lekss361/gendesign
1
0
Fork 0
Code Issues 141 Pull requests 2 Projects Releases Packages Wiki Activity Actions 1

fix(mera/estimate): перефит хедоники поверх area-бакетного ratio — крупное жильё занижалось на 21% (#3255)
All checks were successful
Deploy Trade-In / changes (push) Successful in 12s
Details
Deploy Trade-In / build-frontend (push) Has been skipped
Details
Deploy Trade-In / build-browser (push) Has been skipped
Details
Deploy Trade-In / test (push) Successful in 4m5s
Details
Deploy Trade-In / build-backend (push) Successful in 1m1s
Details
Deploy Trade-In / deploy (push) Successful in 1m34s
Details
Deploy Trade-In / deploy-status (push) Successful in 1s
Details
Deploy Trade-In / perimeter-smoke (push) Successful in 12s
Details

Browse source
This commit is contained in:
lekss361 2026-08-29 20:19:27 +00:00
parent e523c8949c
commit 5da226a271
3 changed files with 126 additions and 81 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

35
tradein-mvp/backend/app/core/config.py
View file

@ -431,10 +431,37 @@ class Settings(BaseSettings):
# После фактора заново применяется le_asking-кламп (expected_sold ≤ asking). # После фактора заново применяется le_asking-кламп (expected_sold ≤ asking).
# OFF ⇒ точно старое поведение expected_sold. # OFF ⇒ точно старое поведение expected_sold.
estimate_hedonic_correction_enabled: bool = True estimate_hedonic_correction_enabled: bool = True
estimate_hedonic_b0: float = 0.6146 # fit log(sold/es) ~ year + ln(area), n=2366 (#2002) # #3248 (перефит 2026-08-30, n=1269 из свежей прод-фикстуры ЕКБ 1600 сделок).
estimate_hedonic_year_coef: float = 0.1220 # per (year-2000)/20 #
estimate_hedonic_larea_coef: float = -0.1603 # per ln(area_m2) # Прежние значения (b0=0.6146, year=0.1220, larea=-0.1603, first=-0.1248) зафичены
estimate_hedonic_first_floor_coef: float = -0.1248 # floor==1 ground-floor ≈ -12%; #2002 n=2366 # 2026-06-27 (#2002), когда asking→sold ratio ключевался ПО КОМНАТАМ. 2026-08-05
# (#2620) ratio переключили на area-бакеты — то есть под хедонику подставили ДРУГУЮ
# базу, а её саму не пересчитали. Хедоника по определению чинит ОСТАТОК
# log(actual_sold / expected_sold), поэтому её коэффициенты верны только для той
# базы, на которой фитились. Итог: площадь штрафовалась дважды, крупные лоты
# занижались на 21% (bias 4+ комнат = -21.4%).
#
# larea = 0.0 ВЫСТАВЛЕН НАМЕРЕННО, это не «не задан». После #2620 площадь несёт
# area-бакетный ratio, и его форма — перевёрнутая U (факт sold/ask по бакетам:
# 0.786 / 0.829 / 0.899 / 0.954 / 0.864), которую монотонный ln(area) выразить не
# может в принципе: он тянет крупное жильё вниз ровно там, где рынок его не
# дисконтирует. Член стал избыточным и вредным — обнуляем, оставляя код-путь.
#
# Замер вариантов на той же фикстуре (bias по комнатам, median):
# текущие MAPE 13.90 | студия +7.6 1к -0.5 2к -2.8 3к -5.3 4+ -21.4
# OLS все 4 члена MAPE 14.11 | студия -0.9 1к -1.6 2к -0.7 3к -4.0 4+ -14.1
# БЕЗ larea (тут) MAPE 14.28 | студия -3.1 1к -2.3 2к -0.4 3к -1.9 4+ -10.4
# хедоника OFF MAPE 15.88 | студия -1.5 1к +4.4 2к +4.2 3к -2.3 4+ -5.0
# Берём «без larea»: худший перекос вдвое меньше, все прочие классы в пределах
# ±3.1%, цена — +0.38 п.п. общего MAPE (хедоника сжимает разброс за счёт year).
#
# ОСТАТОК -10.4% по 4+ конфигом НЕ закрывается: ratio бакета 4 = 0.8211 при
# фактических sold/ask = 0.8640, т.е. занижен на ~5% ДО всякой хедоники. Это
# пересчёт самой таблицы (app/tasks/asking_to_sold_ratio.py), см. follow-up.
estimate_hedonic_b0: float = -0.0140 # #3248 перефит поверх area-бакетного ratio
estimate_hedonic_year_coef: float = 0.0769 # per (year-2000)/20
estimate_hedonic_larea_coef: float = 0.0 # НАМЕРЕННО 0 — площадь несёт ratio (#2620)
estimate_hedonic_first_floor_coef: float = -0.0745 # floor==1 ground-floor ≈ -7%
estimate_hedonic_factor_min: float = 0.75 estimate_hedonic_factor_min: float = 0.75
estimate_hedonic_factor_max: float = 1.30 estimate_hedonic_factor_max: float = 1.30
# ── #1795: premium headline anti-inflation (4 фикса, каждый за флагом) ────── # ── #1795: premium headline anti-inflation (4 фикса, каждый за флагом) ──────

146
tradein-mvp/backend/tests/fixtures/backtest_baseline.json vendored
View file

@ -1,20 +1,20 @@
{ {
"calibration": { "calibration": {
"high": { "high": {
"coverage_pct": 50.0, "coverage_pct": 75.0,
"mape_pct": 27.87, "mape_pct": 25.94,
"n": 8, "n": 8,
"n_covered": 4 "n_covered": 6
}, },
"low": { "low": {
"coverage_pct": 86.68, "coverage_pct": 87.41,
"mape_pct": 13.76, "mape_pct": 14.06,
"n": 1562, "n": 1562,
"n_covered": 1067 "n_covered": 1076
}, },
"medium": { "medium": {
"coverage_pct": 73.33, "coverage_pct": 73.33,
"mape_pct": 17.91, "mape_pct": 18.64,
"n": 30, "n": 30,
"n_covered": 22 "n_covered": 22
} }
@ -26,111 +26,111 @@
], ],
"expected_sold": { "expected_sold": {
"overall": { "overall": {
"mape_pct": 13.9, "mape_pct": 14.28,
"median_bias_pct": -0.78, "median_bias_pct": -1.61,
"n": 1269, "n": 1269,
"n_no_analogs": 0, "n_no_analogs": 0,
"p25_pct": -13.39, "p25_pct": -14.08,
"p75_pct": 14.46 "p75_pct": 14.6
}, },
"per_area_bucket": { "per_area_bucket": {
"0 <30": { "0 <30": {
"mape_pct": 14.55, "mape_pct": 16.51,
"median_bias_pct": 7.59, "median_bias_pct": -3.13,
"n": 161, "n": 161,
"p25_pct": -4.16, "p25_pct": -14.08,
"p75_pct": 29.4 "p75_pct": 22.87
}, },
"1 30-44": { "1 30-44": {
"mape_pct": 13.12, "mape_pct": 14.06,
"median_bias_pct": -0.51, "median_bias_pct": -2.33,
"n": 503, "n": 503,
"p25_pct": -11.98, "p25_pct": -14.49,
"p75_pct": 14.88 "p75_pct": 12.58
}, },
"2 44-62": { "2 44-62": {
"mape_pct": 12.99, "mape_pct": 12.48,
"median_bias_pct": -2.82, "median_bias_pct": -0.39,
"n": 379, "n": 379,
"p25_pct": -14.44, "p25_pct": -11.23,
"p75_pct": 10.26 "p75_pct": 14.65
}, },
"3 62-85": { "3 62-85": {
"mape_pct": 15.32, "mape_pct": 15.24,
"median_bias_pct": -5.33, "median_bias_pct": -1.94,
"n": 179, "n": 179,
"p25_pct": -18.99, "p25_pct": -14.98,
"p75_pct": 7.16 "p75_pct": 15.27
}, },
"4 >=85": { "4 >=85": {
"mape_pct": 23.21, "mape_pct": 20.88,
"median_bias_pct": -21.36, "median_bias_pct": -10.36,
"n": 47, "n": 47,
"p25_pct": -29.97, "p25_pct": -24.08,
"p75_pct": 6.3 "p75_pct": 13.38
} }
}, },
"per_rooms": { "per_rooms": {
"0": { "0": {
"label": "студия", "label": "студия",
"mape_pct": 14.55, "mape_pct": 16.51,
"median_bias_pct": 7.59, "median_bias_pct": -3.13,
"n": 161, "n": 161,
"n_no_analogs": 0, "n_no_analogs": 0,
"p25_pct": -4.16, "p25_pct": -14.08,
"p75_pct": 29.4 "p75_pct": 22.87
}, },
"1": { "1": {
"label": "1к", "label": "1к",
"mape_pct": 13.12, "mape_pct": 14.06,
"median_bias_pct": -0.51, "median_bias_pct": -2.33,
"n": 503, "n": 503,
"n_no_analogs": 0, "n_no_analogs": 0,
"p25_pct": -11.98, "p25_pct": -14.49,
"p75_pct": 14.88 "p75_pct": 12.58
}, },
"2": { "2": {
"label": "2к", "label": "2к",
"mape_pct": 12.99, "mape_pct": 12.48,
"median_bias_pct": -2.82, "median_bias_pct": -0.39,
"n": 379, "n": 379,
"n_no_analogs": 0, "n_no_analogs": 0,
"p25_pct": -14.44, "p25_pct": -11.23,
"p75_pct": 10.26 "p75_pct": 14.65
}, },
"3": { "3": {
"label": "3к", "label": "3к",
"mape_pct": 15.32, "mape_pct": 15.24,
"median_bias_pct": -5.33, "median_bias_pct": -1.94,
"n": 179, "n": 179,
"n_no_analogs": 0, "n_no_analogs": 0,
"p25_pct": -18.99, "p25_pct": -14.98,
"p75_pct": 7.16 "p75_pct": 15.27
}, },
"4": { "4": {
"label": "4+", "label": "4+",
"mape_pct": 23.21, "mape_pct": 20.88,
"median_bias_pct": -21.36, "median_bias_pct": -10.36,
"n": 47, "n": 47,
"n_no_analogs": 0, "n_no_analogs": 0,
"p25_pct": -29.97, "p25_pct": -24.08,
"p75_pct": 6.3 "p75_pct": 13.38
} }
}, },
"per_segment": { "per_segment": {
"бизнес": { "бизнес": {
"mape_pct": 11.85, "mape_pct": 16.86,
"median_bias_pct": -9.98, "median_bias_pct": -16.84,
"n": 144, "n": 144,
"p25_pct": -20.69, "p25_pct": -25.83,
"p75_pct": -1.72 "p75_pct": -9.36
}, },
"комфорт": { "комфорт": {
"mape_pct": 12.71, "mape_pct": 12.14,
"median_bias_pct": -2.39, "median_bias_pct": -5.9,
"n": 425, "n": 425,
"p25_pct": -16.15, "p25_pct": -16.66,
"p75_pct": 9.46 "p75_pct": 4.9
}, },
"премиум": { "премиум": {
"mape_pct": null, "mape_pct": null,
@ -140,18 +140,18 @@
"p75_pct": null "p75_pct": null
}, },
"эконом": { "эконом": {
"mape_pct": 14.98, "mape_pct": 15.17,
"median_bias_pct": 2.46, "median_bias_pct": 6.26,
"n": 697, "n": 697,
"p25_pct": -9.54, "p25_pct": -6.89,
"p75_pct": 26.95 "p75_pct": 28.97
}, },
"элит": { "элит": {
"mape_pct": 18.45, "mape_pct": 31.18,
"median_bias_pct": -18.45, "median_bias_pct": -31.18,
"n": 3, "n": 3,
"p25_pct": -40.32, "p25_pct": -46.94,
"p75_pct": -15.7 "p75_pct": -28.1
} }
} }
}, },
@ -162,20 +162,20 @@
}, },
"range_coverage": { "range_coverage": {
"overall": { "overall": {
"coverage_pct": 86.13, "coverage_pct": 87.0,
"n": 1269, "n": 1269,
"n_covered": 1093 "n_covered": 1104
}, },
"per_confidence": { "per_confidence": {
"high": { "high": {
"coverage_pct": 50.0, "coverage_pct": 75.0,
"n": 8, "n": 8,
"n_covered": 4 "n_covered": 6
}, },
"low": { "low": {
"coverage_pct": 86.68, "coverage_pct": 87.41,
"n": 1231, "n": 1231,
"n_covered": 1067 "n_covered": 1076
}, },
"medium": { "medium": {
"coverage_pct": 73.33, "coverage_pct": 73.33,

26
tradein-mvp/backend/tests/test_estimator_hedonic.py
View file

@ -126,7 +126,15 @@ def test_mid_case_shifts_by_expected_factor(monkeypatch: pytest.MonkeyPatch) ->
def test_factor_clamps_to_min_for_huge_area(monkeypatch: pytest.MonkeyPatch) -> None: def test_factor_clamps_to_min_for_huge_area(monkeypatch: pytest.MonkeyPatch) -> None:
"""Very large area → raw factor < factor_min → clamped to the floor.""" """Very large area → raw factor < factor_min → clamped to the floor.
#3248: тест проверяет МЕХАНИЗМ клэмпа, а не конкретную подгонку. С момента
перефита `estimate_hedonic_larea_coef` = 0 (площадь несёт area-бакетный ratio),
поэтому площадь сама по себе фактор вниз больше не гонит. Задаём площадной
коэффициент явно — иначе тест молча перестаёт проверять клэмп при каждом
перефите вместо того, чтобы падать.
"""
monkeypatch.setattr(estimator.settings, "estimate_hedonic_larea_coef", -0.1603)
monkeypatch.setattr(estimator.settings, "estimate_hedonic_correction_enabled", False) monkeypatch.setattr(estimator.settings, "estimate_hedonic_correction_enabled", False)
off = _price(area_m2=10_000.0, target_year=None, ratio=0.85) off = _price(area_m2=10_000.0, target_year=None, ratio=0.85)
monkeypatch.setattr(estimator.settings, "estimate_hedonic_correction_enabled", True) monkeypatch.setattr(estimator.settings, "estimate_hedonic_correction_enabled", True)
@ -145,7 +153,14 @@ def test_factor_clamps_to_max_for_small_new_lot(monkeypatch: pytest.MonkeyPatch)
le_asking is held OFF so the raw ceiling factor is observable on the point le_asking is held OFF so the raw ceiling factor is observable on the point
(otherwise the re-clamp would cap it at the asking headline). (otherwise the re-clamp would cap it at the asking headline).
#3248: коэффициенты задаются явно — тест про МЕХАНИЗМ потолка, а не про
текущую подгонку (после перефита b0 = -0.0140 и потолка сам по себе не
достаёт).
""" """
monkeypatch.setattr(estimator.settings, "estimate_hedonic_b0", 0.6146)
monkeypatch.setattr(estimator.settings, "estimate_hedonic_larea_coef", -0.1603)
monkeypatch.setattr(estimator.settings, "estimate_hedonic_year_coef", 0.1220)
monkeypatch.setattr(estimator.settings, "estimate_expected_sold_le_asking", False) monkeypatch.setattr(estimator.settings, "estimate_expected_sold_le_asking", False)
monkeypatch.setattr(estimator.settings, "estimate_hedonic_correction_enabled", False) monkeypatch.setattr(estimator.settings, "estimate_hedonic_correction_enabled", False)
off = _price(area_m2=15.0, target_year=2025, ratio=0.85) off = _price(area_m2=15.0, target_year=2025, ratio=0.85)
@ -199,7 +214,7 @@ def test_le_asking_off_allows_hedonic_above_asking(monkeypatch: pytest.MonkeyPat
def test_ground_floor_applies_extra_discount(monkeypatch: pytest.MonkeyPatch) -> None: def test_ground_floor_applies_extra_discount(monkeypatch: pytest.MonkeyPatch) -> None:
"""floor==1 → the extra exp(-0.1248)≈0.88 discount vs floor=3 (#2002). """floor==1 → дополнительная скидка exp(first_floor_coef) против floor=3.
Same year/area/ratio; only the floor differs. The mid case sits strictly inside Same year/area/ratio; only the floor differs. The mid case sits strictly inside
the clamp band so the first-floor term is fully observable (no clamp confound). the clamp band so the first-floor term is fully observable (no clamp confound).
@ -219,8 +234,11 @@ def test_ground_floor_applies_extra_discount(monkeypatch: pytest.MonkeyPatch) ->
f_upper = _expected_factor(50.0, 2010, floor=3) f_upper = _expected_factor(50.0, 2010, floor=3)
extra = math.exp(estimator.settings.estimate_hedonic_first_floor_coef) extra = math.exp(estimator.settings.estimate_hedonic_first_floor_coef)
# ground-floor multiplies the year+area factor by the extra ~0.88 discount. # ground-floor multiplies the year+area factor by the extra discount.
assert extra == pytest.approx(0.8827, abs=1e-3) # #3248: сверяем со ЗНАЧЕНИЕМ НАСТРОЙКИ, а не с зашитым 0.8827 — коэффициент
# подгоняемый и меняется при каждом перефите, а проверяем мы связь термина с
# фактором. Границы держат тест осмысленным: скидка, но не обвал.
assert 0.80 < extra < 1.0
assert f_ground == pytest.approx(f_upper * extra) assert f_ground == pytest.approx(f_upper * extra)
# both factors strictly inside the clamp band → the term is fully observable. # both factors strictly inside the clamp band → the term is fully observable.
assert estimator.settings.estimate_hedonic_factor_min < f_ground < f_upper assert estimator.settings.estimate_hedonic_factor_min < f_ground < f_upper