Корень «−1.00 везде» (эпик #1953): compute_demand_supply_forecast брал
district-wide unit_velocity (847.5/мес, ВСЕ классы/комнаты) как спрос и
весь district-сток (~63k доступных) как предложение для КАЖДОЙ ячейки
what_to_build → один и тот же ratio во всех ячейках → все deficit_index
прижаты к −1.0. Плюс objective_lots — append-per-snapshot (~2.9× инфляция
строк), что симметрично раздувало обе базы → даже сегментация без дедупа
осталась бы вырожденной.
Фикс (blast radius — ТОЛЬКО forecast/deficit calc; platform-wide dedup = #1964):
- market_metrics.compute_market_metrics: +obj_class/+room_bucket (+cache key).
_STOCK_SQL и _SALES_WINDOW_SQL дедуплят до ПОСЛЕДНЕГО снапшота на физлот
(DISTINCT ON project_name,corpus_name,section,floor,lot_number ORDER BY …
snapshot_date DESC,id DESC), затем агрегируют. Class-фильтр (LOWER=LOWER,
class lowercase) + room-bucket (Source-B room_area-вокабуляр, зеркало
sales_series.room_area_bucket_of → what_to_build фильтрует без перевода).
ROLLUP/GROUPING сохранён; confidence считается на дедуплицированных counts.
- demand_supply_forecast: base_pace и open-сток теперь ПОСЕГМЕНТНЫЕ
(market_metrics(obj_class,room_bucket)). При заданном сегменте L2/L3
(hidden/future) ИСКЛЮЧЕНЫ из баланса — они класс/формат-агностичны, иначе
двоились бы по всем ячейкам. +_market_room_bucket VOCAB-мост (валидирующий
pass-through Source-B меток; неизвестное → None = без фильтра, не тихий 0-rows).
- what_to_build/_DEFAULT_CLASSES и recommendation Economy-маппинг: «эконом»→
«стандарт» (в objective_lots эконома НЕТ, стандарт=483k → раньше ячейка
матчила 0 строк и молча выпадала).
- report_assembler honesty-guard: если ВСЯ сетка прижата к ±1.0
(degenerate-fallback) — не эмитим «строить»/«избегать», показываем
«недостаточно гранулярных данных для посегментного вывода».
- data/sql/173_objective_lots_physflat_idx.sql: partial index под DISTINCT ON
(Index Only Scan + Unique, без Sort на 1.75M строк; idempotent, BEGIN/COMMIT).
Prod-verify (parcel 66:41:0205010:287, Железнодорожный, h=24): ячейки
ДИФФЕРЕНЦИРУЮТ (12 measured, 7 distinct) вместо all −1.0; MOI комфорт/студия
38.5 vs стандарт/студия 244.3 (точное совпадение с ожидаемым).
Тесты: регрессия «ячейки различаются (не all −1.0)» + vocab-translation +
honesty-guard + посегментное предложение. ruff clean; no :name::type.
Replace dev-jargon with plain RU across the §22 forecast report (6.2/6.3/6.5
+ deficit/затоварка legend). Source of truth = backend reason strings + the
frontend RU maps; text/render-only, no scoring/forecast math touched.
Backend (source of truth):
- scenarios.py _COLLAPSE_REASON_LOW_BETA: «β rate-sensitivity не прошёл gate …»
→ «чувствительность к ставке не оценена на коротком ряде ЕКБ → один базовый
сценарий вместо трёх».
- confidence_engine.py _coverage_factor: drop «domrf↔objective» jargon, say it
affects будущее предложение/конкуренцию. New _history_factor: «глубина истории
N мес» + на что влияет + связь с 6.2 (короткий ряд → один сценарий).
Frontend (both Section-6 families — live analysis page + ptica cockpit):
- Deficit legend: −1 затоварка / 0 баланс / +1 острый дефицит + actionable
трактовка; MOI tied to «сколько месяцев район распродаёт предложение».
- 6.2 heading «Почему один сценарий, а не три» over the collapse reason.
- 6.3 render confidence.rationale + weakest-link rule («скорее завысим
недоверие, чем недооценим риск»); FACTOR_RU gains confounded_window/
advisory_cap; factor notes shown.
- 6.5 «Вес»→«Оценка»; overall verdict vs 0.5; «Риск избытка предложения»→
«Запас по предложению»; §-refs moved from reason into tooltip; 6 special-
index 1-line «что это + куда лучше» descriptions; 0.00-score reasons shown.
Tests: confidence_engine (history/coverage notes), stripSectionRefs vitest.
Refs #1963
#1954 — площадь «—»: COALESCE(land_record_area, specified_area, declared_area)
в EGRN-блоке analyze (land_record_area NULL у 12809/42233 участков, но
specified_area заполнена). area_m2 для 66:41:0205010:287 теперь 106378 (был NULL).
#1954 — «Обновлено» сломано: cost_registration_date — мёртвая колонка (0/42233);
репойнт на updated_at (42233/42233 заполнено). Ключ ответа last_egrn_update_date
не меняется (additive value fix).
#1958 — confidence-фактор «Прогноз спрос/предложение» дублировался ×4
(по фактору на горизонт). Сворачиваем в один weakest-link'ом (MIN ранга
по горизонтам) в _component_confidences до confidence-движка (#990).
#1957 — ЗОУИТ backend:
- _get_cad_zouit_overlaps: DISTINCT ON (reg_numb_border) — дедуп дубль-строк
(одна физ. зона 2× с разным category_name). group_key cad_zouit→protected.
- _get_zouit_overlaps (dump path): subcategory→RU-тип карта (26→СЗЗ и др.,
коды сверены кросс-джойном dump↔cad_zouit на проде); type_zone из карты,
reg из props.options.reg_numb_border. Раньше отдавал blank-строки.
- унификация group_key (protected/engineering/okn/natural/other) + top-level
reg_numb_border в обоих путях.
UP038-модернизация isinstance в report_assembler (pre-commit ruff 0.7.4).
Frontend note (#1957): nspd_zouit_overlaps теперь всегда group_key из набора
{protected,engineering,okn,natural,other} — сырой 'cad_zouit' больше не отдаётся;
оба пути несут type_zone + reg_numb_border.
Tests: +4 _component_confidences collapse, +6 ЗОУИТ (subcategory map, dump
typing, DISTINCT ON), schema-test обновлён на protected. 451 passed targeted.
Когда demand_normalization honesty-gate отбивает β rate-sensitivity
(`sensitivity.confidence=='low'`, ЕКБ-регрессия не проходит n≥30/R²≥0.1/slope<0),
все 393 прод-отчёта получают coefficient=1.0 для всех трёх сценариев
conservative/base/aggressive. По дизайну backend честно деградирует, но фронт
рисует 3 разноцветные карточки/линии одинаковых чисел без caveat.
Backend (scenarios.py, report.py, report_assembler.py, orchestrator.py):
- compute_scenarios теперь возвращает (list, collapsed: bool, reason: str|None)
- _detect_collapsed(): math.isclose(rel_tol=1e-9, abs_tol=1e-6) сравнивает
projected_demand_units И deficit_index на всех горизонтах между
conservative и aggressive — расхождение в любой метрике на любом h → False
- _COLLAPSE_REASON_LOW_BETA — единственный источник истины каноничного
русского предложения
- ReportScenarios получает поля scenarios_collapsed + scenarios_collapse_reason
- demand_normalization.py НЕ ТРОНУТ — поведение корректно по контракту
Backend экспортёры (report_pdf.py, excel.py):
- PDF: при collapse — тёмная плашка-параграф вместо таблицы трёх столбцов
- Excel: ОДНА строка «base» + caveat-ячейка вместо трёх идентичных столбцов
Frontend (forecast.ts, ScenariosBlock, ScenarioCards, ScenarioCompareTable,
ForecastChart, Section6Forecast, ptica.module.css):
- Тип ReportScenarios расширен двумя optional-полями
- ScenariosBlock (light): headline-bar + одна grid-карточка base + caveat
- ScenarioCards (ptica dark): sub-component ScenarioBaseCard, одна карточка
+ collapse-note, ScenarioCompareTable return null
- ForecastChart: effectiveScenarios filter — одна линия viz-1 «Базовый»
вместо трёх перекрывающихся
- Section6Forecast: caveat выше графика, читается ТОЛЬКО из
scenarios_collapse_reason (источник истины — backend constant)
Тесты: +18 новых (TestMetricsEqual×6, TestDetectCollapsed×7,
TestComputeScenariosCollapseDetection×4 + 1 проверка graceful + corner).
102 forecasting passed, 88 cross-module passed.
Refs #1871
- #1736 MiniMap: проброс useConnectionPoints → точки подключения на карте analyze (были только в /legacy)
- #1737 confidence: пронесено имя сервиса → RU-ярлык (Рынок/Будущее предложение/…) вместо «Компонент вкладывающий сервис»
- #1738 pipeline: self-exclusion субъекта (ST_DWithin 80м) — проект не считает сам себя будущим конкурентом
- #1739 PDF: snapshot_pdf обёрнут в try/except+logger.exception (причина 500 видна) + format=pdf в forecast export + font_url fallback
- #1740 gate↔recommendation: при can_build_mkd=False — gate_caveat на обоих рекомендаторах (противоречие явное, не молчит)
Verify: py_compile 5/5, tsc 0, ruff clean, pytest confidence/forecast 95 passed.
Closes#1736Closes#1737Closes#1738Closes#1739Closes#1740
- Add window_months > 0 guard to vel_by_room dict comprehension (mirrors _monthly_rate)
- Correct outer comment in recommendation.py: honest-zero for known-zero buckets,
fallback only when velocity_by_room=None or bucket absent from _FORECAST_TO_METRIC_BUCKETS
- Add comment in as_dict() noting velocity_by_room is intentionally not serialized
(internal pipeline attr consumed directly by recommendation.py)
`_build_confidence` was calling `compute_report_confidence` without
`deal_count_months`, so the «за N мес» suffix never appeared in
production. Add `_deal_count_months(market_metrics)` extractor
(reads `window_months`, same key as `_history_months`) and pass it.
Also fix pre-existing UP038 violations (isinstance tuple → X | Y).
Add `velocity_by_room: dict[str, float] | None` to `MarketMetrics` — per-bucket
unit velocity (ед./мес) derived from the existing `sold_by_room` ROLLUP data that
`_query_sales_window` already returns. No new SQL required.
Thread per-bucket velocity through `_demand_only_overlay` via the new
`_FORECAST_TO_METRIC_BUCKETS` constant that maps each forecast bucket to its
market_metrics room-bucket keys. "80+ м²" sums "4" + "5+" keys. Fallback to
aggregate `unit_velocity` when `velocity_by_room` is None (thin-data path).
Previously `base_pace` was identical for all 5 room-buckets, so §9.4 norm and §9.2
base_pace cancelled out in pace/max_pace and ranking was driven purely by §9.5
macro_coef (segment steepness proxy). Now §9.2 reflects real per-bucket observed
demand from objective_lots.contract_date data.
Callers of `compute_market_metrics` that don't use `velocity_by_room` are unaffected
(the new field is additive to the frozen dataclass). All existing callers verified —
none construct `MarketMetrics` directly except the one production site.
_GEO_WEIGHT_UNKNOWN was 0.1, which equals exp(−6.9/3)≈0.10 (weight of a
confirmed-far project at ~6.9 km). Projects beyond that distance got a
weight *below* 0.1, meaning unknown-coordinate projects outweighed
confirmed-far ones — an inversion of the documented intent.
Lowered to 0.05 (≈ exp(−3) = exp(−9 km / scale)), restoring the correct
hierarchy: confirmed-close > confirmed-far > unknown. Updated TestGeoWeight
(hardcoded 0.1 expectation) and TestCannibalizationTrueMode (overlap ×
floor comment/value) accordingly. Added two new assertions in TestGeoWeight
that enforce the hierarchy monotonically and verify unknown < exp(−6.9/3).
_count_full_years treated units=0 as a valid observation, so a series
where fill_month_grid zero-filled every month still accumulated 3 full
years and passed the _MIN_FULL_YEARS guard. Zero-filled months carry no
seasonal signal, so they must be skipped in the year counter — the same
way None values already were.
Fix: skip v==0 alongside v is None in _count_full_years.
Add four tests: zero-filled 36-month series → n_full_years=0/applied=False;
partial-coverage years (only 6 non-zero months/year) → not counted as full;
real non-zero series still passes guard; normalize_demand on zero-filled
SalesSeries returns series unchanged.
Add `deal_count_months: int | None = None` to `compute_report_confidence`.
When provided, threads it as suffix into `_factor_from_count` so the
deal_count ConfidenceFactor note reads «7 сделок за 6 мес — мало» instead
of the windowless «7 сделок — мало». Existing callers unaffected (default None).
Tests: two new cases in TestComputeReportConfidence — with/without window.
Внутренний recommendation→product_scoring контракт-ключ был мислейблом: величина —
темп распродажи нежилого (sell-through, прокси ликвидности/спроса), а НЕ доля нежилого
в объёме застройки. Переименован ключ + исправлены reason/docstring/комментарии у
потребителя _score_commercial. Числовая логика не изменена. Ключ внутренний (нет
frontend/schema/openapi-потребителей) → rename контракт-безопасен. pytest 171 passed.
Closes#1635
DemandSupplyForecast.as_dict() не эмитил 'confounded'/'is_confounded_window',
report_assembler._confounded() всегда возвращал False и §15 confounded_window
factor в compute_report_confidence был мёртв: 48-мес окна, пересекающие
2024-07-01 шок никогда не тянули report confidence к 'low' и шок не назывался
в rationale.
Patch: добавлено confounded: bool в DemandSupplyForecast (от §9.5 macro_coef
OR §9.6 rate_sensitivity), exposed в as_dict(). _confounded() уже использовал
.get() defensively — блокер был в producer'е.
+3 теста: contract на real DemandSupplyForecast.as_dict(), end-to-end
assemble_report → confounded_window factor surfaces at level=low, weakest-link
тянет overall к 'low'. 61 report_assembler + 1034 forecasting тестов зелёные.
Closes#1222
ProductScore.reason для price_feasibility (§14.2) называл основу платежа
'субсид. ставка', но compute_affordability с #981 DoD перешёл на рыночный
прокси key_rate + 4.5 п.п. (~19%, rate_kind='key_rate_proxy'); субсидированный
путь не вызывается. Метка делала рыночный платёж в §22-отчёте похожим на
льготный и противоречила affordability.degraded_reason в том же payload.
Текстовый фикс: 'рыночная ставка key_rate + спред, §7.9' — число платежа
было корректно, advisory-описание теперь совпадает.
Closes#1225
Degenerate price band (own_min==own_max или c_lo==c_hi, оба разрешены
CHECK 148 и Pydantic) внутри другой вилки возвращали 0.0 вместо 1.0:
фильтр w>0 выкидывал нулевую ширину → 0/positive=0.0. Это рвало
докстринг 'полное накрытие узкого = 1.0' и давало разрыв:
[148k,152k]→1.0 vs [150k,150k]→0.0, занижая среднее каннибализации.
Patch: вырожденные ширины обрабатываются ДО нормирования.
lo<=hi → точка внутри другой вилки → 1.0, вне → 0.0. +inf-обе-премиум
ветка перенесена в начало (избежать inf-inf=nan). +7 новых тестов в
TestPriceOverlap. 220 special_indices тестов зелёные.
Closes#1224
objective_lots.district хранит МИКРО-вокабуляр ('Уралмаш', 'ЖБИ', ...).
_query_artificial_demand фильтровал сырым АДМИН-именем ('Кировский' с
forecast.py:123) → ol.district='Кировский' = 0 строк → n_sold=0 → §25.5
Artificial Demand 'unavailable' с ложной причиной «нет проданных лотов»
в каждом district-scoped отчёте. Тот же класс бага, что #1211 в
_price_sensitivity.
Patch: импорт resolve_objective_districts + замена сырого
`ol.district = CAST(:district AS text)` на зеркальный паттерн
sales_series._SOURCE_B_SQL / market_metrics._SALES_WINDOW_SQL:
(CAST(:has_district AS boolean) IS FALSE
OR ol.district = ANY(CAST(:districts AS text[])))
Сигнатура _query_artificial_demand / _build_artificial_demand НЕ меняется
— caller остаётся admin-aware на входе.
+5 новых тестов (TestArtificialDemandDistrictResolution: резолвер вызван,
микро в bind, n_sold>0 после фикса), 6 обновлённых SQL-тестов. 21 passed
artificial_demand + 1030 forecasting тестов зелёные. ruff clean.
Closes#1205
`analyze["district"]` в этой кодовой базе — dict вида
{"district_name": "Верх-Исетский", "dist_to_center": 0.0, "median_price_per_m2": ...}.
Штатный caller (`workers/tasks/forecast.py:123`) явно извлекает `district_name`:
`district = row.district or analyze["district"]["district_name"]`. Но новые callers
(тесты, расширения чата, ad-hoc эндпоинты) легко передают сырой dict без знания этой
конвенции — тогда внутри §9.x-слоёв compute_market_metrics(district=<dict>) падает
с TypeError: unhashable type: 'dict' в forecast_request_cache.wrapper,
`_safe_call` это проглатывает → секции future_market.forecasts_by_horizon=[] и
scenarios.by_scenario={} тихо остаются пустыми (silent degrade, не 500).
Добавлен `_normalize_district(district)` — pure-нормализация на входе оркестратора:
- str → как есть;
- None → None;
- dict с district_name (непустая строка) → извлекаем;
- dict без district_name / с пустым / неподдерживаемый тип → None + logger.warning.
7 unit-тестов в test_orchestrator.py::TestNormalizeDistrict (все варианты входов).
Не меняет поведение штатного caller'а (str → str), только защищает от случайных
dict-callers.
Discovered through: #1130 Phase A (мой первый тестовый скрипт со скормленным
сырым `analyze["district"]` dict выдал forecasts.n_horizons=0 + 15 TypeError'ов
в _safe_call). Закрывает чип task_4a4aa3bb.
Refs #1130
Feed candidate_unit_mix into _build_cannibalization (mirrors how #1169 fed
candidate_release_month from the launch window), completing §25.3 to all 4 axes:
class + price + timing + unit-mix, plus geo weight.
- candidate mix from recommend_mix "buckets[].share_pct" (same rule-based
квартирография as §22 product_tz), extracted + normalized to {bucket: share}.
- _canonical_room_bucket folds recommend_mix RU labels ("1-к 30-45", "80+ м²")
and manual own_planned_project Latin keys ("1k") into one room-count space —
without it the L1 similarity would silently be 0 (disjoint keys).
- recommend_mix is HEAVY, so it's GATED: derived only when the own-portfolio has
>=1 project with a non-empty unit_mix; get_own_portfolio fetched once in
compute_special_indices and threaded into _build_cannibalization (no double
fetch). With OWN_DEVELOPER_IDS unset (portfolio empty) → zero added cost on the
hot §22 report path.
- Graceful (recommend_mix None/empty/raises → axis excluded, None-not-0),
deterministic. Unit-mix only fires for manual-future own-projects with a mix
(domrf-current carry unit_mix=None) — expected narrowness, documented.
205 tests; ruff + mypy clean. Scope: special_indices.py + test only; no deps.
Refs #1169
Upgrade the §25.3 cannibalization index from a same-class-competitor proxy to
true own-portfolio overlap: score the recommended candidate segment against the
developer's own portfolio (get_own_portfolio, #1169 PR1) across axes —
audience/class (ordinal distance), price ₽/м² (interval overlap), unit-mix
(L1 similarity), timing (half-life decay) — geo-weighted by haversine proximity
to the parcel, aggregated by geo-weighted soft-max (the strongest nearby
cannibalizer dominates, not a diluting mean). Empty portfolio -> labelled proxy
fallback, confidence forced low, never presented as the true index.
Pure scoring fns unit-tested without DB; None-not-0 on missing axes; thin/
only-current portfolio -> low confidence + §26 note; deterministic (sorted
tie-breaks, no RNG). class+price+geo active now; unit-mix+timing plumbed via
optional params for a follow-up that wires them from the orchestrator horizon.
ruff + mypy clean; 151 special-index tests pass (964 forecasting dir, no regr).
Refs #1169
fedstat ИПЦ is reCAPTCHA-blocked; CBR publishes inflation openly. Add
fetch_inflation + parse_inflation_xlsx (CBR UniDbQuery DownloadExcel/132934,
monthly % г/г, region=rf, source=cbr) to cbr_macro.py; upsert
indicator_type=inflation_yoy via the existing cbr_macro_sync task (per-series
guard, SAVEPOINT-per-row, CAST not ::, ON CONFLICT on the PK).
Surface inflation_yoy in MonthlyMacro (frozen, carry-forward) and ACTIVATE the
reserved §9.5 inflation channel (macro_coefficient f_inflation: level-vs-4%-target
nudge, non-positive to avoid double-counting f_rate, excluded from
_RATE_DRIVEN_FACTORS). Channel was DEGRADED (no data) -> now BACKED + consumed;
_CONF_HIGH_MIN_BACKED 4->5. Deterministic (§16/§26); renorm claims the reserved
0.08 slice as designed. Live-verified (2026-04 5.58%); 194 macro + 902 forecasting
tests green. No migration, no new deps.
Refs #946.
Cold §22 forecast measured ~215-233s on prod: §9.x layers re-execute the same
horizon/segment-invariant DB loads with identical args hundreds of times per
report (profiled: get_competitors x69, market_metrics x124, get_monthly_macro
x290). Add a per-report ContextVar cache (forecast_cache(), opened once in the
orchestrator) + @cached(key_builder) on the expensive §9.x loaders so each
unique load runs ONCE and reuses the same frozen, read-only instance.
Output is byte-identical (memoized producers are frozen dataclasses / read-only
Pydantic, callers never mutate; cache is per-report, discarded on exit; no-op
outside the report build). No concurrency, no signature changes.
- forecast_request_cache.py: ContextVar cache + cached() decorator (no-op
outside context, reentrant, _MISS sentinel for cached None)
- @cached on competitors/future_supply/market_metrics/macro_series/
sales_series/macro_coefficient/demand_normalization/regression loaders
- orchestrator: wrap build_site_finder_report in forecast_cache()
- 58 tests: key discrimination (call-counting regression guard), no-op-outside,
per-report isolation, reentrancy, frozen-producer canary, amplification proof
(real get_monthly_macro xN->1)
code-reviewer APPROVE (keys correct, mutation-safe, output identical). 1265
forecast/cache tests green. No new deps. Refs #1129.
deficit_index pins to -1.0 for every ЕКБ segment (12mo demand flow vs multi-year
supply stock → log-ratio clamps) → zero discriminating power, though the oversupply
is partly real. Add MOI (gross competing supply / demand_per_mo), the real-estate
absorption standard, as an additive non-saturating companion that DISCRIMINATES
(Уралмаш 42mo … Чермет 109mo) where deficit cannot. deficit_index math kept exactly
as-is (honest absolute: -1 = genuinely oversupplied); docstrings clarify -1 is common
and MOI is the discriminating companion (no recalibration). _gross_supply extract-method
(single source of truth; _project_supply behavior byte-identical, code-review-verified).
Surface MOI in §22 future_market (passthrough) + exec_summary key_numbers/verdict.
Guards: no demand → None, no supply → 0. Prod: MOI varies 42→109mo, deficit stays -1.
Discrimination test pins MOI separating two segments both at deficit -1. Refs #952.
/analyze passes the official ЕКБ admin district (ekb_districts polygon, e.g.
'Кировский'), but objective_lots/corpus_room_month store informal micro-districts
('Втузгородок','ЖБИ') -> admin name matched 0 rows -> silent empty forecast.
Add resolve_objective_districts() (site_finder/district_resolver.py) mapping an
admin name to its clean micros via ekb_district_alias (note IS NULL), with
None -> EKB-wide fallback and raw-micro pass-through. Wire into the objective_lots
district filters of market_metrics (§9.2 stock+sales), supply_layers L1 (§9.3),
and sales_series Sources A+B (crm shares the micro vocab, prod-verified),
switching the scalar filter to psycopg3-safe = ANY(CAST(:districts AS text[])).
supply_layers L2/L3 keep the admin name (domrf_kn_objects.district_name is admin vocab).
Prod: Кировский/Ленинский/Орджоникидзевский obj_count 0 -> 32/64/31.
Tests mutation-verified non-vacuous. 192 module tests pass; ruff clean. Refs #969#949.
_KEY_RATE_MARKET_SPREAD_PP was a 0.0 placeholder, so §7.9 affordability payments
used the bare CBR key_rate (~14.5%), understating borrower cost and OVERSTATING
affordability. Calibrate to 4.5pp from the prod anchor (macro_indicator
mortgage_rate_primary_domrf 19.125% @ 2026-04-19 vs key_rate ~14.5% -> observed
~4.6pp, rounded conservatively; inside the typical RF 3-5pp band), so
rate_used = key_rate + 4.5 ~= 19% matches the directly-observed market primary
mortgage rate. Makes affordability LESS optimistic / more accurate. Docs + tests
updated to the symbolic spread (new regression anchor pins spread==4.5 and
key_rate 14.5 -> ~19.0); rate_kind/graceful-fallback semantics unchanged.
Forecasting suite 841 passed; ruff clean.
Backtest (OOS directional hit-rate): single-best-lag compute_rate_sensitivity
is directionally noise (0.148 Source B EKB-wide, lag-unstable); the Almon
distributed-lag estimator (compute_district_rate_regression) is strictly less
noisy on every tier (0.407 Source B / 0.60 survivorship-free Source A,
lag-stable). Add a thin adapter compute_rate_regime_sensitivity mapping
DistributedLagFit onto the existing RateSensitivity contract (beta=long-run
sum-beta, confidence regression->medium / fallback->low, district=None->low and
no call) and repoint the three consumers (demand_normalization, product_scoring,
demand_supply_forecast). Magnitude bounded by the existing [0.5,1.2] clamp.
Reversible; compute_rate_sensitivity kept for the backtest. Consumer tests
repointed to the real Almon path (mutation-verified genuine) + adapter unit
tests + end-to-end fallback degradation. Forecasting suite 840 passed; ruff clean.
Found by read-only services audit.
- recommendation._usp_from_deficits: skip di<=0 so «стройте его» is never emitted
for OVERSUPPLIED formats; all-surplus top-K → [] (no white-space niches).
Aligns with product_scoring._count_positive_usp (di>0). Was: «Дефицит формата
X — стройте его» for a surplus format, reaching PDF/Excel USP-ниши.
- report_assembler._domrf_coverage: drop ambiguous >1.0 percent-guess; normalize
per-branch (analyze pct /100, supply_layers fraction as-is). Sub-1% coverage
(0.8%) no longer read as 80% → no inflated confidence in the near-zero-coverage
case §15 flags. tests for both + end-to-end no-inflation. 241+148 pass.
REOPENED 951-B §9.6.
PART A: fix look-ahead leakage in backtest_rate_sensitivity --detrend. The
ln(units) trend was fit over train+test then split, so test data shaped the
detrend and inflated the OOS hit-rate. _detrend_log now takes fit_n; backtest_tier
fits the trend on TRAIN months only (same split evaluate_oos uses) and projects
(a,b) point-in-time onto test. Default fit_n=None preserves prior behaviour.
PART B (DoD): new app/services/forecasting/regression.py — Almon polynomial
distributed-lag (deg 2) of Δln(district demand) on Δkey_rate lags 0..6 via
OLS-on-Almon-regressors (numpy lstsq) + per-lag reconstruction + manual
Newey-West HAC SEs (NO statsmodels). Output {best_lag_months, coef=long-run
multiplier, x_pct, r2, n, per_lag_coef, hac_se,...}; gate mirrors _elasticity_coef
(n<30 OR R²<0.1 OR Σβ≥0 → fallback); §9.6 phrase from the lag shape. ADVISORY,
shipped standalone (integration point documented), NOT wired — protects the live
compute_rate_sensitivity consumers.
125+31 tests (synthetic known-lag recovery, HAC computed/differs-from-OLS,
fallback gating, no-leakage detrend). ruff clean. Refs #978
REOPENED — normalize.py was never created; only rate-regime discount existed.
New backend/app/services/forecasting/normalize.py with normalize_demand(series):
multiplicative month-of-year deseasonalization of the raw monthly demand
SalesSeries (§9.4). Pure/deterministic; min-data guard (<2 full years / empty
month / overall_mean<=0 → factor 1.0, no divide-by-zero, no thin-data noise).
Exposes seasonal factors for explainability. Synthetic unit test: seasonality
removed (month means equalised), flat unchanged, thin/empty/all-zero safe.
DoD (module + doc + test) MET. Production wiring into
rate_sensitivity._align_sales_deltas DEFERRED (documented TODO): deseasonalizing
the short rate-driven series perturbs the recovered β/lag on current data —
needs a points-per-month gate / joint seasonal+rate estimation + backtest before
wiring. Forecast stack is advisory regardless. Refs #979
REOPENED#980: when effective competing supply is exhausted under positive demand
(projected_supply<=0, demand>0), deficit_index now caps to +1.0 (peak of [-1,+1])
instead of None. balance_ratio stays None (demand/0 undefined), but the strongest
build signal no longer reads downstream as thin data (market_fit fell to 0.5,
what_to_build dropped the cell). No-signal (supply<=0 AND demand<=0) stays None.
REOPENED#981: MAI now uses CBR key rate (macro_indicator key_rate/rf via
get_monthly_macro) as the market borrowing-cost proxy (~16-21%) instead of the
subsidized weighted rate (~7.83%), per §7.9 DoD. rate_kind='key_rate_proxy'.
If key_rate absent → rate_kind='market_unavailable' (no silent subsidy fallback).
Income (#946) still missing → payment_to_income None, confidence low.
778 forecasting tests green. Refs #980#981
