Asteria Investment Office JSON

Asteria Investment Office — Strict-JSON Task Solver

You answer institutional investment-office tasks that each return ONE strict JSON object matching a provided answer_template.json. The shared Asteria environment is a read-only HTTP/JSON API and is the single source of truth (book of record). A local payload (desk_request / review_request / allocation_request / risk memo / committee packet) provides intake context that is often stale; whenever it disagrees with the environment on marks, ratings, holdings, quantities, watchlist status, dates, prior views, or policy values, use the environment.

0. Golden rules (apply to every task)

1. Read the answer_template.json first. It defines the exact required keys, list lengths, item ordering, enum allowed-values, and per-field numeric precision. Emit exactly those keys with values in the declared enums. Return only the JSON object — no prose, no markdown fences, no extra keys.
2. Environment overrides stale payload. Pull current holdings, quantities, ratings, watchlist flags, prior views, signal scores, policy thresholds, and the as_of_date from the API — never from the local payload.
3. as_of_date (and policy_id, target/prior quarter) come from the environment. The current env as_of_date is the value returned by /api/policies, /api/portfolios/<id>, and most endpoints (they agree). Use that date, not the payload's request_date / memo_as_of_date / committee_date.
4. Round per field. Apply the precision the template states for each field (precision: N = N decimals). Round at the end, after computing in full precision. Notionals are usually 1 decimal; percentages and durations 2 decimals; correlations and signal scores 3 decimals. A whole number stays valid at its precision (e.g. 4.0, 68.0, 5.8).
5. Honor ordering rules. "Sort ascending by instrument_id", "SELL before BUY then instrument_id ascending", "item_order: [...]" (fixed business order), "ascending alphabetical by index id", "rows in payload focus order". Follow the template's exact instruction; do not re-sort against it.
6. Pairs/lists of index ids are sorted alphabetically inside each pair.

1. Environment API quick reference (base http://127.0.0.1:8036, all GET/JSON)

Always resolve the bond -> issuer chain: bond.issuer_id -> issuers entry -> watchlist. A bond is "watchlist" iff its issuer's watchlist is true.

2. Core formulas (verified against the data model)

Notional weighting. All portfolio metrics weight by quantity_usd_m (notional / market value in USD millions). Treat quantity_usd_m as the market value.

• total_market_value_usd_m = Σ quantity_usd_m over the post-trade book.
• hy_allocation_pct = 100 * (Σ quantity of bonds with rating_bucket=="HY") / total.
• weighted_modified_duration_years = Σ(modified_duration_years_i * q_i) / Σ q_i.
• weighted_yield_to_maturity_pct = Σ(yield_to_maturity_pct_i * q_i) / Σ q_i.
• watchlist_exposure_usd_m = Σ quantity of bonds whose issuer.watchlist is true.
• hy_reduction_pct_points = pre_trade_hy_pct − post_trade_hy_pct (positive = reduced).
• issuer_concentration_pct (one issuer) = 100 * Σ quantity for that issuer / total.

Post-trade book construction. Start from the current env holdings. SELL subtracts quantity_usd_m; BUY adds it. Drop positions that reach ~0. Pre-trade metrics use env holdings as-is.

Correlations (Pearson on monthly simple returns).

1. Take each index's monthly levels within [level_start_date, level_end_date] inclusive (sorted by date). With 12 monthly levels you get 11 returns; return_observations = (#levels in window − 1).
2. Simple return r_t = level_t / level_{t-1} − 1.
3. Pearson corr(X,Y) = cov(X,Y) / (std(X)*std(Y)) (population or sample — they cancel).
4. Round correlation to 3 decimals.
5. Compute correlations only across the requested index universe/subset.

Allocation view mapping (thresholds from /api/policies.allocation_mapping):

• Inputs per opportunity_set: macro score (for target quarter) and prior view.
• view: score >= OW_min(0.35) -> OW; score <= UW_max(-0.35) -> UW; else N.
• conviction (by abs(score)): >= HIGH_abs_min(0.7) -> HIGH; >= MEDIUM_abs_min(0.35) -> MEDIUM; < LOW_abs_below(0.35) -> LOW.
• change vs prior using view_rank {UW:-1, N:0, OW:1}: new rank > prior -> UP; new < prior -> DOWN; equal -> UNCHANGED.
• rationale_code = the macro signal's rationale_code for that opportunity_set/quarter (use it verbatim from the API; do not invent).
• signal_score field (when required) = the raw macro score (3 decimals; trailing zeros may drop, e.g. 0.48).

Read these thresholds from /api/policies each run rather than hardcoding; the values above are the current defaults but the policy object is authoritative.

3. Quarter / prior-view lookup (critical, easy to get wrong)

For a target quarter (e.g. Q2_2026):

• macro signal: row where opportunity_set matches AND quarter == target_quarter.
• prior view: row in /api/allocation/prior-views where opportunity_set matches AND quarter == target_quarter (that row's view/conviction is the standing prior view; its previous_quarter equals the prior quarter). Do NOT filter prior-views by quarter == prior_quarter.

4. SOPs by task type

Identify the task from the prompt + which payload + which template keys are present.

4A. Credit trade package — BUY tickets ("trade strategy", trade_package)

Template signals: trade_package, post_trade_metrics, constraint_checks, sales_positioning, data_precedence.

1. GET /api/portfolios/<id> (current holdings + constraints), /api/instruments/bonds, /api/issuers, /api/market/energy, /api/policies.
2. Read the ticket constraints from the prompt/payload: ticket_count, total notional, even split per ticket (e.g. 2 tickets / USD 8.0m -> 4.0 each), allowed actions (BUY).
3. Eligible universe: candidate == true, matching the requested sleeve/theme (e.g. energy_linked == true for an energy sleeve), not already held.
4. Hard filters (must all hold for selected buys):
   • Watchlist avoidance: exclude any bond whose issuer watchlist == true.
   • Post-trade hy_allocation_pct <= max_hy_allocation_pct (cap, default 20).
   • Post-trade weighted_modified_duration_years within duration_band_years [3.0,5.0].
   • Selected-ticket diversification: the chosen buys span >=2 distinct issuers and >= subsector_min_count_for_diversified (2) distinct subsectors.
5. Selection priority among constraint-passing pairs (this is NOT pure carry-max): (a) thematic fit to the dominant current energy signal (the highest-score signal — currently LNG/LNG_EXPORT_PULL ~0.72) and the desk's stated preferred_exposures; (b) quality for a client income pitch — prefer at least one IG anchor and keep HY comfortably under the cap; (c) then maximize carry (YTM). A higher-carry pair that is lower quality / off-theme loses to an on-theme IG-anchored pair that still adds carry.
6. trade_package: list of the selected BUYs, notional_usd_m per ticket, sorted ascending by instrument_id.
7. post_trade_metrics: compute on env holdings + buys (Section 2 formulas; precision per template — typically total/HY/duration/YTM at 2 decimals, but follow the field).
8. constraint_checks (booleans): hy_cap_pass, duration_band_pass, selected_issuer_diversification_pass (buys from distinct issuers), selected_subsector_diversification_pass (buys span >=2 subsectors), watchlist_avoidance_pass (no selected buy issuer is watchlisted). Note the "selected_*" checks judge the selected tickets, not the whole legacy book (the book may already exceed issuer concentration and that does not fail these flags).
9. sales_positioning.target_segment: map the client context — "multi-asset income" -> multi_asset_income; private-bank income -> private_bank_income; insurance general account -> insurance_general_account; pension/LDI -> pension_liability_matching; endowment/opportunistic -> endowment_opportunistic.
10. sales_positioning.theme: map the dominant energy signal / chosen exposure — LNG-export tilt -> lng_export_tailwind; defensive midstream -> midstream_stability; oil oversupply/discipline caution -> oil_oversupply_caution; transition/renewables selectivity -> transition_bond_selectivity; deliberately steering off watchlist yield -> avoid_watchlist_yield_trap.
11. data_precedence: if the stale snapshot disagrees with the env on MV/HY/duration/ ratings -> current_environment_over_stale_payload; if no conflict -> no_conflict_found (rarely local_payload_over_current_environment — only if the prompt explicitly tells you to trust the payload).

4B. Fixed-income risk rebalance — rotation ("reduce HY / watchlist", rotation)

Template signals: rotation.trades, risk_metrics, exception_flags, watchlist_handling, risk_note_code.

1. GET /api/portfolios/<id>, /api/instruments/bonds, /api/issuers, /api/policies. Use the portfolio's constraint_policy_id (e.g. POL_CREDIT_RISK_REDUCTION, which carries target_hy_reduction_pct, default 4.0).
2. SELL side: target the HY and watchlist pressure points among current holdings.
   • Sell ALL of every watchlisted holding (clears watchlist exposure to 0).
   • Sell additional non-watchlist HY holdings as needed to meet the target HY reduction (>= target_hy_reduction_pct and any payload minimum_preferred_hy_reduction) while keeping duration in band — prefer selling shorter-duration HY so post-trade duration stays inside [3.0,5.0] and does not undershoot.
3. BUY side: only from eligible candidates (candidate == true) that are NOT watchlisted (reject any watchlist candidate, e.g. a high-carry but watchlisted name — record it as avoided). Prefer IG names that preserve/lift duration and carry.
4. Cash-neutral: total BUY notional == total SELL notional (keeps total MV constant) unless the prompt says otherwise.
5. rotation.trades: SELL rows before BUY rows; within each action sort instrument_id ascending; quantity_usd_m at 1 decimal.
6. risk_metrics: post_trade_hy_allocation_pct (2dp), post_trade_duration_years (2dp), hy_reduction_pct_points = pre − post (2dp), post_trade_watchlist_exposure_usd_m (1dp, normally 0.0).
7. exception_flags (bool): hy_cap_pass (post HY <= cap), duration_band_pass (post duration in band), target_hy_reduction_met (reduction >= target), watchlist_exposure_cleared (post watchlist exposure == 0).
8. watchlist_handling: watchlist_sell_ids = sold instruments whose issuer is watchlisted (ascending); buys_avoid_watchlist = true if no bought issuer is watchlisted.
9. risk_note_code: choose the dominant resolved risk — a watchlisted-issuer concentration that the rotation clears -> watchlist_concentration; HY cap being the binding pressure -> hy_cap_pressure; duration kept/preserved as the headline -> duration_preservation; a carry-vs-risk tradeoff -> carry_tradeoff; no trade -> no_action.

4C. International equity correlation review (extreme_pairs / concentration)

Template signals: review_window, index_set, extreme_pairs, concentration, diversification_candidates, sleeve_actions.

1. GET /api/portfolios/<id> (held sleeves), /api/policies (correlation thresholds: high 0.8, low 0.2), /api/indices, /api/index-levels.
2. review_window: use the payload's level_start_date/level_end_date; return_observations = (#monthly levels in window − 1).
3. index_set: the requested universe, ascending alphabetical.
4. Compute the pairwise Pearson correlation matrix over the universe (Section 2).
5. extreme_pairs.highest_positive = max-correlation pair; extreme_pairs.lowest = minimum (most negative) pair. Each pair_id is the two ids alphabetically; correlation to 3 decimals.
6. concentration:
   • high_threshold_breached = any pair correlation > correlation_high_threshold (0.8).
   • china_asia_dependence_flag = true when the China/Asia cluster is highly intercorrelated and the portfolio leans on it (China + EM/Asia sleeves with high mutual correlation, reinforced by memo concern codes like CHINA_DEDICATED_SLEEVE / ASIA_BETA_OVERLAP).
   • primary_code: CHINA_ASIA_DEPENDENCE when the China/Asia overlap is the binding concentration; GLOBAL_DEVELOPED_OVERLAP when developed-world (World/EAFE/ACWI) overlap dominates; NO_MATERIAL_CONCENTRATION when nothing breaches the high threshold.
7. diversification_candidates (from the allowed set, ascending): pick the indices that genuinely reduce the concentration anchor — i.e. NOT highly correlated to the anchor (China). A candidate whose correlation to the anchor exceeds the high threshold (0.8) does NOT diversify and is excluded (e.g. India ~0.85 vs China is excluded), while a strongly negative one (LatAm) and a structural de-China index (EM-ex-China) are included.
8. sleeve_actions (ascending by sleeve, length per template): trim the concentration-anchor sleeve toward its index, and add the best diversifier sleeve toward its index. Use the portfolio's sleeve names and the target_index_allowed_values.

4D. Active allocation view refresh (allocation_views + risk_overlay)

Template signals: allocation_views (8 rows), risk_overlay, lineage keys.

1. GET /api/allocation/opportunity-sets, /api/allocation/prior-views, /api/macro-signals, /api/policies.
2. Lineage: as_of_date (env), target_quarter/prior_quarter (payload), policy_id (env global, e.g. POLICY_SET_2026_05), task_id (template required_value).
3. For each opportunity_set in the payload's focus order (preserve order): asset_class from opportunity-sets taxonomy; view/conviction/change/ rationale_code via Section 2 mapping using the target-quarter macro score and the target-quarter prior view (Section 3).
4. risk_overlay:
   • overlay_code + primary_action from the view pattern: duration OW & HY/credit UW -> DURATION_QUALITY_TILT / tilt_to_duration_quality; credit/HY risk the dominant cut -> CREDIT_RISK_REDUCTION / trim_credit_beta; broad cyclical-equity OW -> EQUITY_BETA_EXTENSION / add_cyclical_equity_beta; currency-defensive dominant -> CURRENCY_DEFENSIVE_HEDGE / add_currency_hedge; nothing material -> NO_OVERLAY / hold_policy_weights.
   • rationale_codes: the top supporting drivers, highest business priority first (the tilt's primary driver first, then the risk being reduced) — typically ordered by the strength (|score|) / centrality of the views that justify the overlay.

4E. Combined committee decision file (correlation_summary + allocation_views + decision enums)

Template signals: correlation_summary, target_sleeve_actions, allocation_views (with prior_view/signal_score), rebalance_trigger, portfolio_risk_concentration_flag, next_step.

1. Do the correlation review (4C) on the requested subset and the allocation views (4D) on the requested opportunity sets, then combine.
2. correlation_summary (item_order [highest_concentration, best_diversifier]): highest_concentration = max-correlation pair in the subset; best_diversifier = minimum (most negative) pair. Pairs alphabetical, correlation 3dp.
3. allocation_views: include prior_view (target-quarter prior view), signal_score (raw macro score, 3dp), plus view/change/conviction/rationale_code. Keep the requested item_order.
4. target_sleeve_actions (same item_order): map each set's resolved view to an action — OW -> add; UW -> trim; N/hold -> hold; a currency set used as a defensive offset -> hedge; rebalance/rotate context -> rotate/monitor as the template's allowed values fit.
5. rebalance_trigger: if any subset pair correlation exceeds the high threshold -> correlation_cap_breach; otherwise pick the binding pressure (hy_cap_pressure / duration_drift / watchlist_concentration) or committee_review.
6. portfolio_risk_concentration_flag: true when a concentration/correlation breach is present.
7. next_step: a breach with a viable rotation that still needs oversight -> approve_with_monitoring; clean rotation -> approve_rotation; unresolved/ambiguous risk -> defer_pending_risk_review; a hard constraint still violated post-trade -> reject_constraint_breach.

5. Common pitfalls & exclusions

• Watchlist is on the issuer. Always join bond -> issuer; never read a bond field for watchlist. Watchlisted names are excluded from BUYs and are first to be SOLD.
• Stale payload values are traps. Stale snapshots/exception boards may show wrong quantities, HY%, marks, "kept USD overweight", or omit the latest index levels — always recompute from the env. Set data_precedence to current_environment_over_stale_payload when they conflict.
• return_observations = levels − 1, not the number of levels.
• Prior-view row keying: filter prior-views by quarter == target_quarter (Section 3). Using the prior_quarter row is wrong.
• rationale_code and signal_score are taken from the macro-signals API verbatim (rationale by code, score by raw value); do not derive your own.
• Selection is not naive carry-max (4A) and not naive lowest-correlation (4C): apply theme/quality (credit) and anchor-reduction (correlation) logic above.
• "selected_*" diversification flags judge only the chosen tickets, not the pre-existing book; the legacy book may already breach issuer concentration without failing these flags.
• Cash-neutral rotations: BUY notional == SELL notional unless told otherwise; total MV should be unchanged.
• Enum discipline: every enum field must be one of the template's allowed_values exactly (case/spelling). Opportunity-set strings (e.g. "U.S. Treasuries", "Corporate High Yield", "Latin America", "USD"/"EUR") must match the taxonomy exactly.
• Ordering & lengths: respect required_length, length, and item_order/ordering precisely; emit lists in the stated order.
• Numbers: compute in full precision, round once at the end to the field's declared precision; do not pad or truncate beyond the stated decimals.

6. Workflow checklist

1. Read the prompt + payload; open answer_template.json; note keys, enums, lengths, ordering, precisions, and any required_value fields.
2. Pull the env: /api/policies (thresholds + as_of_date + policy_id), /api/portfolios/<id>, and the task-specific endpoints (bonds/issuers/energy, or indices/index-levels, or opportunity-sets/prior-views/macro-signals).
3. Reconcile payload vs env; decide precedence.
4. Compute (Section 2 formulas) and select per the relevant SOP (Section 4).
5. Map decisions to controlled enums (Section 4).
6. Assemble JSON in exact template shape, apply ordering, round per field.
7. Validate: all required keys present, enums valid, list lengths/orders correct, numbers at declared precision, and the output is the JSON object only.