knowledge: MAPE-aligned point prediction — median models overpredict under MAPE; per-row exp(−λσ²) shrinkage (new backend/common/ml)

[choiyounggi]

Knowledge flush — 1 insight ingested (2 dropped as pending-duplicates of open PR #51)

Queue drained: 3 pending candidates across 2 session files.

Verified best-practice

Ingested — MAPE-aligned point prediction (from avm-hackathon-jycho session):

• Claim: A regression model trained to predict the conditional median (log target + L1 loss) and scored by MAPE is a structural overpredictor; the MAPE-optimal point prediction is the median of the 1/y-reweighted predictive distribution, which equals median × exp(−σ²) for a lognormal conditional. Correct per-row with pred × exp(−λσ²), σ estimated from q16/q84 quantile spread, λ selected from 0.5 by multi-period holdout consistency rather than the theoretical 1.0.
• Sources checked:
  • https://arxiv.org/abs/0912.0902 — Gneiting, "Making and Evaluating Point Forecasts" (JASA 106:746–762, 2011). Table 5: the Bayes rule under absolute percentage error is the β-median with β = −1 (median of the y⁻¹-reweighted predictive distribution). Confirmed via WebSearch against arxiv.org and the tandfonline JASA record.
  • https://arxiv.org/abs/1605.02541 — de Myttenaere, Golden, Le Grand, Rossi, "Mean Absolute Percentage Error for regression models" (Neurocomputing 2016). Confirms MAPE-optimal regression ≡ 1/y-weighted MAE regression. The citation given in the raw candidate was checked and is real (also mirrored at hal.science/hal-01312590).
  • Lognormal algebra re-derived independently: for Y ~ LN(μ, σ²), y⁻¹·f(y) normalizes to LN(μ − σ², σ²), whose median is exp(μ − σ²) = median(Y)·exp(−σ²).
• Verification result: the shrinkage mechanism and its direction are verified against the two papers plus the closed-form derivation. The practical λ = 0.5 starting point and the "quantile spread overestimates σ" claim rest on the session's holdout evidence only (Seoul commercial-building AVM, LightGBM + CatBoost: λ = 0.5 improved MAPE on all 4 holdout years, 19.18% → 18.86% excluding 2025 outliers; λ = 1.0 degraded 2 of 4 years) — that part is field-tested and the page's Sources section says so explicitly. Page frontmatter is confidence: verified for the sourced mechanism, with the field-tested scope of the λ practice stated inline.

Dropped candidates (both from the linkly r3 orchestrate session) — see Open-PR check; neither was ingested, so no verification pass was spent upgrading them.

Existing-layer check

Routing went INDEX.md → backend → domain index. A full-text sweep of the checkout's wiki/ for mape, percentage error, lognormal, quantile, lightgbm, calibration returned zero hits — nothing in any layer covers regression-model point-prediction calibration, so this is a create, not a merge. The closest existing category is backend/common/llm, whose pages cover consuming LLM completion APIs, not training/evaluating predictive models; I opened its nearest page to confirm the trigger space does not overlap (it gates on HTTP completion responses, not on metrics or point forecasts).

Pages read: backend-common-llm-completion-response-validation

No conflicts flagged. related: left empty — no genuinely adjacent page exists yet (the new category is a seed; future ML ingests should link here). Plumbing updated: wiki/backend/index.md gained a ### ml section with a load-when line, the common-subtree route line and root INDEX.md backend line both mention MAPE-aligned point-prediction calibration, and log.md has the ingest entry.

Open-PR check

Open knowledge/* heads listed via gh pr list --search "head:knowledge/": #47 (knowledge/dch0202-20260806-130040), #49 (knowledge/dch0202-rsquare-20260806-142309), #50 (knowledge/dch0202-20260806-172420), #51 (knowledge/dch0202-20260806-183029), #52 (knowledge/dch0202-rsquare-20260807-100149).

Per-candidate verdicts:

• Candidate 1 — worktree_escape guardrail escalation on cross-worktree reads (linkly): drop. Fetched and diffed origin/knowledge/dch0202-20260806-183029 (knowledge: 9 insights — closed value table widening, guardrail read-vs-write correction, dispatch binding taxonomy #51) against main: its wiki/infrastructure/agent-orchestration/worktree-isolated-workers.md additions already carry this insight in strictly better form — including the exact "budget the escalation round trip (read → approve → clear escalations/ → restart the watcher) and state in the worker's first brief that reads are approved" row. knowledge: 9 insights — closed value table widening, guardrail read-vs-write correction, dispatch binding taxonomy #51 additionally corrects the raw candidate: its local reproduction (guardrails 1.2.0) shows pure reads pass and the rule fires only when a main-root mention co-occurs with a write verb or an absolute-path redirect — the candidate's blanket "fires on read-only access" phrasing is the pre-correction version. Nothing unique to fold; retired as pending-duplicate.
• Candidate 2 — Orca dispatch binding taxonomy (idle-prompt check; runtime_unavailable vs agent_unconfigured vs terminal_worktree_mismatch) (linkly): drop. The same knowledge: 9 insights — closed value table widening, guardrail read-vs-write correction, dispatch binding taxonomy #51 diff's wiki/infrastructure/agent-orchestration/pane-delivery-confirmation.md additions carry all four rows (bind only on idle prompt because "done" is a report not the turn's end; wait-and-rebind for occupied runtime; close pane + new worker-mode agent for a dead one; always pass worktree with pane) citing the same three 2026-08-06 field incidents the candidate cites. Nothing unique to fold; retired as pending-duplicate.
• Candidate 3 — MAPE-aligned point prediction (avm-hackathon): new. No overlap with any open head — knowledge: bats/bash-3.2 assertion trap + 4 orchestration/guard edges; 5 duplicates retired #47/knowledge: 3 verified testing-quality insights (value-preserving refactor tests, unasserted return fields, stale artifact baselines) #49/knowledge: one test per success-return site when a handler applies one policy at several returns #50/knowledge: 4 verified insights — surviving-mutant triage, source-text wiring assertions, query state vs fetch state, python text-io encoding #52 are testing/shell/encoding-themed and knowledge: 9 insights — closed value table widening, guardrail read-vs-write correction, dispatch binding taxonomy #51's wiki diff (11 files) touches no ML content. Ingested here.

Routing decision

• Candidate 3 → backend/common/ml/mape-aligned-point-prediction.md (id backend-common-ml-mape-aligned-point-prediction), new category ml under backend/common. Justification for the new category: the harvested domain hint was backend, and the backend domain's 12 existing categories (api-design, auth, caching, change-impact, concurrency, errors, integrations, jobs, llm, orm, reliability, storage) all cover server-side application code concerns; none covers training or evaluating a predictive model. llm is the nearest name but its scope is consuming LLM completion APIs from server code — putting metric-aligned regression calibration there would corrupt its load-when gate. No other domain fits better (databases owns SQL/schema; qa owns release process). The category seeds with one page.
• Candidates 1–2: no routing — dropped as pending-duplicates of open PR knowledge: 9 insights — closed value table widening, guardrail read-vs-write correction, dispatch binding taxonomy #51 (see above); their queue rows are retired to .processed.jsonl so the auto-flush cannot re-surface them.