combines correlation review and allocation views into a single JSON.

Correlation summary:

• Compute Pearson correlations on the subset of indices named in the request.
• highest_concentration: the pair with the highest positive correlation.
• best_diversifier: the pair with the lowest (most negative) correlation.
• Pairs sorted alphabetically within each pair list.

Target sleeve actions:

• Based on correlation findings and allocation views: "trim" concentrated sleeves, "add" diversifying sleeves, "hedge" currency sleeves with negative signals.

Allocation views for each opportunity set:

• Include prior_view (from /api/allocation/prior-views), signal_score (from /api/macro-signals), view (derived from score), change (vs prior), conviction (from score magnitude), rationale_code (from macro-signals).
• Round signal_score to 3 decimal places.

Rebalance trigger:

• "correlation_cap_breach" — when a correlation pair exceeds the high threshold.
• "hy_cap_pressure" — when HY allocation is near/above cap.
• "duration_drift" — when duration is outside the band.
• "watchlist_concentration" — when watchlist exposure is material.
• "committee_review" — when escalated for general review.

portfolio_risk_concentration_flag: true when any material concentration or threshold breach is identified.

next_step: "approve_with_monitoring" when there are actions but no hard constraint breach; "approve_rotation" for clean rotations; "defer_pending_risk_review" when material risks need further review; "reject_constraint_breach" for hard breaches.

Precision rules (by field type)

Critical: Match the precision declared in the template's properties for each field. If the template says precision: 2 for a field and the answer for train_001 shows 5.8 not 5.80, the format strips trailing zeros — match the number of significant decimal places shown in the template specification but don't pad trailing zeros unless the template explicitly formats them.

Ordering rules

1. Trade lists: SELL before BUY, then alphabetical by instrument_id within each action group. If the template says "sort ascending by instrument_id" without action grouping, sort globally.
2. Index lists: Always alphabetical ascending by index ID (IDX_...).
3. Pair IDs: Both IDs within a pair sorted alphabetically.
4. Allocation views: Follow the order in the request payload's focus_opportunity_sets or the template's item_order.
5. Sleeve actions: Alphabetically by sleeve name, or by template's item_order.
6. Diversification candidates: Alphabetically by index ID.
7. Watchlist sell IDs: Alphabetically by instrument ID.
8. Rationale codes in risk_overlay: Business priority order, highest/c strongest signal first.

Enumeration conventions

• Actions: BUY, SELL, HOLD, NO_TRADE
• Sleeve actions: trim, add, hold, hedge, monitor, rotate
• Views: UW (underweight), N (neutral), OW (overweight)
• Change: UP, DOWN, UNCHANGED
• Conviction: LOW, MEDIUM, HIGH
• Rating buckets: IG (investment grade: AAA through BBB-), HY (high yield: BB+ and below)
• Data precedence: current_environment_over_stale_payload, local_payload_over_current_environment, no_conflict_found
• Sales target segments: insurance_general_account, pension_liability_matching, multi_asset_income, private_bank_income, endowment_opportunistic
• Themes: lng_export_tailwind, oil_oversupply_caution, midstream_stability, transition_bond_selectivity, avoid_watchlist_yield_trap
• Risk note codes: hy_cap_pressure, watchlist_concentration, duration_preservation, carry_tradeoff, no_action
• Rebalance triggers: correlation_cap_breach, hy_cap_pressure, duration_drift, watchlist_concentration, committee_review
• Next steps: approve_rotation, defer_pending_risk_review, approve_with_monitoring, reject_constraint_breach
• Correlation concentration codes: CHINA_ASIA_DEPENDENCE, GLOBAL_DEVELOPED_OVERLAP, NO_MATERIAL_CONCENTRATION
• Overlay codes: DURATION_QUALITY_TILT, CREDIT_RISK_REDUCTION, EQUITY_BETA_EXTENSION, CURRENCY_DEFENSIVE_HEDGE, NO_OVERLAY
• Primary actions: tilt_to_duration_quality, trim_credit_beta, add_cyclical_equity_beta, add_currency_hedge, hold_policy_weights

Common pitfalls

1. Stale payload data: The local payload may have stale marks, old quantities, or outdated worksheet snapshots. Always reconcile against the API's current portfolio and instrument records. The stale snapshot is intake context — not the source of truth.

2. Wrong policy: Each portfolio has a constraint_policy_id. Credit default (POL_CREDIT_DEFAULT) and credit risk reduction (POL_CREDIT_RISK_REDUCTION) have different target_hy_reduction_pct values (0.0 vs 4.0). Multi-asset policies (POL_MULTI_ASSET_DEFAULT, POL_MULTI_ASSET_RISK) stack sub-policies. Always check which policy applies.

3. Correlation window: The window is level_start_date through level_end_date inclusive at both ends. The number of return observations = (number of levels) - 1. A 12-date window yields 11 return observations. Use only the dates within the window — do not include earlier or later observations.

4. Simple returns, not log returns: The formula is (L_t / L_{t-1}) - 1, not ln(L_t / L_{t-1}).

5. HY classification: HY is determined by rating_bucket, not by the spread or yield. A BB bond is HY even if its spread is modest.

6. Weighted metrics are market-value-weighted: Each holding's contribution to portfolio duration or YTM is proportional to its quantity divided by total portfolio market value.

7. Change direction: Compare current view rank minus prior view rank. If prior was N (0) and current is OW (+1), that's UP. If prior was OW (+1) and current is N (0), that's DOWN.

8. Signal score vs conviction: Signal score determines the view AND conviction independently. View uses the signal score threshold (±0.35). Conviction uses the absolute signal score magnitude (≥0.7 = HIGH, ≥0.35 = MEDIUM, else LOW).

9. Post-trade MV for rotation: In a rotation (sell to fund buy), total MV does NOT change — you are swapping one set of holdings for another. In a new-money buy, total MV = current MV + new buy notional.

10. Watchlist avoidance is per issuer: The /api/issuers endpoint has the watchlist boolean. If any issuer is watchlisted, all its bonds are off-limits for BUY actions when watchlist avoidance is required.

11. Output template is authoritative: Always read the template JSON for exact key names, required fields, enum values, and ordering rules. Templates differ between tasks even within the same archetype.

12. Trailing zeros in JSON numbers: Python's json.dumps may output 5.8 for a float 5.80. If the template requires exactly 2 decimal places (precision: 2), you may need to format the number as 5.8 (minimal representation) or 5.80 (padded). Follow the convention shown in the template's answer examples — train outputs show minimal decimal representation (e.g., 5.8 not 5.80).