Fewshot Attempt 01

Asteria Investment Office — Portfolio-Risk Evaluation Skill

This skill distills reusable execution knowledge for solving Asteria Investment Office portfolio-risk tasks. The test task gives you a prompt, an input/payloads/answer_template.json, and one or more request payloads. You also have the remote Asteria environment (the current book of record) and this skill. Use the environment as the source of truth and follow the SOPs below.

The tasks draw on three workflows that may also be combined:

• Workflow A — Energy / fixed-income credit trade strategy. Build a BUY/SELL bond package under credit-risk constraints, then report post-trade metrics, constraint booleans, and sales positioning.
• Workflow B — International equity correlation review. Compute pair correlations from monthly index levels, flag concentration, and propose sleeve actions.
• Workflow C — Cross-asset active allocation view refresh. Translate macro signal scores into active views (UW/N/OW) with conviction, change vs prior quarter, rationale, plus risk overlay and committee fields.

0. Golden rules (read first)

1. Environment is the current book of record. The remote service at <remote-env-url> is authoritative for current holdings, holding quantities, issuer status, market signals, index levels, policies, prior views, and macro signals. Local input/payloads/*.json files are intake context that may be stale (old worksheet dates, stale marks, stale quantities, old shortlists). When a local payload conflicts with the environment, prefer the environment unless the task prompt explicitly overrides. Set data_precedence accordingly (current_environment_over_stale_payload when a conflict exists; no_conflict_found only when nothing stale is present). Watch for explicit "stale" markers in payloads (snapshot_date, stale_*, memo_as_of_date, stale_local_note, "Operations has not reconciled..." comments) — these signal you must refresh from the environment.
2. Always read answer_template.json first. It defines required keys, allowed enum values, field precision, ordering rules, and list lengths. Conform exactly: a field declared precision: 1 must be rounded to 1 decimal; a list declared length: 2 must have exactly 2 items; an enum must use one of the allowed values verbatim.
3. as_of_date is the environment's current date, NOT the payload's request date or memo date. Read it from GET /api/policies (top-level as_of_date) or any GET /api/portfolios/<id> (as_of_date). All train examples use the environment as_of_date (e.g. 2026-05-29), not the older memo/snapshot dates.
4. Correlations are never precomputed. Compute them yourself from GET /api/index-levels (monthly simple returns, see Workflow B).
5. Be precise about ordering. Lists are sorted as the template declares: trade packages by instrument_id ascending (or SELL-before-BUY then instrument_id), index sets/pair_ids alphabetically, allocation rows in the request payload's focus order, sleeve_actions by sleeve name ascending.

1. Environment endpoints (GET only)

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

2. Cross-cutting conventions

Precision (follow answer_template.json declarations)

• notional_usd_m / quantity_usd_m → 1 decimal.
• post_trade_metrics numeric fields (*_usd_m with precision 2, hy_allocation_pct, weighted_modified_duration_years, weighted_yield_to_maturity_pct) → 2 decimals.
• correlation → 3 decimals.
• signal_score → 3 decimals.
• total_market_value_usd_m → 2 decimals (per template).
• Booleans are true/false. Round half-up to the declared precision; JSON drops trailing zeros naturally (e.g. 5.80 renders as 5.8).

Policy snapshot (GET /api/policies)

Key fields the SOPs reference (concrete values observed; re-read at test time in case the environment changed):

• policy_id (top-level): the policy-set id used as the allocation policy_id field (e.g. POLICY_SET_2026_05).
• credit_default / 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. credit_risk_reduction adds target_hy_reduction_pct: 4.0 (the minimum HY reduction a rebalance must achieve). Portfolios opt into one of these via constraints.policy_id.
• correlation: correlation_high_threshold: 0.8, correlation_low_threshold: 0.2, review_window_start, review_window_end.
• allocation_mapping: view_score_thresholds (OW_min: 0.35, UW_max: -0.35, neutral_between: [-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).

Which policy a portfolio uses

Read GET /api/portfolios/<id> → constraints.policy_id. A pure credit sleeve uses POL_CREDIT_DEFAULT; a risk-reduction rotation uses POL_CREDIT_RISK_REDUCTION (which adds the target_hy_reduction_pct floor). Multi-asset sleeves compose: POL_MULTI_ASSET_DEFAULT references credit + correlation + allocation; POL_MULTI_ASSET_RISK references correlation + credit-risk-reduction.

3. Workflow A — Energy / fixed-income credit trade strategy

Applies to: standalone energy sleeve trades and fixed-income risk-reduction rotations. Two sub-shapes exist:

• Income-add shape (two BUY tickets): trade_package (BUY list), post_trade_metrics, constraint_checks, sales_positioning, data_precedence.
• Rotation shape (SELL+BUY): rotation.trades (SELL before BUY), risk_metrics, exception_flags, watchlist_handling, risk_note_code.

Step-by-step SOP

1. Read the portfolio GET /api/portfolios/<portfolio_id>. Capture current holdings (instrument_id, quantity_usd_m), market_value_usd_m, as_of_date, and constraints (duration band, HY cap, target HY reduction). Use these environment quantities, not any stale quantities in the local memo. (Example pattern: a local memo may show a holding at 10.0 while the environment shows 12.0 — trade off the 12.0.)
2. Read all bonds GET /api/instruments/bonds (and ?candidate=true for the buy universe). Read all issuers GET /api/issuers to get the watchlist flag per issuer (join on issuer_id).
3. Apply exclusion filters to candidate buys:
   • Watchlist avoidance: drop any bond whose issuer watchlist == true. This is a hard exclusion (the desk is "sensitive to headline carry from watchlisted issuers"). Candidates flagged with theme tags like WATCHLIST_RISK / REFINANCING_RISK or refining issuers (negative refining signal) are exclude-by-default.
   • Duration-band eligibility: keep only bonds whose modified_duration_years is within the policy band [3.0, 5.0] (inclusive). Long-dated distractors (dur 5.8, 6.7, 5.9) and very short ones (dur 2.3) are excluded. For an income-add where the package must keep the portfolio weighted duration in band, individual buy durations should sit inside the band.
   • HY capacity: HY buys are allowed only if post-trade HY % stays under the cap (20%). Do not blind-buy the highest-YTM HY name if it breaches the cap or concentrates a watchlisted issuer.
   • Sector/energy linkage: for an energy sleeve task, buy bonds with energy_linked == true. (energy_linked is a bond-level field, not the issuer sector; e.g. a utilities-issuer power bond can be energy-linked.)
   • Issuer concentration: no single issuer should exceed issuer_concentration_limit_pct (12%) of post-trade market value.
4. Select the package honoring the desk's stated preferences (e.g. LNG exporters / gas demand / non-watchlist carry) and the "improve carry" goal (higher YTM helps). A robust pattern for a two-ticket income add:
   • Pick one IG bond that matches the preferred theme (LNG/gas) for ballast + client pitch suitability.
   • Pick one higher-carry non-watchlist energy-linked bond (often HY) that lifts portfolio YTM while keeping HY % well under cap and adding issuer + subsector diversification.
   • Verify post-trade weighted duration stays inside the band and that new issuers/subsectors increase diversification. For a rotation (risk-reduction) task:
   • Sell watchlist holdings first (entire positions); the watchlist sells are reported in watchlist_sell_ids.
   • Sell additional HY as needed to bring post-trade HY % under cap and meet target_hy_reduction_pct. Prefer keeping higher-carry non-watchlist HY and trimming lower-carry HY when both are eligible, so carry is preserved.
   • Fund duration ballast by buying IG candidates from the shortlist that are non-watchlist and inside the duration band; weight notionals toward longer duration to replace duration lost when selling short-duration HY.
   • Keep the package roughly self-funding (total sell notional ≈ total buy notional) so total market value is preserved.
5. Order the trades as the template declares. Income-add: ascending by instrument_id. Rotation: SELL before BUY, then instrument_id ascending within each action. Notional field name follows the template (notional_usd_m for income-add; quantity_usd_m for rotation).

Post-trade metrics computation (mechanical, market-value-weighted)

Build the post-trade book = current holdings adjusted by trades (SELL subtracts, BUY adds; use environment quantities). Then:

• total_market_value_usd_m = sum of post-trade quantity_usd_m.
• hy_allocation_pct = (sum of HY quantity_usd_m) / total_market_value × 100, rounded to 2 decimals. (HY defined by bond rating_bucket == "HY".)
• weighted_modified_duration_years = Σ(quantity · modified_duration_years) / total_market_value, rounded to 2.
• weighted_yield_to_maturity_pct = Σ(quantity · yield_to_maturity_pct) / total_market_value, rounded to 2.

For rotation outputs:

• post_trade_hy_allocation_pct = post-trade HY % (2 decimals).
• post_trade_duration_years = post-trade weighted modified duration (2 dp).
• hy_reduction_pct_points = pre_trade_HY% − post_trade_HY% (2 dp, can be large when a watchlist HY block is removed).
• post_trade_watchlist_exposure_usd_m = sum of post-trade quantities whose issuer watchlist == true (1 decimal; should be 0.0 if all watchlist names were sold).
• Pre-trade HY % is computed the same way on the un-adjusted holdings.

Verify every metric by recomputing; do not trust stale worksheet snapshots.

Constraint-check derivation (booleans)

• hy_cap_pass = post-trade HY % ≤ max_hy_allocation_pct (20.0).
• duration_band_pass = post-trade weighted duration within duration_band_years (inclusive [3.0, 5.0]).
• issuer_diversification_pass / selected_issuer_diversification_pass = no single issuer exceeds issuer_concentration_limit_pct (12%) of post-trade MV (and, for "selected" variants, the newly selected issuers are distinct from each other / add diversification).
• subsector_diversification_pass / selected_subsector_diversification_pass = the holding set (or selected buys) spans at least subsector_min_count_for_diversified (2) subsectors.
• watchlist_avoidance_pass = no bought bond has a watchlisted issuer (and for rotation, watchlist_exposure_cleared = post-trade watchlist exposure == 0).
• target_hy_reduction_met (rotation only) = hy_reduction_pct_points ≥ target_hy_reduction_pct (4.0 for credit_risk_reduction policy).

Sales positioning (energy income-add)

Read GET /api/market/energy. Pick the strongest positive signal that matches your selected bonds' theme tags:

• LNG export pull (highest positive score) → theme: lng_export_tailwind.
• Midstream defensive carry → midstream_stability.
• Oil discipline / oversupply caution → oil_oversupply_caution.
• Watchlist yield avoidance → avoid_watchlist_yield_trap.
• Selective transition/renewables → transition_bond_selectivity.

target_segment follows the client context in the desk request (e.g. "multi-asset income update" → multi_asset_income; insurance / pension / private-bank / endowment contexts map to their respective enum). When refining/refiner signals are negative and watchlisted, do NOT pick a refining theme; the refining signal being negative + AVOID_REFINING_WATCHLIST pitch theme reinforces watchlist avoidance.

risk_note_code (rotation)

Pick the dominant driver of the rotation: watchlist_concentration (when the main issue was a watchlisted holding), hy_cap_pressure (when HY cap drove sells), duration_preservation (when the rotation's point was keeping duration in band), carry_tradeoff (when selling carry to reduce risk), no_action (when nothing needed fixing). If watchlist exposure was the trigger, watchlist_concentration is correct.

4. Workflow B — International equity correlation review

No correlations are precomputed. Compute from GET /api/index-levels.

Correlation method (confirmed deterministic)

1. Take the review window from the request payload (review_window.level_start_date / level_end_date) — these match the policy correlation.review_window_start/end for the default window. For each index in the requested index_universe, extract the monthly levels in date order across the window.
2. Compute monthly simple returns: r[t] = level[t] / level[t-1] − 1 for consecutive observations. A window with 12 month-end levels yields 11 return observations (review_window.return_observations = 11).
3. For each unordered pair of indices, compute the Pearson correlation of their return series: corr = cov(r_x, r_y) / (σ_x · σ_y) (population, dividing by n).
4. extreme_pairs.highest_positive = the pair with max correlation; extreme_pairs.lowest = the pair with min correlation (most negative). pair_id lists the two index ids sorted alphabetically. Round correlation to 3 decimals.
5. index_set = the requested universe, sorted ascending alphabetically.

Concentration flags

Using the policy thresholds correlation_high_threshold (0.8) and correlation_low_threshold (0.2):

• high_threshold_breached = any pair correlation ≥ 0.8.
• china_asia_dependence_flag = true when IDX_CHINA correlation with the Asia/EM/developed block (IDX_EM, IDX_AC_ASIA_PAC_EX_JP, IDX_ACWI_IMI, IDX_WORLD, IDX_EAFE, IDX_EM_EX_CHINA, IDX_INDIA) is ≥ 0.8 — i.e. China is tightly coupled to the regional/global complex.
• primary_code:
  • CHINA_ASIA_DEPENDENCE when China↔Asia/EM pairs drive the concentration (CIO memos flag ASIA_BETA_OVERLAP / CHINA_DEDICATED_SLEEVE).
  • GLOBAL_DEVELOPED_OVERLAP when the dominant high-correlation cluster is developed/global (IDX_WORLD, IDX_EAFE, IDX_ACWI_IMI) rather than China.
  • NO_MATERIAL_CONCENTRATION when no pair breaches the high threshold.

Diversification candidates

From the template's allowed candidate set, select indices that diversify the concentration center (typically China). Rank allowed candidates by their correlation with the concentration center ascending (most negative / lowest first). Include those that meaningfully diversify:

• Indices with negative correlation to the center (e.g. IDX_LATAM, which is negatively correlated to everything) always qualify.
• Indices that structurally remove the concentration (e.g. IDX_EM_EX_CHINA) qualify as the "ex-China" diversifier.
• Exclude an allowed candidate whose correlation with the concentration center breaches the high threshold in a way that adds to (not relieves) the concentration (e.g. IDX_INDIA when it is ≥0.8 with China does not diversify China risk). Sort the resulting list ascending alphabetically.

Sleeve actions

Produce the number of sleeve-action objects the template requires, ordered ascending by sleeve name. Map from the review's conclusions:

• trim the concentrated sleeve (the source of concentration, e.g. the China dedicated sleeve).
• add the best diversifier sleeve (e.g. Latin America).
• hold / monitor / hedge / rotate for other sleeves as appropriate. target_index_id and action must use the template's allowed enums.

5. Workflow C — Cross-asset active allocation view refresh

This is largely deterministic — derive views mechanically from macro signal scores and the allocation_mapping policy. Do not editorialize.

Per opportunity set (one allocation_views row)

1. prior_view = the view from GET /api/allocation/prior-views for the entry where quarter == target_quarter AND previous_quarter == prior_quarter. (This is the standing view set last quarter for the current quarter.)
2. signal_score = GET /api/macro-signals score for (opportunity_set, quarter == target_quarter) (round to 3 decimals if the template exposes it).
3. rationale_code = that macro-signal entry's rationale_code (verbatim from the environment — e.g. CHINA_DEPENDENCE, INDIA_OFFSET, LATAM_DIVERSIFIER, DURATION_SUPPORT, HY_VALUATION_RISK, EUROPE_RECOVERY, JAPAN_POLICY_RISK, DOLLAR_DEFENSIVE, NEUTRAL_BALANCE, RATE_CUT_SUPPORT, CREDIT_SPREAD_RISK, GROWTH_IMPROVES).
4. view from signal_score via allocation_mapping.view_score_thresholds:
   • score ≥ OW_min (0.35) → OW
   • score ≤ UW_max (−0.35) → UW
   • otherwise → N
5. conviction from |signal_score| via conviction_thresholds:
   • |score| ≥ HIGH_abs_min (0.7) → HIGH
   • MEDIUM_abs_min (0.35) ≤ |score| < 0.7 → MEDIUM
   • |score| < LOW_abs_below (0.35) → LOW
6. change from view_rank (UW=−1, N=0, OW=1): compare rank(new_view) vs rank(prior_view):
   • new > prior → UP
   • new < prior → DOWN
   • equal → UNCHANGED
7. asset_class = the asset_class from GET /api/allocation/opportunity-sets for that opportunity set (Equities / Duration / Credit / Currency).
8. Row order = the request payload's focus_opportunity_sets order (not alphabetical). required_length matches the focus list size.
9. policy_id = the top-level policy_id from GET /api/policies (e.g. POLICY_SET_2026_05). target_quarter / prior_quarter = the request values. as_of_date = environment as_of_date.

Risk overlay (judgment, but patterned)

Choose overlay_code / primary_action from the macro picture formed by the rows (paired enums: 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):

• If duration/quality is supported (U.S. Treasuries OW, DURATION_SUPPORT) and HY is a risk (Corporate HY UW, HY_VALUATION_RISK) → DURATION_QUALITY_TILT / tilt_to_duration_quality.
• If HY/credit spread risk dominates → CREDIT_RISK_REDUCTION / trim_credit_beta.
• If defensive currency is the theme → CURRENCY_DEFENSIVE_HEDGE / add_currency_hedge.
• If cyclical growth recovery dominates with no risk offset → EQUITY_BETA_EXTENSION / add_cyclical_equity_beta.
• If views net to roughly neutral with no material risk → NO_OVERLAY / hold_policy_weights.

rationale_codes = the rationale codes driving the overlay choice, ordered by business priority (highest priority first). Favor the risk-reduction and duration-support rationales first (e.g. DURATION_SUPPORT, HY_VALUATION_RISK, CHINA_DEPENDENCE), then growth/diversifier codes. Keep the list to the 2–4 most material drivers.

6. Combined workflow — correlation + allocation (committee JSON)

Some tasks (multi-asset committee files) combine B and C: a small index sub-universe correlation summary plus allocation views for the matching sleeves.

• correlation_summary (length 2, ordered [highest_concentration, best_diversifier]): compute Pearson over the requested index subset; highest_concentration = the max-correlation pair; best_diversifier = the min (most negative) pair. Each pair sorted alphabetically; correlation rounded to 3 decimals.
• target_sleeve_actions: one row per requested opportunity set, in the template's item_order. Map view → action: OW → add, UW → trim. For a currency sleeve (USD) that is downgraded from OW toward neutral, use hedge (reduce the defensive long). Stable OW → add/hold; stable N → hold/ monitor.
• allocation_views: same deterministic computation as Workflow C (prior_view, signal_score, view, change, conviction, rationale_code), in the template's item_order.
• rebalance_trigger:
  • correlation_cap_breach when a high-threshold pair breach drove the review (most common when the correlation summary shows a pair ≥ 0.8).
  • hy_cap_pressure / duration_drift / watchlist_concentration for those drivers; committee_review as the generic catch-all.
• portfolio_risk_concentration_flag: true when the correlation summary or constraint picture shows material concentration (a pair ≥ high threshold, or a China/Asia dependence flag).
• next_step:
  • approve_with_monitoring when actions reduce concentration but residual monitoring is warranted (common when the package trims the concentrated sleeve and adds diversifiers without a hard breach remaining).
  • approve_rotation when the proposed rotation fully resolves the issue.
  • defer_pending_risk_review when material exceptions remain unresolved.
  • reject_constraint_breach when a constraint breach cannot be fixed.

7. Common misjudgments to avoid

• Trusting stale holding quantities / marks. The local memo is intake context. A "stale_exception_board" with stale_quantity_usd_m or a "stale_holding_snapshot" is explicitly unreliable — recompute off the environment portfolio. Sell quantities must equal the environment's current holding quantity when exiting a position.
• Buying a watchlisted issuer. Always join bonds → issuers and check watchlist. High-carry HY names (yield 9–11%) are often traps: a watchlisted refiner or E&P issuer with a HIGH_CARRY/WATCHLIST_RISK tag should be excluded even if it maximizes YTM.
• Duration-ineligible distractors. Long-dated IG names (dur 5.8–6.7) and very short HY names (dur 2.3) sit outside the [3.0, 5.0] band. Do not pick them as buys; flag the band as the reason.
• Misreading energy_linked. It is a bond-level boolean, not the issuer's sector label. Filter buys for the task's sector requirement using this flag.
• Forgetting the HY cap after adds. A high-YTM HY buy can push post-trade HY % over 20%; recompute and choose a smaller notional or an IG alternative.
• Computing correlations wrong. Use simple monthly returns (level[t]/level[t-1]−1), not log returns or level changes; use Pearson (cov/σσ); the number of return observations is one fewer than the number of levels. Round to 3 decimals; sort pair ids alphabetically.
• Inventing allocation views. Views/conviction/rationale are computed from macro-signals + the allocation_mapping thresholds + prior-views — not judgment. A score of 0.732 is OW/HIGH/INDIA_OFFSET, not a subjective call.
• Wrong change. change compares the new view rank to the prior standing view (from prior-views for the target quarter), not to a guessed baseline. OW(1)→UW(−1) is DOWN; N(0)→OW(1) is UP; OW→OW is UNCHANGED.
• Wrong as_of_date. Use the environment's current as_of_date, not the request_date / memo_as_of_date / snapshot_date in the local payload.
• Wrong trade ordering. Income-add packages sort by instrument_id ascending; rotation packages put SELL before BUY then instrument_id ascending within each action.
• Including rationale_code not in the enum. Only use rationale codes that appear in the template's allowed_values and are returned by the macro-signals endpoint verbatim.
• Over- or under-rounding. Match each field's declared precision exactly.

8. Final output checklist

Before emitting the JSON:

• Every required top-level key present; no extra narrative outside the JSON (unless the prompt allows it — most say "Return only JSON").
• as_of_date = environment as_of_date; portfolio_id / task_id / target_quarter / prior_quarter / review_quarter match the template's required_value.
• All enums use allowed values verbatim; all booleans are real booleans.
• Every numeric field rounded to its declared precision.
• Lists sorted and sized per template (ordering, length, required_length, item_order).
• trade_package / rotation.trades ordering correct (instrument_id asc; SELL-before-BUY where applicable).
• Post-trade metrics recomputed from environment holdings + trades (not the stale snapshot); constraint booleans derived from recomputed metrics vs policy thresholds.
• Watchlist joins performed; no watchlisted issuer bought; watchlist sells reported correctly.
• Correlations computed from monthly simple returns (Pearson, 3 decimals).
• Allocation views computed from macro-signals + allocation_mapping policy + prior-views; change vs the matching prior-views cohort.
• data_precedence reflects whether a local payload was stale vs the environment.
• Validate the final object against answer_template.json field-by-field.