Self Attempt 03

Asteria Investment Office — Institutional Portfolio-Risk Solver Skill

A transferable SOP for solving Asteria Investment Office portfolio-risk tasks. Three workflows feed the test set: (A) energy/fixed-income credit trade strategy, (B) international equity correlation review, (C) cross-asset active allocation view updates. Some test tasks combine B+C.

0. Golden rules (read first)

1. The remote environment is the book of record. Base URL <remote-env-url>. Every input/payloads/*.json is intake context and may be stale (old worksheet marks, stale quantity boards, prior-week shortlists, old desk notes). When a payload conflicts with the environment, prefer the environment. The answer field data_precedence / your reasoning must reflect this.
2. Fetch fresh. Do not trust locally-cached payload numbers for holdings, quantities, or marks. Always pull /api/portfolios/<id> for current holdings and /api/instruments/bonds, /api/issuers for security master. Stale-quantity boards in particular disagree with the portfolio service; use the portfolio service quantities.
3. No correlations are stored. Compute every correlation yourself from /api/index-levels (monthly simple returns, Pearson). Never invent or reuse a correlation.
4. Precision is declared per-field in answer_template.json. Respect it exactly:
   • notional_usd_m / quantity_usd_m / post_trade_watchlist_exposure_usd_m → 1 decimal.
   • All post_trade_metrics, risk_metrics percents/years → 2 decimals.
   • correlation and allocation signal_score → 3 decimals.
   • Round half-up at the end; do not round intermediate inputs.
5. Output exactly one JSON object matching the template's required keys, enums, and ordering. Sort trade lists as the template dictates (ascending instrument_id; SELL-before-BUY then instrument_id). Pair index ids alphabetically inside each pair.
6. No judge, no hidden gold. Derive every value from environment + template. Do not assume a "correct" pre-decided answer; the environment + policy thresholds define correctness.

1. Endpoints that matter

All GET, all on <remote-env-url>:

• /api/catalog — every portfolio/bond/issuer/index/policy/opportunity-set id. Start here to confirm ids.
• /api/portfolios — summaries (id, objective, strategy, constraint_policy_id, MV, holding count, as_of_date).
• /api/portfolios/<id> — objective, constraints, current holdings (instrument_id, quantity_usd_m, sleeve, notes), MV. This is authoritative for holdings & quantities.
• /api/instruments/bonds (filters: ?candidate=true, ?rating_bucket=HY) — bond universe. Per bond: candidate, energy_linked, rating, rating_bucket (IG/HY), modified_duration_years, yield_to_maturity_pct, coupon_pct, spread_bps, sector, subsector, issuer_id, maturity, recommended_theme_tags.
• /api/issuers — per issuer: sector, subsector, rating_bucket, watchlist (bool), credit_outlook, research_tags. Join to bonds via issuer_id.
• /api/policies — one object with sub-policies: credit_default, credit_risk_reduction, correlation, allocation_mapping, multi_asset, multi_asset_risk. Each portfolio's constraint_policy_id names which applies.
• /api/market/energy — commodity signals (oil/gas/LNG/refining/renewables) with score, direction, signal_id, pitch_themes, stale_data_warning. Use for energy-desk theme/sales positioning.
• /api/indices — index metadata (region, frequency, level window).
• /api/index-levels — {index_id: [{date, level}, ...]} monthly. Source for all correlations.
• /api/allocation/opportunity-sets — taxonomy mapping opportunity_set → asset_class (Equities/Duration/Credit/Currency) + display_order.
• /api/allocation/prior-views — records {quarter, previous_quarter, opportunity_set, view, conviction}. Holds the PRIOR-period view (see §6 for the exact selection rule).
• /api/macro-signals — {quarter, opportunity_set, score, rationale_code, drivers}. Source for new allocation views.

2. Policy reference (environment facts, re-fetch to confirm)

• credit_default / credit_risk_reduction (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, target_hy_reduction_pct (0 for credit_default; 4.0 for credit_risk_reduction — a required percentage-point reduction).
• correlation (POL_CORRELATION_DEFAULT): correlation_high_threshold: 0.8, correlation_low_threshold: 0.2, window start/end align with index level window.
• allocation_mapping (POL_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.
• multi_asset (POL_MULTI_ASSET_DEFAULT) uses_allocation_mapping/uses_correlation_default/uses_credit_default: true.
• multi_asset_risk (POL_MULTI_ASSET_RISK) uses_correlation_default/uses_credit_risk_reduction: true, escalates on two_or_more_material_exceptions.

3. Precision & ordering conventions (universal)

• Trade lists: sort as template states. Energy trade package = ascending instrument_id. Lumen rotation trades = SELL before BUY, then ascending instrument_id within each action.
• notional_usd_m / quantity_usd_m: 1 decimal (e.g. 4.0, not 4).
• total_market_value_usd_m, hy_allocation_pct, weighted_modified_duration_years, weighted_yield_to_maturity_pct, post_trade_hy_allocation_pct, post_trade_duration_years, hy_reduction_pct_points: 2 decimals.
• post_trade_watchlist_exposure_usd_m: 1 decimal.
• Correlation, signal_score: 3 decimals.
• as_of_date: use the portfolio summary as_of_date (currently 2026-05-29) unless the task says otherwise.
• Pair ids: alphabetical within the pair; sleeve/lists ordered ascending unless an item_order is given.

WORKFLOW A — Energy / fixed-income credit trade strategy

Covers: BUY/SELL bond tickets under credit constraints; post-trade metrics; sales positioning; data-precedence (train_001 BUY-only energy income; train_004 HY/watchlist risk-reduction rotation).

A1. Load current state (order matters)

1. GET /api/portfolios/<portfolio_id> → holdings (instrument_id → quantity_usd_m), MV, constraints.
2. GET /api/instruments/bonds → join each holding to its bond master (rating_bucket, modified_duration_years, yield_to_maturity_pct, subsector, issuer_id, energy_linked).
3. GET /api/issuers → join issuer_id → watchlist, credit_outlook.
4. GET /api/market/energy → signal scores + pitch themes (for energy-desk tasks).
5. GET /api/policies → the portfolio's constraint_policy_id sub-policy.

Re-read holdings from the portfolio service even if a stale snapshot / exception board is in the payload — those boards routinely disagree by a few $m (e.g. a watchlist HY bond shown as 10.0 on the stale board is 12.0 in the service). The service wins.

A2. Post-trade metric formulas (market-value-weighted)

For the post-trade book (existing holdings ± trades):

• total_market_value_usd_m = Σ quantity_usd_m (after trades).
• hy_allocation_pct = (Σ quantity where rating_bucket=="HY") / total_market_value_usd_m * 100.
• weighted_modified_duration_years = Σ(quantity * modified_duration_years) / total_market_value_usd_m.
• weighted_yield_to_maturity_pct = Σ(quantity * yield_to_maturity_pct) / total_market_value_usd_m.

Weight by quantity (=$m par/market value here; the service gives one figure). Round outputs to 2 dp. For rotation tasks: hy_reduction_pct_points = pre_trade_hy_pct − post_trade_hy_pct (2 dp); post_trade_watchlist_exposure_usd_m = Σ quantity of holdings whose issuer is watchlist (1 dp).

A3. Constraint pass/fail

• hy_cap_pass: post-trade hy_allocation_pct <= max_hy_allocation_pct (20.0).
• duration_band_pass: duration_band_years[0] <= weighted_modified_duration_years <= [1] (3.0–5.0).
• selected_issuer_diversification_pass: the SELECTED (new) tickets' issuers are distinct. (Issuer concentration limit 12% also exists; check a selected buy isn't an existing-issuer overload.)
• selected_subsector_diversification_pass: the selected tickets span >= subsector_min_count_for_diversified (2) distinct subsectors.
• watchlist_avoidance_pass: no selected BUY whose issuer watchlist==true. For rotation tasks also report watchlist_exposure_cleared (post-trade watchlist $m == 0) and buys_avoid_watchlist.
• target_hy_reduction_met (risk-reduction policy only): hy_reduction_pct_points >= target_hy_reduction_pct (4.0).

A4. BUY-only energy selection (train_001-style)

Inputs: exactly N BUY tickets (usually 2) for a fixed total notional, split evenly; energy-linked; improve carry; keep inside constraints; client-facing income pitch.

Selection filter for candidates:

1. candidate==true AND energy_linked==true AND not already a holding.
2. Watchlist filter: issuer watchlist==false (credit committee is sensitive to headline watchlist carry — drop Driftwood/Pacific Refining/Juniper even though their YTM is highest).
3. Duration-band filter on the post-trade book, but a good heuristic is to drop candidates whose own modified_duration_years is outside [3.0, 5.0] — long-dated LNG/oil (dur ~5.8–6.7) and very short HY (dur <3.0) are duration-ineligible distractors; including them risks breaching the band. Prefer candidates whose own duration sits inside the band.
4. Prefer the desk's preferred themes (LNG exporters / gas demand): match recommended_theme_tags (LNG_EXPORTS, GAS_DEMAND) and the strongest market/energy signal (LNG export pull is the most positive energy signal).
5. Improve carry: among eligible, prefer higher YTM. One IG anchor (LNG/gas, IG) + one non-watchlist HY carry bond is the canonical "improve carry while staying diversified" pair; verify the HY addition keeps hy_allocation_pct well under the 20% cap and that the two picks are distinct issuers and distinct subsectors.

Output sales_positioning: target_segment from the request's client_context (multi_asset_income etc.); theme from the energy pitch themes — pick the one matching the anchor (e.g. lng_export_tailwind when the IG anchor is an LNG exporter). data_precedence = current_environment_over_stale_payload whenever a stale worksheet mark differs from the service (MV, HY%, duration, or quantity) — which is the normal case.

A5. HY/watchlist risk-reduction rotation (train_004-style)

Goal: lower HY, remove avoidable watchlist risk, keep duration in the CIO band, fund IG candidates.

1. Identify watchlist holdings (issuer watchlist==true) — sell them (fully) to clear watchlist.
2. The portfolio is usually far over the 20% HY cap. Compute pre-trade HY%. To clear the cap with MV held roughly constant, sell at least (pre_hy_pct − 20)% * total_mv of HY. Selling only the watchlist bond may not reach the cap — sell additional non-watchlist HY "pressure points" too.
3. Keep at least the higher-rated HY for carry if "preserve carry" is an objective; don't strip all HY unless the desk asks.
4. Fund IG candidates from the desk shortlist / ?candidate=true IG universe. Exclude any candidate whose issuer is watchlist (avoid_new_watchlist_buy). Verify each buy's own duration sits inside [3.0, 5.0]; spread buys across distinct issuers (issuer conc 12%).
5. Preserve duration: selling short-duration HY (dur ~2.8–3.4) raises portfolio duration; buying longer IG (dur ~4.0–4.8) raises it further. Re-check weighted_modified_duration_years <= 5.0. If it would breach 5.0, tilt buys toward shorter IG / sell some longer IG.
6. risk_note_code: pick the dominant story — watchlist_concentration (watchlist was the issue), hy_cap_pressure (HY still over cap after rotation), duration_preservation (rotation held duration), carry_tradeoff (gave up carry to reduce risk), no_action.

Pitfalls: the stale exception board's quantities are NOT the service quantities — re-fetch. The desk shortlist often includes a watchlist HY bond (high carry, "risk team concerned about issuer status") as a trap; exclude it. Duration-ineligible IG candidates (dur >5.0) appear in shortlists as "duration ballast" but breach the band — check, don't trust the label.

WORKFLOW B — International equity correlation review

Covers: pair correlations across an index universe, highest/lowest pairs, China/Asia dependence flags, diversification candidates, sleeve actions (train_002; also the correlation half of train_005).

B1. Compute correlations

1. GET /api/index-levels (and/or /api/index-levels/<id>). Each index has monthly {date, level}.
2. For each index, sort by date, compute simple monthly returns r_t = level_t / level_{t-1} − 1 across the review window (level_start_date … level_end_date). 12 monthly levels ⇒ 11 returns.
3. Pearson correlation per pair: cov(r_i, r_j) / (std(r_i)*std(r_j)) over the shared return dates.
4. Round to 3 decimals. return_observations = number of returns (e.g. 11).

Use the window declared in the request payload's review_window (start/end dates) — these match the index metadata level_start_date/level_end_date.

B2. Extremes & pairs

• highest_positive: the pair with the maximum correlation. lowest: the pair with the minimum correlation (can be negative). pair_id / pair = the two index ids sorted alphabetically.
• For combined B+C tasks (train_005): correlation_summary has exactly two items in order [highest_concentration, best_diversifier]. highest_concentration = the highest absolute correlation pair (the pair that moves together most = concentration risk). best_diversifier = the lowest (most negative) correlation pair (best diversification). Both pairs sorted alphabetically internally.

B3. Concentration flags (POL_CORRELATION_DEFAULT)

• high_threshold_breached: any pair correlation >= correlation_high_threshold (0.8). Usually true for regional equity universes — many pairs exceed 0.8.
• china_asia_dependence_flag: set when the dedicated China sleeve and the Asia-Pacific-ex-Japan sleeve are both highly correlated with EM/World (Asia beta overlap). Check IDX_CHINA and IDX_AC_ASIA_PAC_EX_JP correlations against IDX_EM / IDX_WORLD — if multiple exceed 0.8, flag true.
• primary_code:
  • CHINA_ASIA_DEPENDENCE when China/AsiaPac/EM inter-correlations dominate the high pairs.
  • GLOBAL_DEVELOPED_OVERLAP when EAFE/World/ACWI developed-overlap pairs dominate instead.
  • NO_MATERIAL_CONCENTRATION only when no pair breaches 0.8.

B4. Diversification candidates & sleeve actions

• diversification_candidates: indices with low (ideally negative) correlation to the concentrated China/Asia cluster. Typical allowed set: IDX_EM_EX_CHINA, IDX_INDIA, IDX_LATAM — list all that qualify, ascending alphabetical. LatAm's negative correlation to China/Asia makes it the strongest diversifier.
• sleeve_actions (length per template, ascending by sleeve): the two most actionable moves. Typical pattern: trim/rotate the concentrated China sleeve toward IDX_EM_EX_CHINA (keep EM beta, drop single-country concentration), and add the best diversifier (IDX_LATAM). action ∈ {trim, add, hold, hedge, monitor, rotate}; target_index_id ∈ the sleeve's allowed set.
• For combined B+C tasks: target_sleeve_actions lists one action per requested sleeve (e.g. EM, India, LatAm, USD) in the template's item_order. Map each sleeve's allocation view to an action: UW+DOWN → trim; OW+UP → add; OW+UNCHANGED-high-conviction → add or hold; N → hold/monitor; a currency sleeve used as defensive offset → hedge.

B5. Correlation pitfalls

• Use simple returns (level ratio − 1), NOT log returns, unless the template says otherwise.
• Ensure both indices share the same monthly dates; align by date before correlating.
• Don't round returns before correlating — round only the final correlation.
• A negative "lowest" pair is normal; don't force it positive.

WORKFLOW C — Cross-asset active allocation view updates

Covers: active allocation views (view/change/conviction/rationale_code) per opportunity set, risk overlay, lineage, concentration flag, rebalance trigger, next-step enums (train_003; the allocation half of train_005).

C1. The view-derivation pipeline (this is the core algorithm)

For a refresh with target_quarter T and prior_quarter P (e.g. T=Q2_2026, P=Q1_2026):

1. New view from /api/macro-signals where quarter==T, using allocation_mapping thresholds:
   • score >= 0.35 → OW; score <= -0.35 → UW; else N.
2. Conviction from |score|:
   • |score| >= 0.7 → HIGH; >= 0.35 → MEDIUM; < 0.35 → LOW.
3. rationale_code = the macro signal's own rationale_code (e.g. EUROPE_RECOVERY, JAPAN_POLICY_RISK, CHINA_DEPENDENCE, INDIA_OFFSET, LATAM_DIVERSIFIER, DURATION_SUPPORT, HY_VALUATION_RISK, DOLLAR_DEFENSIVE, CREDIT_SPREAD_RISK, GROWTH_IMPROVES, RATE_CUT_SUPPORT, NEUTRAL_BALANCE).
4. prior_view from /api/allocation/prior-views. Selection rule (critical): pick the records where quarter == T (these carry previous_quarter == P); their view (and conviction) IS the prior-period (P) view. Do NOT look for a quarter == P batch — it does not exist. Do NOT treat the prior-views view as the new/target view; the new view always comes from the macro signal via step 1.
5. change = compare view_rank of new vs prior: UP if new>old, DOWN if new<old, else UNCHANGED. (view_rank: UW=-1, N=0, OW=1.)
6. asset_class from /api/allocation/opportunity-sets (Europe→Equities, U.S. Treasuries→Duration, Corporate High Yield→Credit, USD→Currency, etc.).

Sanity check: the prior-views batch where quarter == T reflects the incoming (P) view; a later batch where quarter == T+1 would reflect the just-derived (T) view and should match your signal-derived T views for most (not all) opportunity sets — committee judgment can override an occasional name, so trust your signal derivation for the T view, and trust prior-views for the P view.

C2. Output ordering & required rows

• allocation_views ordered by the request payload's focus_opportunity_sets order (NOT alphabetical, NOT display_order). Include exactly the requested opportunity sets (train_003 = 8; train_005 = 4).
• Each row: opportunity_set, asset_class, view, change, conviction, rationale_code.
• train_005 also wants prior_view and signal_score (raw macro score, 3 dp) per row.

C3. Risk overlay (train_003)

risk_overlay: overlay_code, primary_action, rationale_codes (list, business-priority order, highest priority first).

• Decide from the dominant risk in the view set. When HY is UW on HY_VALUATION_RISK and duration (U.S. Treasuries) is OW on DURATION_SUPPORT, the overlay is DURATION_QUALITY_TILT with primary_action: tilt_to_duration_quality — rotate out of credit beta into duration/quality. When credit spread risk dominates with no duration offset, use CREDIT_RISK_REDUCTION / trim_credit_beta. When currency defensiveness dominates (USD/CHF defensive), use CURRENCY_DEFENSIVE_HEDGE / add_currency_hedge. If views net to neutral, NO_OVERLAY / hold_policy_weights.
• rationale_codes priority order: lead with the most actionable risk-reduction code (the one driving the overlay), then supporting codes. Deduplicate.

policy_id (train_003) = POL_ALLOCATION_MAPPING (the policy whose thresholds produce the views). task_id = the template's required_value (e.g. train_003). as_of_date = portfolio/env as_of. target_quarter/prior_quarter = the request's quarters (template holds required_value).

C4. Combined correlation+allocation committee JSON (train_005-style)

Top-level: portfolio_id, as_of_date, review_quarter (the allocation quarter, e.g. Q2_2026), correlation_summary (§B2), target_sleeve_actions (§B4, in item_order), allocation_views (§C1, with prior_view+signal_score), rebalance_trigger, portfolio_risk_concentration_flag, next_step.

• rebalance_trigger: choose from {correlation_cap_breach, hy_cap_pressure, duration_drift, watchlist_concentration, committee_review}. When the concentration pair breaches the 0.8 correlation cap and drives the review, use correlation_cap_breach. When it's a routine committee refresh with no hard breach, committee_review.
• portfolio_risk_concentration_flag: true when the highest-concentration pair breaches the 0.8 threshold OR a dedicated China sleeve creates single-country beta — generally true for these sleeves.
• next_step: approve_rotation when a clear risk-reducing rotation is proposed and no hard constraint is breached; approve_with_monitoring when concentration is flagged but managed; defer_pending_risk_review when signals conflict materially; reject_constraint_breach only when a hard policy cap is violated by the proposal.

C5. Allocation pitfalls

• The prior-view selection rule (§C1.4) is the #1 error source. Re-read it. The view field in prior-views is the PRIOR (incoming) view for the refresh tagged by quarter==T, not the new T view.
• Don't copy conviction from prior-views — derive the NEW conviction from |score|.
• change is about view RANK (UW<N<OW), not score magnitude. OW→OW is UNCHANGED even if conviction rises.
• rationale_code must come from the macro signal, never invented. If two opportunity sets share a rationale code (e.g. Europe and EUR both EUROPE_RECOVERY), that's fine.
• signal_score in output = the raw macro score (3 dp), not the derived view.
• A stale local note (e.g. "USD overweight as defensive offset") describes a PRIOR posture, not the current view — refresh from signals; the USD score is often negative (defensive bid fading).

4. Reusable computation snippets

Pearson over monthly levels (Python):

def returns(levels):  # levels: list of (date, level) sorted by date
    r = []
    for i in range(1, len(levels)):
        r.append(levels[i][1] / levels[i-1][1] - 1.0)
    return r
def pearson(x, y):
    n = len(x); mx = sum(x)/n; my = sum(y)/n
    cov = sum((x[i]-mx)*(y[i]-my) for i in range(n))
    return cov / ((sum((xi-mx)**2 for xi in x)**0.5) * (sum((yi-my)**2 for yi in y)**0.5))

View derivation:

def view_of(s):  return "OW" if s>=0.35 else ("UW" if s<=-0.35 else "N")
def conv_of(s):
    a=abs(s); return "HIGH" if a>=0.7 else ("MEDIUM" if a>=0.35 else "LOW")
RANK={"UW":-1,"N":0,"OW":1}
def change_of(new,old):
    return "UP" if RANK[new]>RANK[old] else ("DOWN" if RANK[new]<RANK[old] else "UNCHANGED")
# prior_view: prior_views record where record["quarter"]==target_quarter (previous_quarter==prior)

Post-trade metrics:

tot = sum(q for q in holdings.values())
hy  = sum(q for id,q in holdings.items() if bond[id]["rating_bucket"]=="HY")
wdur= sum(q*bond[id]["modified_duration_years"] for id,q in holdings.items())/tot
wytm= sum(q*bond[id]["yield_to_maturity_pct"]  for id,q in holdings.items())/tot
hy_pct = hy/tot*100

5. Distillation checklist (run before emitting JSON)

• Did I fetch holdings from /api/portfolios/<id> (not the stale payload board)?
• Are watchlist issuers excluded from BUYS (and sold in rotations)?
• Are duration-ineligible candidates (own mod_dur outside [3,5]) excluded?
• Post-trade HY%, duration, YTM computed market-value-weighted and rounded to 2 dp?
• Correlations computed from monthly simple returns, Pearson, 3 dp, pairs alphabetical?
• Allocation views derived from macro signals via the 0.35 / 0.7 thresholds; prior_view from prior-views where quarter==target; change from view rank?
• Trade/list ordering matches the template (ascending instrument_id; SELL-before-BUY)?
• data_precedence reflects env-over-stale when any payload figure disagreed with the service?
• All enums pulled from the template's allowed_values, no typos?
• One JSON object only, no narrative outside it (unless the prompt asks otherwise)?