Self Attempt 02

Asteria Investment Office — Portfolio-Risk Task Skill

Operating SOP for solving Asteria portfolio-risk evaluation tasks. The test solver receives only a task input/ (prompt + payloads/answer_template.json + a request payload) plus this skill and the remote environment. Three workflow families appear; a task may involve one or combine two.

Base URL: <remote-env-url> (GET only). Use curl or python3+urllib.

0. Golden rules (apply to every task)

1. Environment = book of record. Every input/payloads/*.json is intake context and may be stale (old marks, old quantities, old desk notes). Always reconcile to the environment. When a local payload conflicts with the environment, prefer the environment. The data_precedence answer field is current_environment_over_stale_payload whenever any conflict exists (e.g. stale snapshot MV/quantity differs from /api/portfolios/<id>).
2. as_of_date in your answer = the environment's current as-of date. Read it from /api/portfolios/<id> (as_of_date) or /api/policies (as_of_date). Do not copy a date stamped on a local request payload or a stale worksheet unless the prompt explicitly overrides.
3. No precomputed correlations. The environment stores index levels, never correlations. Compute Pearson correlations yourself from monthly simple returns (see Workflow B).
4. Precision follows each field in answer_template.json. Round the final value to the declared decimals. JSON numerics like 5.8 and 5.80 are equal, but always round to the specified precision (2/3/1 decimals) before emitting.
5. Read the answer template first. It declares required keys, enums, ordering, and precision per field. Conform exactly: missing keys, wrong enum values, wrong sort order, or wrong precision all fail.
6. Output is a single JSON object matching the template — no narrative outside JSON unless the prompt asks for it.

1. Environment endpoints (what each is for)

GET /api/catalog may return empty; ignore it and use the typed endpoints above.

Reuse fetches. /api/instruments/bonds, /api/issuers, /api/index-levels, /api/allocation/prior-views, /api/macro-signals are small and shared across tasks — fetch once, cache to /tmp/*.json, reuse.

2. Policy reference (/api/policies)

• credit_default (POL_CREDIT_DEFAULT): max_hy_allocation_pct=20.0, duration_band_years=[3.0, 5.0], 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-reduction rotations).
• correlation (POL_CORRELATION_DEFAULT): correlation_high_threshold=0.8, correlation_low_threshold=0.2, review window (e.g. 2025-05-30 → 2026-04-30).
• allocation_mapping (POL_ALLOCATION_MAPPING):
  • view_score_thresholds: OW if score ≥ 0.35; UW if score ≤ -0.35; N in (-0.35, 0.35).
  • conviction_thresholds: HIGH if |score| ≥ 0.70; MEDIUM if |score| ≥ 0.35; LOW if |score| < 0.35.
  • view_rank: OW=1, N=0, UW=-1 (for computing change vs prior).
• multi_asset (POL_MULTI_ASSET_DEFAULT): uses allocation_mapping + correlation_default + credit_default. Used by multi-asset sleeves.
• multi_asset_risk (POL_MULTI_ASSET_RISK): uses correlation_default + credit_risk_reduction; committee_escalation_threshold= "two_or_more_material_exceptions" → next_step logic in Workflow C.

The portfolio's constraints.policy_id tells you which policy set governs it.

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

Tasks: build a BUY package (income sleeve) or a SELL+BUY rotation (risk-reduction). Answer fields: trade_package/rotation.trades, post_trade_metrics / risk_metrics, constraint_checks / exception_flags, sales_positioning or watchlist_handling, risk_note_code, data_precedence.

3.1 Gather state

1. GET /api/portfolios/<id> → holdings (instrument_id, quantity_usd_m), MV, constraints.policy_id.
2. GET /api/instruments/bonds → bond metrics (rating_bucket, modified_duration, YTM, spread, energy_linked, subsector, issuer_id).
3. GET /api/issuers → cross-reference watchlist flag per issuer_id and subsector/sector/rating.
4. GET /api/market/energy → signal scores + pitch_themes (energy tasks).

3.2 Pre-trade metrics (market-value-weighted)

Using quantity_usd_m as the market-value weight:

• HY allocation % = Σ(HY quantity) / Σ(all quantity) × 100.
• weighted modified duration = Σ(qty × modified_duration_years) / Σ(qty).
• weighted YTM = Σ(qty × yield_to_maturity_pct) / Σ(qty).

Same weighting applies post-trade after applying buys/sells (post MV = pre MV

• buys − sells; for a "new sleeve allocation" MV grows by the buy notional; for a "fund from proceeds" rotation MV stays ~constant, i.e. Σbuys = Σsells).

3.3 Eligibility & exclusion rules (critical)

• Watchlist avoidance: never BUY a bond whose issuer watchlist is true. Cross-check the issuer_id from /api/issuers. Watchlist flags are on the issuer, not the bond — a non-watchlist-looking bond whose issuer is watchlisted is still forbidden.
• Duration-ineligible distractors: bonds with modified_duration_years outside [3.0, 5.0] are ineligible as BUY candidates (the duration band is enforced on both the portfolio and effectively on eligible candidates; long 2033/2034 paper at dur 5.4–6.7 and short paper at dur 2.3–2.9 are distractors). Always prefer candidates with duration inside the band.
• HY cap: post-trade HY % must be ≤ 20.0 (hy_cap_pass). If the book is already over cap, the rotation must sell enough HY to get under.
• Duration band: post-trade weighted modified duration must be in [3.0, 5.0] (duration_band_pass). For rotations, keep duration roughly flat ("without creating a duration shortfall").
• Diversification (selected_*_diversification_pass): the selected BUY package must span ≥2 distinct issuers AND ≥2 distinct subsectors (subsector_min_count_for_diversified=2). This is a check on the selected trades, not the whole book. Do not pair two bonds from the same issuer or same subsector.
• Avoid doubling up on already-overweight issuers where it would push a single issuer over 12% concentration, but the hard "selected" check is the 2-buy diversification rule above.

3.4 Selection SOP — income BUY package (2 tickets, even split)

1. Compute the package notional from the desk request (e.g. 8.0m total → 4.0 each, 1-decimal precision).
2. Among candidate=true bonds that are energy_linked=true (for energy sleeves), non-watchlist, duration-in-band, pick an anchor matching the desk's top preference and the strongest /api/market/energy signal. The LNG export signal is typically strongest; an IG LNG exporter bond (IG, dur ~4, themes LNG_EXPORTS/GAS_DEMAND) is the usual anchor.
3. Pick the second bond from a different issuer and different subsector, also energy-linked, non-watchlist, duration-in-band, that improves carry (YTM above the pre-trade weighted YTM). If HY budget allows (post HY ≤ 20%), a non-watchlist HY carry bond maximizes carry improvement; choose the highest-YTM eligible HY diversifier. Renewables / merchant-power HY are typical carry diversifiers.
4. Verify all five constraint_checks; compute post_trade_metrics (2 decimals); notional_usd_m to 1 decimal; sort trade_package ascending by instrument_id.
5. sales_positioning.target_segment: take from the request's client_context (e.g. "multi-asset income" → multi_asset_income).
6. sales_positioning.theme: pick the enum matching the dominant pitch_themes/signal (LNG-led → lng_export_tailwind; renewables+LNG → transition_bond_selectivity; avoiding watchlist HY carry → avoid_watchlist_yield_trap; midstream → midstream_stability).
7. data_precedence: current_environment_over_stale_payload if the stale snapshot/MV/quantity differs from the environment, else no_conflict_found.

3.5 Selection SOP — risk-reduction rotation (SELL+BUY)

1. Identify pressure points to sell: watchlist holdings (issuer watchlist=true) first, then HY holdings, prioritizing the lowest-carry HY for sale while keeping the highest-carry non-watchlist HY to preserve carry.
2. Sell enough HY to (a) clear watchlist exposure to 0 and (b) get post-trade HY % ≤ 20 (and meet the target_hy_reduction_pct ≥ 4 pp where the policy is POL_CREDIT_RISK_REDUCTION — getting under the 20% cap always satisfies 4 pp).
3. Buy IG candidate=true bonds from the candidate shortlist that are non-watchlist, duration-in-band. Fund the buys with the sell proceeds (Σbuys = Σsells, MV constant). Pick the highest-YTM IG candidates across different sectors for carry + diversification.
4. Trivially never buy a watchlist HY candidate even if it appears on the desk's shortlist with a "high carry" comment (buys_avoid_watchlist=true).
5. Sort rotation.trades SELL before BUY, then instrument_id ascending within each action. quantity_usd_m to 1 decimal.
6. risk_metrics: post_trade_hy_allocation_pct (2), post_trade_duration_years (2), hy_reduction_pct_points (2) = pre_HY% − post_HY%, post_trade_watchlist_exposure_usd_m (1).
7. watchlist_handling.watchlist_sell_ids: all sold watchlist instrument_ids, ascending.
8. risk_note_code: the dominant theme — watchlist_concentration if a watchlist name was the acute risk removed; hy_cap_pressure if the HY-over- cap was the driver; duration_preservation if keeping duration in-band was the binding constraint; carry_tradeoff if high-carry HY was sacrificed for IG quality. no_action only if no trades.

4. Workflow B — International equity correlation review

Tasks: pair correlations across an index universe from monthly levels; highest/lowest pairs; China/Asia dependence concentration; diversification candidate set; sleeve actions.

4.1 Compute correlations (yourself)

1. GET /api/index-levels → levels per index. Use the window from the request's review_window (level_start_date, level_end_date) — which equals the policy correlation.review_window_start/end and the index metadata window.
2. For each index, take levels with start_date <= date <= end_date, sorted by date. Monthly simple return r_t = (level_t − level_{t-1}) / level_{t-1}.
3. return_observations = number of returns = (number of levels in window) − 1 (12 monthly levels → 11 returns).
4. Pearson correlation between two return series: r = Σ((xᵢ−x̄)(yᵢ−ȳ)) / sqrt(Σ(xᵢ−x̄)² · Σ(yᵢ−ȳ)²).
5. Round correlations to 3 decimals.
6. pair_id / pair = the two index ids in ascending alphabetical order.

4.2 Extreme pairs

• highest_positive = the pair with the maximum correlation.
• lowest = the pair with the minimum correlation (most negative). Both report pair_id (sorted) and correlation (3 decimals).

4.3 Concentration flags

• china_asia_dependence_flag = true when the China index and the Asia-Pacific- ex-Japan index (and/or EM) are correlated above the high threshold (0.8), i.e. the dedicated China + Asia sleeves carry overlapping beta.
• high_threshold_breached = true when the relevant concentration pair crosses correlation_high_threshold (0.8).
• primary_code: CHINA_ASIA_DEPENDENCE if the China/Asia/EM overlap is the dominant concentration (matches the CIO concern codes ASIA_BETA_OVERLAP / CHINA_DEDICATED_SLEEVE); GLOBAL_DEVELOPED_OVERLAP if the dominant high correlations are developed-world (EAFE/WORLD/ACWI); NO_MATERIAL_CONCENTRATION if nothing breaches.

4.4 Diversification candidates & sleeve actions

• diversification_candidates: the ex-China EM diversifier pool (e.g. IDX_EM_EX_CHINA, IDX_INDIA, IDX_LATAM), ascending alphabetical — the sleeves the committee evaluates for diversification. (The genuinely low/negatively- correlated one — typically LatAm vs China — is the one to add.)
• sleeve_actions (length per template, ordered ascending by sleeve name): trim the concentrated sleeve (the high-beta China/Asia sleeve → action trim, target that index) and add the low-correlation diversifier sleeve (→ action add, target the diversifying index). Use monitor/hold/rotate sparingly only when no size change is warranted. target_index_id must be from the template's allowed set.

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

Tasks: produce per-opportunity-set active views (view, change, conviction, rationale_code) from prior views + macro signals; choose a risk overlay; and (combined variant) link correlations to sleeve actions + rebalance trigger

• next step.

5.1 Source the inputs

• Prior (prior-quarter) views: GET /api/allocation/prior-views. Each record has quarter, previous_quarter, opportunity_set, view, conviction. For a target quarter Q with prior quarter P, the prior view = the record where quarter=Q and previous_quarter=P. (Its view/conviction fields ARE the prior-quarter view you compare against — do not treat them as the new view.) The new view is derived from signals below.
• Current signals: GET /api/macro-signals, filter quarter=Q. Each gives score (signal_score) and rationale_code.
• Mapping policy: POL_ALLOCATION_MAPPING (§2).
• Asset class: from /api/allocation/opportunity-sets (opportunity_set → asset_class: Equities / Duration / Credit / Currency).

5.2 Derive each allocation row

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

1. asset_class from the opportunity-sets taxonomy.
2. signal_score = macro-signal score, 3 decimals.
3. view (new) from score via mapping: OW if ≥0.35, UW if ≤−0.35, else N.
4. conviction from |score|: HIGH ≥0.70, MEDIUM ≥0.35, LOW <0.35.
5. rationale_code = macro-signal rationale_code (must be one of the enum).
6. prior_view (combined-variant template) = the prior-views record's view.
7. change = compare new view rank vs prior view rank (OW=1,N=0,UW=−1): UP if new>prior, DOWN if new<prior, UNCHANGED if equal.
8. Order rows exactly as the request's focus_opportunity_sets list (or the template's item_order).

5.3 Choose the risk overlay

Pick the single overlay that captures the dominant cross-asset rotation:

• If credit (Corporate High Yield) is UW and/or duration (U.S. Treasuries) is OW → rotate credit→duration: DURATION_QUALITY_TILT / tilt_to_duration_quality.
• If HY UW is the single clearest risk with no duration bid → CREDIT_RISK_REDUCTION / trim_credit_beta.
• If cyclical equities (Europe/LatAm) are the leading OWs with no defensive rotation → EQUITY_BETA_EXTENSION / add_cyclical_equity_beta.
• If USD/defensive currency view is the headline → CURRENCY_DEFENSIVE_HEDGE / add_currency_hedge.
• Only NO_OVERLAY / hold_policy_weights when no view departs materially from neutral. rationale_codes: the 1–3 enum codes driving the overlay, business priority order (largest |signal| / most material risk first). For a duration-quality tilt use the duration-support and credit-risk codes; for credit-risk reduction use the negative credit/equity-risk codes ordered by severity.

policy_id (where the template requires it) = POL_ALLOCATION_MAPPING.

5.4 Combined correlation + allocation variant (multi-asset committee JSON)

This variant (e.g. PF-MA-HELIO) asks for both a correlation_summary and allocation_views plus target_sleeve_actions, rebalance_trigger, portfolio_risk_concentration_flag, next_step.

1. correlation_summary (length 2, order [highest_concentration, best_diversifier]): from the named index subset, the highest-correlation pair = concentration risk; the most-negative pair = best diversifier. Pair ids sorted alphabetically; correlation 3 decimals; compute as in §4.
2. allocation_views: as §5.2 for the named opportunity sets (typically Emerging Markets, India, Latin America, USD), in the template's item_order.
3. target_sleeve_actions: one per opportunity set in item_order. Map the view+correlation to an action: trim sleeves whose index is in the high-concentration pair and whose view is UW; add sleeves that are the low-correlation diversifier / OW offset; hold for sleeves whose view moved to neutral (revert to policy weights); monitor/hedge only when clearly warranted.
4. rebalance_trigger: correlation_cap_breach if a pair exceeds the 0.8 high threshold (the usual trigger for these reviews); else hy_cap_pressure / duration_drift / watchlist_concentration if a credit exception dominates; committee_review as the catch-all.
5. portfolio_risk_concentration_flag: true if any concentration pair breaches the high threshold (0.8).
6. next_step: count material exceptions (correlation breach + any credit exception). Under POL_MULTI_ASSET_RISK, two-or-more → defer_pending_risk_review; one addressable exception → approve_with_monitoring; clean rotation that resolves the breach → approve_rotation; unresolvable breach → reject_constraint_breach. Under POL_MULTI_ASSET_DEFAULT a single correlation breach that the rotation addresses typically → approve_with_monitoring.

6. Precision & ordering conventions (all workflows)

Ordering rules (strict):

• trade_package: ascending by instrument_id.
• rotation.trades: SELL before BUY, then instrument_id ascending within each action.
• index_set, diversification_candidates, pair/pair_id: ascending alphabetical by index id.
• sleeve_actions: ascending by sleeve.
• allocation_views/target_sleeve_actions: the request's focus_opportunity_sets order (or template item_order).
• watchlist_sell_ids: ascending instrument_id.
• rationale_codes: business priority, highest priority first.

7. Common pitfalls & exclusion rules

• Stale worksheet trap: a desk payload may state MV, HY%, duration, or holding quantities that differ from the environment. Always recompute from /api/portfolios/<id> holdings + /api/instruments/bonds. Set data_precedence = current_environment_over_stale_payload.
• Watchlist is on the issuer, not the bond: a bond's rating/sector can look fine while its issuer_id is watchlist=true. Always join to /api/issuers. Watchlist HY "high carry" candidates on a desk shortlist are deliberate traps — never buy them.
• Duration-ineligible distractors: long-dated 2032–2034 paper (dur 5.4–6.7) and very short 2028 paper (dur 2.3–2.9) sit outside [3,5]; exclude as buys.
• HY cap math: HY % is of post-trade MV (after buys/sells), not pre-trade. For a rotation that keeps MV constant, the denominator is unchanged; for a new-sleeve BUY package, MV grows by the buy notional.
• Diversification is on the selected package: do not reject a pair just because the broader book is concentrated; the selected_*_diversification_pass checks the 2 chosen buys span ≥2 issuers and ≥2 subsectors.
• Correlations are not stored: never look for a "correlation" field in the environment; compute from index-levels. Use simple (not log) returns.
• Prior view vs new view: in /api/allocation/prior-views, the record quarter=Q, previous_quarter=P holds the prior (P) view for a Q target. The new Q view comes from Q's macro-signals + the mapping policy. Getting this backwards inverts every change.
• Conviction uses |score|: a UW view at score −0.373 has MEDIUM conviction (|0.373|≥0.35), not LOW. A score of exactly ±0.35 is MEDIUM (≥ threshold).
• change direction: UP/DOWN by view rank, not by score magnitude. OW(1)→N(0) is DOWN; N(0)→UW(−1) is DOWN; N→OW is UP.
• Enum casing & values: match the template's allowed_values exactly (e.g. multi_asset_income not multi-asset-income; Q2_2026 not Q2-2026).
• Trailing zeros: JSON numerics 5.8 and 5.80 are equal, but always round to the template's declared decimals before emitting.

8. End-to-end solving checklist

1. Read prompt.txt, payloads/answer_template.json, and the request payload. Note required keys, enums, ordering, precision, and which workflow(s) apply.
2. Fetch the portfolio (/api/portfolios/<id>) and the relevant universe endpoints; cache large ones to /tmp.
3. Identify the governing policy_id from the portfolio constraints and read the matching policy thresholds from /api/policies.
4. Apply the workflow SOP (§3 / §4 / §5).
5. Recompute every numeric field from environment data; round to template precision; apply the template's sort order.
6. Re-check every enum value against allowed_values; re-check every boolean (constraint pass/fail) against the policy thresholds.
7. Emit a single JSON object conforming to the template — no extra keys, no narrative.

Reusable correlation snippet (python3)

import json, itertools
from statistics import mean
lv = json.load(open('/tmp/index_levels_all.json'))
START='2025-05-30'; END='2026-04-30'
def rets(idx):
    rows=sorted([r for r in lv[idx] if START<=r['date']<=END],key=lambda r:r['date'])
    L=[r['level'] for r in rows]
    return [(L[i]-L[i-1])/L[i-1] for i in range(1,len(L))]
def pearson(x,y):
    n=len(x); mx=mean(x); my=mean(y)
    sxy=sum((x[i]-mx)*(y[i]-my) for i in range(n))
    return sxy/((sum((a-mx)**2 for a in x)**0.5)*(sum((b-my)**2 for b in y)**0.5))
R={i:rets(i) for i in UNIVERSE}
pairs=sorted(((a,b,pearson(R[a],R[b])) for a,b in itertools.combinations(sorted(UNIVERSE),2)),key=lambda t:t[2])
highest=max(pairs,key=lambda t:t[2]); lowest=pairs[0]  # pairs[0] is min

Reusable weighted-metric snippet

bonds={b['instrument_id']:b for b in json.load(open('/tmp/bonds_all.json'))}
# port = {instrument_id: market_value_usd_m}
mv=sum(port.values())
hy=sum(q for iid,q in port.items() if bonds[iid]['rating_bucket']=='HY')/mv*100
dur=sum(q*bonds[iid]['modified_duration_years'] for iid,q in port.items())/mv
ytm=sum(q*bonds[iid]['yield_to_maturity_pct'] for iid,q in port.items())/mv
print(round(mv,2),round(hy,2),round(dur,2),round(ytm,2))