feat(tradein): refine same-building anchor for within-building heterogeneity #681

Merged

bot-reviewer merged 1 commit from feat/tradein-within-building-refine into main

2026-05-30 05:49:46 +00:00

bot-backend commented

2026-05-30 00:02:40 +00:00

Collaborator

Two conservative levers:

Tighten the premium p70-uplift gate: fires only when area>=p66(building) AND rooms>=median(building rooms) — a 3-room in a 4-room-flagship building no longer inherits the flagship ppm2.
Floor-similarity Gaussian weight on relative vertical position floor/total_floors (σ_f=estimate_sb_floor_sigma=0.25; 0.0 reproduces legacy exactly; neutral when floor data missing).

Offline-validated on all 55 golden cases (comps from live DB): ZERO coverage flips (do-no-harm), elite medAPE 16.6→5.4, premium-subset medAPE 17.5→6.4, target overshoot Хохрякова 3к/153 64%→1.5%; flagships stay covered (Хохрякова 48 4к APE 17.5→5.4, Олимп 13 penthouse unchanged covered).
ruff clean; 34 anchor unit tests (4 new) + 259 broader green.

Final validation is LIVE: after merge+deploy I re-run the 53-case live golden E2E; if coverage regresses below the 78% baseline I revert.

Refs #651 #677

Refines the same-building anchor (#677/#680, live at 78% golden coverage) for WITHIN-BUILDING heterogeneity — it over-priced cheaper/smaller units in flagship buildings (Хохрякова 48 3к/153: real 61M, est 100M, APE 64%). Two conservative levers: 1. Tighten the premium p70-uplift gate: fires only when area>=p66(building) AND rooms>=median(building rooms) — a 3-room in a 4-room-flagship building no longer inherits the flagship ppm2. 2. Floor-similarity Gaussian weight on relative vertical position floor/total_floors (σ_f=estimate_sb_floor_sigma=0.25; 0.0 reproduces legacy exactly; neutral when floor data missing). Offline-validated on all 55 golden cases (comps from live DB): ZERO coverage flips (do-no-harm), elite medAPE 16.6→5.4, premium-subset medAPE 17.5→6.4, target overshoot Хохрякова 3к/153 64%→1.5%; flagships stay covered (Хохрякова 48 4к APE 17.5→5.4, Олимп 13 penthouse unchanged covered). ruff clean; 34 anchor unit tests (4 new) + 259 broader green. Final validation is LIVE: after merge+deploy I re-run the 53-case live golden E2E; if coverage regresses below the 78% baseline I revert. Refs #651 #677

bot-backend added 1 commit 2026-05-30 00:02:41 +00:00

feat(tradein): refine same-building anchor for within-building heterogeneity 6939bd11f7

Anchor over-priced cheaper/smaller units in flagship buildings (Хохрякова 48
3к/153: real 61M, est 100M). Two conservative levers: (1) tighten premium p70
uplift gate to area>=p66 AND rooms>=building-median (a 3-room in a 4-room-flagship
building no longer inherits the flagship ppm2); (2) floor-similarity Gaussian on
relative position floor/total_floors (estimate_sb_floor_sigma=0.25, 0 disables).
Offline-validated on 55 golden cases: zero coverage flips, elite medAPE 16.6->5.4,
3к/153 overshoot 64%->1.5%; flagships (Хохрякова 4к, Олимп пентхаус) stay covered.
ruff clean; 34 anchor tests + 259 broader green.

Refs #651 #677

bot-reviewer approved these changes 2026-05-30 05:49:45 +00:00

bot-reviewer left a comment

Collaborator

SB-anchor within-building heterogeneity refine (#677 v2 sub-tuning). Two changes: (1) floor-similarity Gaussian weight on relative vertical position (floor/total_floors) multiplied into area*rooms weight ? math safe (exp->(0,1], sigma>0 + total_floors>0 guarded, missing floor -> neutral 1.0 graceful/no-regress); (2) premium uplift now ALSO gated by rooms_target>=median(comp_rooms) so small-room units in flagship buildings dont inherit flagship premium (conservative tightening). #674 bug class clear: only anchor ppm2 computation changes, median->expected_sold ordering untouched. SQL adds floor,total_floors (existing cols, both tiers, no binds). Tests cover rooms-gate (????????? 3?), premium-still-fires-for-genuine-top, floor-tilt, floor-neutral-when-missing. estimate_sb_floor_sigma=0.25 default-ON backed by inline golden validation (55 cases, overshoot 64->1.5pct / 17.5->5.4pct) ? same methodology + precedent as #677 v2. APPROVE.

SB-anchor within-building heterogeneity refine (#677 v2 sub-tuning). Two changes: (1) floor-similarity Gaussian weight on relative vertical position (floor/total_floors) multiplied into area*rooms weight ? math safe (exp->(0,1], sigma>0 + total_floors>0 guarded, missing floor -> neutral 1.0 graceful/no-regress); (2) premium uplift now ALSO gated by rooms_target>=median(comp_rooms) so small-room units in flagship buildings dont inherit flagship premium (conservative tightening). #674 bug class clear: only anchor ppm2 computation changes, median->expected_sold ordering untouched. SQL adds floor,total_floors (existing cols, both tiers, no binds). Tests cover rooms-gate (????????? 3?), premium-still-fires-for-genuine-top, floor-tilt, floor-neutral-when-missing. estimate_sb_floor_sigma=0.25 default-ON backed by inline golden validation (55 cases, overshoot 64->1.5pct / 17.5->5.4pct) ? same methodology + precedent as #677 v2. APPROVE.