Asteria Investment Office JSON Desk

Asteria Investment Office — Desk JSON Workflow

You produce a single JSON object that conforms to the task's input/payloads/answer_template.json. The shared Asteria environment is a read-only HTTP API and the current book of record. Local payloads (desk requests, memos, snapshots, "stale notes") are intake context only and may be stale — when they disagree with the environment, the environment always wins.

0. Universal SOPs (apply to every task)

1. Environment is authoritative. Re-pull every number you use from the API. Never trust a quantity, rating, watchlist flag, duration, view, or score that only appears in the local payload. If a payload field conflicts with the env, use the env value and (where the schema asks) flag the precedence as "current_environment_over_stale_payload". If nothing conflicts, "no_conflict_found".
2. Read the answer_template first and obey it literally. It declares the exact top-level keys, required values (e.g. portfolio_id), enum value sets, numeric precision (decimal places), units, list length/required_length, and ordering rules. Mismatched enum spelling, wrong rounding, or wrong sort order are failures even when the analysis is right.
3. as_of_date / lineage. Use the environment's current as-of date for output as_of_date fields (do not copy the payload's stale date). At the time this skill was written the env-wide as-of date was the value returned by /api/policies (as_of_date) and matched each portfolio's as_of_date. Pull it live; do not hard-code.
4. Rounding. Round only at the end, to the precision the template states (e.g. correlations to 3 decimals, %/years to 2, notional to 1). Carry full precision through intermediate math.
5. id ordering. Where the template says "alphabetical", sort index/instrument ids as plain ascending strings. For index pairs, sort the two ids inside the pair alphabetically too. Trade lists usually sort by action then instrument_id; read the exact rule each time.
6. Use python (not python3) and curl. Save responses inside your own working directory. Compute correlations and weighted aggregates in python.
7. Candidate vs held filter. Bonds carry candidate (true = buyable new ticket) and you can see which ids are already in the portfolio's holdings. New BUY tickets normally come from candidate==true instruments not already held. Holdings you keep are not "trades".
8. Watchlist / eligibility exclusions are the most common trap. A high-yield "distractor" candidate is often on the issuer watchlist or duration-ineligible. Read the issuer's watchlist flag (/api/issuers) and the instrument's modified_duration_years/rating_bucket; exclude per the task's policy.

Endpoint map

Most list endpoints accept equality filters by field name (e.g. ?rating_bucket=HY, ?candidate=true, ?quarter=Q3_2026).

Policy thresholds (always re-pull /api/policies; do not hard-code values)

/api/policies returns nested blocks. Key sub-policies and the fields you use:

• allocation_mapping (POL_ALLOCATION_MAPPING):
  • view_score_thresholds: OW_min (e.g. 0.35), UW_max (e.g. -0.35), neutral band in between. Map a signal score → view: score >= OW_min → OW; score <= UW_max → UW; else N.
  • conviction_thresholds: by absolute score — HIGH_abs_min (e.g. 0.7), MEDIUM_abs_min (e.g. 0.35), below that → LOW.
  • view_rank: {UW:-1, N:0, OW:1}. Use it for the change field: new_rank > prior_rank → UP; < → DOWN; equal → UNCHANGED. (UP means more bullish, i.e. toward OW.)
• correlation (POL_CORRELATION_DEFAULT): correlation_high_threshold (e.g. 0.8), correlation_low_threshold (e.g. 0.2), and the official review_window_start / review_window_end.
• credit_default (POL_CREDIT_DEFAULT): duration_band_years (e.g. [3.0,5.0]), max_hy_allocation_pct (e.g. 20.0), issuer_concentration_limit_pct (e.g. 12.0), subsector_min_count_for_diversified (e.g. 2), target_hy_reduction_pct (0.0).
• credit_risk_reduction (POL_CREDIT_RISK_REDUCTION): same bands/caps but target_hy_reduction_pct is a positive number (e.g. 4.0) — used by the FI rebalance task.
• multi_asset / multi_asset_risk: composition policies that say which of the above they reuse (uses_allocation_mapping, uses_correlation_default, uses_credit_default / uses_credit_risk_reduction) and the committee escalation rule (committee_escalation_threshold, e.g. "two_or_more_material_exceptions").

A portfolio's own constraints.policy_id tells you which credit policy governs it (/api/portfolios/<id>). For policy_id output fields, prefer the policy id the portfolio/answer is governed by (e.g. the allocation-mapping policy id for an allocation refresh, or the portfolio's constraints.policy_id for a credit task). The top-level /api/policies policy_id (e.g. a "POLICY_SET_..." id) is the versioned bundle; use it only if the schema clearly wants the set id.

Core computations

Pearson correlation from monthly simple returns

1. Pull monthly level series for each index over the requested window (inclusive of both endpoints). N monthly levels → N-1 returns.
2. Simple return for month i: r_i = level_i / level_{i-1} - 1.
3. Pearson correlation of two return vectors x,y: corr = Σ(x-mean_x)(y-mean_y) / ( sqrt(Σ(x-mean_x)^2) * sqrt(Σ(y-mean_y)^2) ).
4. return_observations in the output = number of returns = N-1 (NOT N levels).
5. Round correlation to the template's precision (typically 3 decimals).
6. "Highest positive" pair = max corr; "lowest" / "best diversifier" = min corr (most negative). Within each pair, sort the two ids alphabetically.

def simple_returns(levels):           # levels: [{"date","level"}, ...] in date order
    return [levels[i]['level']/levels[i-1]['level']-1 for i in range(1,len(levels))]
def pearson(x,y):
    import math; n=len(x); mx=sum(x)/n; my=sum(y)/n
    cov=sum((a-mx)*(b-my) for a,b in zip(x,y))
    sx=math.sqrt(sum((a-mx)**2 for a in x)); sy=math.sqrt(sum((b-my)**2 for b in y))
    return cov/(sx*sy)

Portfolio credit aggregates (notional/market-value weighted)

For holdings list of (instrument_id, quantity_usd_m), joined to bond master:

• total_market_value = Σ quantity.
• hy_allocation_pct = 100 * Σ(quantity where rating_bucket=="HY") / total.
• weighted_modified_duration = Σ(quantity * modified_duration_years) / total.
• weighted_ytm = Σ(quantity * yield_to_maturity_pct) / total.
• issuer_concentration_pct(issuer) = 100 * Σ(quantity for that issuer) / total; compare the max to issuer_concentration_limit_pct.
• watchlist_exposure_usd_m = Σ(quantity where issuer.watchlist == true). Recompute these on the post-trade book (apply SELLs/BUYs to the live holdings).

Archetype (a) — Energy-credit recommendation / trade-package selection

Example: "Prepare the proposed energy-credit trade strategy for PF-EN-ALTA … two BUY tickets totaling USD 8.0m split evenly … improve carry while staying inside credit constraints … suitable for a client income pitch."

Endpoints: /api/portfolios/<id>, /api/instruments/bonds, /api/issuers, /api/market/energy, /api/policies (credit_default for this portfolio).

Selection recipe:

1. Build the eligible BUY universe: bonds with candidate==true, energy_linked==true (for an energy sleeve), issuer not on watchlist, and not already held.
2. Honour the ticket structure exactly (e.g. 2 tickets, USD 8.0m total split evenly → 4.0m each). Action enum is BUY/SELL/HOLD/NO_TRADE.
3. Post-trade constraint checks (apply the buys to live holdings, recompute):
   • hy_cap_pass: post-trade hy_allocation_pct <= max_hy_allocation_pct.
   • duration_band_pass: post-trade weighted modified duration within duration_band_years.
   • selected_issuer_diversification_pass: the selected tickets are from different issuers (i.e. don't double up one issuer within the package).
   • selected_subsector_diversification_pass: the selected tickets span >= subsector_min_count_for_diversified distinct subsectors.
   • watchlist_avoidance_pass: no selected ticket's issuer is on the watchlist. NOTE the word "selected": issuer/subsector diversification checks are about the chosen package, not the whole portfolio. A pre-existing large legacy holding may already exceed the portfolio-wide issuer-concentration limit; that is a pre-existing condition and is not what selected_issuer_diversification_pass measures.
4. Choosing among feasible packages: prefer the package that improves expected carry (higher post-trade weighted YTM / spread) and is defensible for a client income pitch. Use /api/market/energy signals: the strongest positive commodity signal and pitch_themes should steer theme/segment. Watch the tension: HY-heavy packages push carry up but bump against the HY cap and the committee's stated sensitivity to "headline carry from watchlisted issuers"; IG LNG/midstream names give a cleaner, lower-HY income story. Pick the package that maximises carry subject to all constraints and the client/theme guidance rather than raw yield alone.

Output schema conventions:

• trade_package: list of length 2, sorted ascending by instrument_id, each {action, instrument_id, notional_usd_m}; notional_usd_m precision 1.
• post_trade_metrics: total_market_value_usd_m, hy_allocation_pct, weighted_modified_duration_years, weighted_yield_to_maturity_pct — all precision 2.
• constraint_checks: the five booleans above.
• sales_positioning: target_segment ∈ {insurance_general_account, pension_liability_matching, multi_asset_income, private_bank_income, endowment_opportunistic}; theme ∈ {lng_export_tailwind, oil_oversupply_caution, midstream_stability, transition_bond_selectivity, avoid_watchlist_yield_trap}. Pick the theme that matches the dominant energy signal and the client context in the request (e.g. a multi-asset income client + strong LNG signal → multi_asset_income + lng_export_tailwind).
• data_precedence: ∈ {current_environment_over_stale_payload, local_payload_over_current_environment, no_conflict_found}. If the stale snapshot's MV/HY/duration differ from the live portfolio (they usually do), use current_environment_over_stale_payload.

Archetype (b) — International equity correlation review

Example: PF-INT-NEXVEN correlation review over a stated monthly-level window for a 9-index universe.

Endpoints: /api/portfolios/<id> (sleeves/holdings), /api/policies (correlation thresholds + official window), /api/indices (window dates), /api/index-levels (levels).

Recipe:

1. Window: use the request's review_window (reconcile against the policy's review_window_start/end and the indices' level_start_date/level_end_date; they should agree — env wins if not). Compute monthly simple returns → return_observations = N_levels - 1.
2. Compute the full pairwise Pearson correlation matrix over the requested index_universe. Identify highest_positive (max) and lowest (min, most negative) pairs.
3. Concentration logic: if the highly-correlated cluster is the China + Asia sleeves (China, Asia-Pac, EM are mutually >high_threshold), set china_asia_dependence_flag = true and primary_code = CHINA_ASIA_DEPENDENCE. high_threshold_breached = true if any relevant pair exceeds correlation_high_threshold. (Other codes: GLOBAL_DEVELOPED_OVERLAP if the dominant overlap is developed-world betas; NO_MATERIAL_CONCENTRATION if nothing breaches.)
4. Diversification candidates: from the allowed set (e.g. EM_EX_CHINA, INDIA, LATAM), the genuine diversifiers are those with low/negative correlation to the concentrated cluster. LATAM-type indices that are negatively correlated to the China/Asia cluster are the real diversifiers; EM-ex-China and India often remain >0.8 correlated to the core and are weaker diversifiers. Sort the chosen candidate ids alphabetically.
5. Sleeve actions: typically two — trim/hedge the concentrated sleeve (target the China index) and add the best diversifier (target the LATAM index). action ∈ {trim, add, hold, hedge, monitor, rotate}; target_index_id restricted to the template's allowed set.

Output schema conventions:

• index_set: alphabetical list of the universe ids.
• review_window: {level_start_date, level_end_date, return_observations}.
• extreme_pairs: {highest_positive:{pair_id:[a,b],correlation}, lowest:{...}}, pair ids alphabetical within the pair, correlation to 3 decimals.
• concentration: {china_asia_dependence_flag(bool), primary_code(enum), high_threshold_breached(bool)}.
• diversification_candidates: alphabetical subset of the allowed ids.
• sleeve_actions: list (length per template, e.g. 2), sorted by sleeve, each {sleeve, action, target_index_id}.

Archetype (c) — Active allocation view refresh

Example: Q2 2026 CIO active allocation refresh for a focus list of opportunity sets; output view / change / conviction / rationale per set plus a portfolio-level risk overlay.

Endpoints: /api/allocation/opportunity-sets (asset_class + display_order), /api/allocation/prior-views, /api/macro-signals, /api/policies (allocation_mapping).

Recipe (per opportunity set, for the target quarter):

1. Pull the macro signal for that opportunity_set and quarter==target_quarter: gives score, rationale_code, drivers.
2. view = map score through view_score_thresholds (OW/UW/N).
3. conviction = map abs(score) through conviction_thresholds (HIGH/MEDIUM/LOW).
4. prior_view = the standing view from /api/allocation/prior-views for that opportunity set whose quarter == target_quarter (its previous_quarter equals the prior quarter). That recorded view is the baseline you refresh from.
5. change = compare new view rank vs prior view rank using view_rank (UP/DOWN/UNCHANGED).
6. rationale_code = the macro signal's rationale_code (already from the allowed enum, e.g. EUROPE_RECOVERY, JAPAN_POLICY_RISK, CHINA_DEPENDENCE, INDIA_OFFSET, LATAM_DIVERSIFIER, DURATION_SUPPORT, HY_VALUATION_RISK, NEUTRAL_BALANCE, …).
7. asset_class = from the opportunity-set taxonomy (Equities/Duration/Credit/ Currency).

Risk overlay: synthesize from the dominant theme across the focus set:

• overlay_code ∈ {DURATION_QUALITY_TILT, CREDIT_RISK_REDUCTION, EQUITY_BETA_EXTENSION, CURRENCY_DEFENSIVE_HEDGE, NO_OVERLAY}.
• primary_action ∈ {tilt_to_duration_quality, trim_credit_beta, add_cyclical_equity_beta, add_currency_hedge, hold_policy_weights}.
• Choose the overlay that matches the strongest cross-set message: e.g. if duration/treasuries score strongly positive while HY scores strongly negative, that is a duration-quality tilt away from credit beta (DURATION_QUALITY_TILT + tilt_to_duration_quality). If HY risk dominates, it's CREDIT_RISK_REDUCTION + trim_credit_beta.
• rationale_codes: ordered list, highest business priority first, drawn from the rationale codes of the most decisive sets (e.g. DURATION_SUPPORT, HY_VALUATION_RISK, then supporting codes).

Output schema conventions:

• allocation_views: list ordered by the request's focus_opportunity_sets order (NOT alphabetical, NOT display_order — follow the payload), required length = number of focus sets. Each row: {opportunity_set, asset_class, view, change, conviction, rationale_code}, all enums exactly as allowed.
• Lineage: task_id (required_value from template), as_of_date (env current), target_quarter, prior_quarter, policy_id.
• risk_overlay: {overlay_code, primary_action, rationale_codes[]}.

Archetype (d) — Fixed-income risk rebalance

Example: PF-FI-LUMEN rotation that reduces HY and watchlist pressure while keeping duration in the CIO band; sell pressure points, fund eligible candidates.

Endpoints: /api/portfolios/<id>, /api/instruments/bonds, /api/issuers, /api/policies (the portfolio's constraints.policy_id, typically credit_risk_reduction with a positive target_hy_reduction_pct).

Recipe:

1. Compute pre-trade hy_allocation_pct, weighted duration, and watchlist_exposure_usd_m from live holdings.
2. SELLs: prioritise (i) watchlist HY holdings (these are the "avoidable watchlist risk" to remove), then (ii) additional non-watchlist HY as needed. You usually must sell more than just the watchlist name — removing one watchlist HY often still leaves HY above the cap, so keep selling HY until both the target_hy_reduction_pct reduction AND the max_hy_allocation_pct cap are satisfied.
3. BUYs: fund from candidate==true, non-watchlist, IG names not already held (e.g. data-center / materials / LNG IG). Keep the rotation roughly notional-neutral (Σ buys ≈ Σ sells) unless told otherwise, and choose buy durations that keep post-trade weighted duration inside duration_band_years.
4. Reject the watchlist HY distractor candidate. A shortlisted candidate that is HY and watchlist (e.g. a higher-coupon name flagged "issuer status concern") must NOT be bought when the memo says avoid new watchlist buys — buys_avoid_watchlist must stay true.
5. Recompute post-trade metrics and the boolean exception flags.

Output schema conventions:

• rotation.trades: sorted SELL before BUY, then ascending instrument_id within each action; each {action(BUY|SELL), instrument_id, quantity_usd_m}, quantity precision 1.
• risk_metrics: post_trade_hy_allocation_pct (prec 2), post_trade_duration_years (prec 2), hy_reduction_pct_points = pre_HY% − post_HY% (prec 2), post_trade_watchlist_exposure_usd_m (prec 1).
• exception_flags: hy_cap_pass, duration_band_pass, target_hy_reduction_met (reduction ≥ target_hy_reduction_pct), watchlist_exposure_cleared (post-trade watchlist exposure == 0).
• watchlist_handling: watchlist_sell_ids (ascending instrument_id list of the watchlist names you sold), buys_avoid_watchlist (bool).
• risk_note_code ∈ {hy_cap_pressure, watchlist_concentration, duration_preservation, carry_tradeoff, no_action} — pick the dominant theme of the rebalance (e.g. clearing HY over the cap → hy_cap_pressure, or removing a watchlist concentration → watchlist_concentration).

Archetype (e) — Multi-asset committee decision

Example: PF-MA-HELIO committee JSON linking non-US equity correlation findings to active allocation views for EM, India, Latin America, and USD.

This archetype combines (b) and (c) for one compact file.

Endpoints: /api/portfolios/<id>, /api/index-levels, /api/policies (multi_asset / multi_asset_risk → which reuses correlation_default and the allocation mapping), /api/allocation/prior-views, /api/macro-signals.

Recipe:

1. Correlation summary over the requested equity index ids and window (same Pearson-from-monthly-simple-returns recipe). Produce exactly two items in the fixed order [highest_concentration, best_diversifier]:
   • highest_concentration = the max-correlation pair (the concentration risk).
   • best_diversifier = the min-correlation (most negative) pair. Each item {pair_role, pair:[a,b] alphabetical, correlation (prec 3)}.
2. Allocation views for the requested opportunity sets (EM, India, Latin America, USD) using archetype (c)'s mapping. Output rows in the template's fixed item_order (e.g. Emerging Markets, India, Latin America, USD). Each row needs {opportunity_set, prior_view, signal_score (prec 3), view, change, conviction, rationale_code}. The stale local note often keeps USD overweight "as a defensive offset" — refresh it: if the current USD signal is in the neutral band, USD becomes N with change=DOWN from a prior OW. This is the canonical "environment overrides stale local note" case.
3. target_sleeve_actions for the same sets (same fixed order), action ∈ {trim, add, hold, hedge, monitor, rotate}. Tie actions to the views/correlation: trim the concentrated EM/China-driven sleeve, add the diversifier (LatAm), hold/monitor India if already OW, hedge/hold USD per its refreshed view.
4. rebalance_trigger ∈ {correlation_cap_breach, hy_cap_pressure, duration_drift, watchlist_concentration, committee_review}: if a correlation pair breaches correlation_high_threshold, correlation_cap_breach is apt; otherwise committee_review.
5. portfolio_risk_concentration_flag (bool): true if the equity book shows a material concentration (high-corr cluster breaching the threshold).
6. next_step ∈ {approve_rotation, defer_pending_risk_review, approve_with_monitoring, reject_constraint_breach}. Use the multi_asset_risk escalation rule: e.g. two or more material exceptions → escalate (defer/approve_with_monitoring) rather than approve outright.

Output schema conventions: top-level keys portfolio_id (required_value), as_of_date (env current), review_quarter (required_value), correlation_summary (len 2, fixed role order), target_sleeve_actions (fixed set order), allocation_views (fixed set order), rebalance_trigger, portfolio_risk_concentration_flag, next_step. No prose outside the JSON.

Common misjudgments to avoid (checklist)

• Using stale payload marks (MV, HY%, duration, prior view, USD overweight) instead of live env values. Always re-pull.
• Counting levels instead of returns for return_observations (it is N_levels − 1).
• Forgetting to sort pair ids / index sets / trade lists per the template's ordering rule, or sorting allocation rows alphabetically when the template says follow the request's focus order.
• Treating EM-ex-China or India as a strong diversifier when they remain

  0.8 correlated to the China/Asia core; the negatively-correlated index (LatAm-type) is the real diversifier.

• Buying a watchlist or already-held or non-candidate or duration-ineligible instrument as a "new ticket."
• Selling only the single watchlist name and assuming the HY cap is cleared — it often is not; sell enough HY to satisfy both the reduction target and the cap.
• Mis-mapping change: remember view_rank UW(-1) < N(0) < OW(1); UP = toward OW.
• Wrong rounding precision or emitting an enum value not in the template's allowed_values (exact spelling, case, and punctuation matter).
• Confusing the "selected" diversification checks (within the chosen package) with the portfolio-wide issuer-concentration limit.

Self-consistency cross-checks before emitting

• Re-add post-trade notional: Σ holdings ± trades should reconcile to the MV you report.
• Confirm every enum value you output appears verbatim in the template's allowed set; confirm every required top-level key is present.
• Confirm correlations are within [-1, 1] and that highest_positive ≥ every other pair and lowest ≤ every other pair.
• Confirm boolean flags actually match your computed numbers (e.g. hy_cap_pass ⇔ post-trade HY% ≤ cap; duration_band_pass ⇔ duration in band; watchlist_exposure_cleared ⇔ post-trade watchlist exposure == 0).
• Confirm list lengths and ordering match the template exactly.