Repository navigation
feat(supply): D3b v3.11 adopted 单一事实源 + 趋势/季节模型与趋势诊断(P0+P1) - #346
Merged
Merged
Conversation
- 3_分类 只 join 4_回测_asof 的 adopted:单源精确、多源一致/历史量最大并留痕; 采用门槛回退后重新 join;断言 fail closed,不再按产品历史量另选模型 - 新增 ETS(A,Ad,N)、ETS(A,A,A)/STL+ETS、简化 Theta(numpy 实现,无新依赖) - 新增趋势诊断:2026 vs 2025 同期、3/6/12 月斜率、结构突变标记; 输出 15_趋势诊断 + trend CSV - 新增 P0 ADR:H_key 公式、98/154/161/91/112 标准桶、A/B/C/D 服务先验、 rolling-origin + FVA 选择协议、春节供给日历 + bridge 原则 - 同步 supply_netting 文档索引与根索引 Co-Authored-By: Claude Code <noreply@anthropic.com>
P1: - Holt 缺测月按日历推进状态(level=预测值, trend=phi*trend),尾部缺测推进到预测起点 - validation(202512/202601, h=3) 选模;holdout(202604, h=6, 202605..202610) 只评估, select() 不读取 holdout;3a/4_回测 增加 split 与 validation/holdout 双指标 - 自检 FAIL 不写工作簿/报告/manifest 且非零退出 - ABC 实际分母改为共同键对账;A 缺 1 键/100 件单列,自检全 PASS P2: - ETS(A,A,A) 先还季节再截零,避免去季节负值被提前截断抬高预测 - 多源键含 none 标 adopted_partial,不再算“一致”;adopted_sources 留痕 - 同比只比较同月两边都有效观测的配对月,新增 paired_months 验收:3_分类 1749 / 3a 1853 / 4_回测 49,997(val 29,487 + holdout 20,510); 单源 763/763 精确、多源 85 条(agree 1 / partial 67 / maxqty 17)、0 不一致; 自检 PASS、0 fail;holdout FVA(中位数)0.2%,adopted 仅 23/640 键优于最佳基线。 Co-Authored-By: Claude Code <noreply@anthropic.com>
Owner
Author
|
已按 Codex 复审意见修完 P1/P2,commit P1
P2
验收: 独立 holdout 的结论:adopted 640 条,holdout WAPE 中位数 0.850 vs 最佳基线 0.851,holdout FVA 0.2%,仅 23/640 键优于基线。即 validation 选出的模型在真正独立的 6 个月里几乎没有超出基线的增量——这条已写进 PR 描述,作为后续 P2 重新考虑候选集/门槛的依据。 |
- 候选资格/分类/模型集只用 SELECTION_END=202604 之前的数据;holdout 202605..202610 不参与 eligibility、分类或模型集 - 3_分类 增加 n_obs_train/n_pos_train/demand_pattern_train/sbc_class_train; 3a_分类_asof 增加 *_full,训练期与全历史可对照 - A/B/C 仓级回测不再同窗选模:沿用前置 validation 的 adopted, 统一在 holdout origin 202604(h=6,202605..202610)评估 - A/B/C 汇总增加 common_keys/common_actual/common_wape/excluded_actual; 共同样本与 full 并列,差量单列 - 自检 PASS;4_回测 46,483(validation 27,701 + holdout 18,782) Co-Authored-By: Claude Code <noreply@anthropic.com>
Owner
Author
|
已修完 Codex 第二轮三条 P1,commit
最终验收: 独立 holdout:640 条可用,WAPE 中位数 0.848 vs 最佳基线 0.851,FVA 0.3%,22/640 键优于基线。这条作为后续 P2 重审候选集/门槛的依据,不再用 selection 窗口的改善代替。 |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
摘要
P0 + P1,已按 Codex 两轮复审修复。未改 D5/D6/CNY。
Codex 第二轮 P1 修复(commit
e1f66ec3)SELECTION_END=202604;eligible、demand_pattern、sbc_class、models_for只用 ≤202604 的数据。3_分类增加n_obs_train/n_pos_train/demand_pattern_train/sbc_class_train,3a_分类_asof增加*_full,训练期与全历史可对照。holdout(202605..202610)不参与 eligibility/分类/模型集。current_adoptionadopted,不在 ABC 内重选;A/B/C 统一在 holdout origin202604(h=6,test 202605..202610)评估。common_keys/common_actual/common_wape/common_bias/common_stockout_rate/excluded_actual;full 与 common 并列,差量单列。本轮 A/B/C actual 均为 20,499,common actual 20,499,差量 +0;A WAPE 0.8193,B/C90/C180 0.8171。Codex 第一轮 P1/P2 修复(commit
75a9aa00)adopted_partial;同比按同月配对。验收(最终 clean run)
3_分类1749 /3a1853 /4_回测46,483(validation 27,701 + holdout 18,782)/15_趋势诊断1749adopted_sources源键)3avs4_回测(validation)独立 holdout 的诚实结果:adopted 中 holdout 可用 640 条;holdout WAPE 中位数 0.848,同批最佳基线 0.851;holdout FVA(中位数)0.3%,adopted 仅 22/640 键优于最佳基线。即 validation 选出的模型在真正独立的 6 个月里几乎没有超出基线的增量——已写进报告,作为后续 P2 重审候选集/门槛的依据。
软包墙围 4 SKU(validation / holdout WAPE)
说明
orders_unified_20261111_v2.parquet;输出:D:/Work/fzh-order-store/analysis/D3b-v3_11-20261112/。3_分类的 none 为 901:只认该键自己的 as-of 采用结果;旧版按产品历史量借模型的做法已移除。D5 消费时必须保留 none 的 fallback。index.md由scripts/update_index.py重新生成(首 commit);后续 commit 只改脚本 + ADR + log。🤖 Generated with Claude Code