低估值研究：历史口径覆盖与排名比较
金融模板覆盖补充：独立单日诊断恢复部分覆盖，保留本页原基线。
五种口径目前还不足以做一份全市场可比的“最低估值榜”。 2026 年 10 月 16 日，5,567 只具备有效市值的股票中，只有 1,743 只的五种分母都为正、资料齐全且报告期一致，占 31.3%。在这批股票中，PE 与 PB 的排名相关系数为 0.507，两者最低五分之一名单的交并比为 28.1%。本轮只比较覆盖和排名，没有回测收益。
面板覆盖 2016 年 1 月 29 日至 2026 年 10 月 16 日的 130 个按月形成日，共 556,948 条股票—日期记录；10 月尚未结束。下载已复核的汇总数据。各分母的定义见口径表与六家公司样本。
最近一期能算多少？
| 口径 | 可计算股票 | 占 5,567 只比例 |
|---|---|---|
| 归母净利润 TTM 对应 PE | 3,839 | 69.0% |
| 扣非归母净利润 TTM 代理 PE | 3,703 | 66.5% |
| 归母净资产 PB | 5,447 | 97.8% |
| 归母净资产减合并商誉的代理 PB | 2,650 | 47.6% |
| 归母净资产减合并商誉、无形资产的代理 PB | 2,601 | 46.7% |
| 五种均可算，报告期不限 | 1,744 | 31.3% |
| 五种均可算，且报告期一致 | 1,743 | 31.3% |
缺失商誉、无形资产不能当作零。分母非正也不算“特别便宜”。利润表、指标表和资产负债表分别选最新合格报告，避免缺一张表就把其他指标一并丢掉或退回旧期；五种口径共同排名另要求报告期一致。
金融类报表模板的覆盖尤其不足： 87 只使用金融类 comp_type 的股票都缺少可计算 PE TTM 的完整数据，没有一只进入五口径共同样本。检查发现是保留来源中的报告腿缺失、资产负债表过旧或缺失，不能解释成这些公司都在亏损。报表模板也不等于行业分类；1,743 只共同样本全部使用一般企业模板，因此本轮不能声称已完成金融与非金融行业比较。
换分母，排名会变多少？
以下均用 10 月 16 日同一批 1,743 只股票。排名相关使用 Spearman 系数；最低五分之一保留边界并列。交并比是两份名单交集除以并集，不是原组合保留比例。
| 两种口径 | 排名相关 | 最低五分之一交并比 |
|---|---|---|
| PE / 扣非 PE | 0.947 | 75.4% |
| PE / PB | 0.507 | 28.1% |
| PB / 扣商誉 PB | 0.983 | 89.7% |
| PB / 扣商誉和无形资产 PB | 0.934 | 75.8% |
按报表模板、市值五分组、ROE 五分组进一步分层，每层至少 20 只共同样本，共有 19 层合格，全部来自一般企业模板。层内相关系数中位数分别为：PE/扣非 PE 0.903，PE/PB 0.778，PB/扣商誉和无形资产 PB 0.930。这只是排名诊断，没有完成行业中性化、收益回归或因果检验，披露缺失造成的样本选择仍然明显。
历史覆盖
下表按年汇总股票—日期记录，不代表可交易覆盖或收益。
| 年份 | 形成日 | 股票—日期记录 | 五口径共同 | PE 可算 | PB 可算 | 扣商誉/无形资产 PB 可算 |
|---|---|---|---|---|---|---|
| 2016 | 12 | 31,697 | 33.0% | 65.3% | 99.5% | 53.3% |
| 2017 | 12 | 36,755 | 42.3% | 86.8% | 99.7% | 53.4% |
| 2018 | 12 | 40,514 | 43.8% | 87.4% | 98.9% | 54.4% |
| 2019 | 12 | 43,599 | 41.1% | 82.4% | 98.5% | 55.6% |
| 2020 | 12 | 46,850 | 36.7% | 78.6% | 98.0% | 52.7% |
| 2021 | 12 | 52,614 | 36.0% | 80.8% | 99.1% | 50.2% |
| 2022 | 12 | 58,244 | 34.5% | 78.9% | 99.2% | 48.7% |
| 2023 | 12 | 62,503 | 32.5% | 72.6% | 99.3% | 47.8% |
| 2024 | 12 | 64,128 | 32.3% | 71.8% | 99.3% | 47.6% |
| 2025 | 12 | 64,876 | 31.4% | 69.9% | 99.4% | 47.4% |
| 2026 截至 10 月 16 日 | 10 | 55,168 | 31.2% | 68.9% | 99.0% | 47.0% |
2016 年初还受 TTM 预热缺口影响：保留报表从 2015 年开始，缺少此前年度报告腿，不能把这种缺失解释成亏损。
历史可得性与保留版本核对
历史面板由 10 月 18 日保留版本重建。ann_date 与 f_ann_date 取较晚者，必须早于形成日；只接纳标准合并报表 report_type=1。TTM 用本期累计加上年全年、减上年同期，缺任何报告腿都不补数。每张表分别剔除报告期距形成日超过 550 个自然日的记录；实际观测时间必须存在且不在未来。
重建仍可能使用后来修订过的历史财报，不能认证为历史时点数据。另一版本再要求实际观测时间不晚于形成日上海时间 18:00，只用当时已有的合格保留快照。
| 形成日 / 保留版本 | 同一价格下的核对结果 |
|---|---|
| 8 月 31 日 / 8 月 29 日 | 原样清单校验链失败，拒绝接纳，没有用重建值冒充当时观测 |
| 9 月 28 日 / 9 月 26 日 | 84 只资产负债表报告期不同；双方都可算的股票中，25 只 PB 数值不同，两种调整 PB 各 15 只；可计算范围也不同 |
| 10 月 16 日 / 10 月 16 日 | 11 只资产负债表报告期不同；双方都可算的数值没有差异，但可计算范围不同 |
9 月实际观测约束下五口径共同样本为 1,738 只，10 月为 1,743 只。双方共同样本中的同指标排名相关为 1.0，不代表被排除的股票也正确，更不构成未接触样本外检验。
29 个保留日期目录中，4 个缺少完整顶层封存，8 个原样元数据链失败。另做只读取证后，逆向还原一次历史目录路径迁移即可重现全部 32 个原始子清单哈希；32 个底层文件哈希及父级内容聚合都一致。这支持数据内容未变，但原样存储的校验链仍然失败。本轮没有重写清单、封存或时间戳，严格加载仍拒绝这些版本。目录元数据盘点不等于核验全部历史数据文件；实际接纳参与本轮计算的输入文件另有哈希检查。
原始财报核对与边界
独立核对了分众传媒 2026 年半年报及长江电力 2026 年半年报：归母净利润、扣非归母净利润、归母权益、少数股东权益、商誉、无形资产六项均与来源数据相符到分，公告日期也相符。两份均为未经审计的半年报；两家公司一期核对不能替代全市场历史认证。
扣除项使用合并商誉、无形资产，却没有完整归母与少数股东分摊，仍然是代理口径。最近一期 3,854 只少数股东权益为正；优先股相关信息 21 只为正、5,401 只未知。分子是 A 股价格推算的总市值，没有认证多类别股份的经济总价值。本轮没有加入历史 ST、流动性或执行筛选，股票范围来自每个日期的报价记录，没有拿今天的证券清单回筛历史。排名不是交易建议。
下一步应先补齐金融类模板报表及历史保留版本，核实归母、少数股东和优先股范围，再固定收益实验。分析代码为 src/market_research/valuation_panel.py，复现入口为 scripts/analyze_valuation_panel.py。公开下载只包含复核后的汇总，供应商逐行数据留在数据所有者存储。
复现
显式提供所有者的财务版本根目录 FINANCIAL_VINTAGES、日频估值资产目录 DAILY_BASIC_ASSET 及仓库外输出根目录 RESEARCH_OUTPUT；输出目录必须尚不存在。复现前须核对汇总中的输入哈希，读取更新后的日频别名不是同一次实验。
uv run --locked --extra duckdb --extra dev python scripts/analyze_valuation_panel.py \
--vintages "$FINANCIAL_VINTAGES" \
--daily-basic "$DAILY_BASIC_ASSET" \
--reference-vintage vintage=20261018 --start 20160101 --end 20261016 \
--output "$RESEARCH_OUTPUT/valuation-panel-20261018"Low valuation: historical denominator coverage and rank comparison
Financial-template coverage follow-up: a separate single-date diagnosis restores coverage without replacing this baseline.
The five valuation definitions cannot yet support a comparable whole-market cheapest-stock list. On October 16, 2026, only 1,743 of 5,567 priced names (31.3%) have all five positive, available denominators from the same reporting period. On those names, PE and PB rank correlation is 0.507; the Jaccard overlap of their cheapest fifth is 28.1%. These are descriptive rank results, with no return backtest.
The panel contains 556,948 stock-date records at 130 monthly formations from January 29, 2016 to October 16, 2026. October is partial. Download the reviewed aggregate. The definition and six-company checks explain the original accounting choices.
Latest coverage
| Definition | Available names | Share of 5,567 |
|---|---|---|
| PE: parent profit TTM | 3,839 | 69.0% |
| PE: deducted parent profit TTM proxy | 3,703 | 66.5% |
| PB: parent equity | 5,447 | 97.8% |
| PB: parent equity less consolidated goodwill proxy | 2,650 | 47.6% |
| PB: parent equity less consolidated goodwill and intangibles proxy | 2,601 | 46.7% |
| All five available, any reporting periods | 1,744 | 31.3% |
| All five available, matching reporting periods | 1,743 | 31.3% |
Missing goodwill or intangibles are not zero. A nonpositive denominator is unavailable, not an exceptionally cheap valuation. Income, indicator and balance-sheet tables each select their latest admissible report independently, so one missing table does not erase another computable metric. Matching periods are additionally required for the five-way rank sample.
Coverage is particularly incomplete for financial reporting templates: all 87 names with financial comp_type have unavailable PE TTM and none enter the five-way sample. This reflects missing report legs and stale or missing balance sheets in the retained source, not evidence that these companies made losses. Reporting-template groups are not industry classifications; all 1,743 common names use the general-company template. Consequently, this is not a comparison across financial and nonfinancial industries.
What changes when the denominator changes?
All comparisons below use the same 1,743 names on October 16. Spearman correlation compares ranks. The cheapest fifth includes boundary ties; Jaccard is intersection divided by union, not the share of one portfolio retained.
| Pair | Rank correlation | Cheapest-fifth Jaccard |
|---|---|---|
| PE / deducted-profit PE | 0.947 | 75.4% |
| PE / PB | 0.507 | 28.1% |
| PB / PB less goodwill | 0.983 | 89.7% |
| PB / PB less goodwill and intangibles | 0.934 | 75.8% |
Within reporting-template × size-quintile × ROE-quintile strata with at least 20 common names, 19 strata qualify, all from the general-company template. Median within-stratum correlations are 0.903 for PE/deducted PE, 0.778 for PE/PB and 0.930 for PB/PB less goodwill and intangibles. These are rank diagnostics, not industry neutralization, a return regression or a causal explanation. Asset-disclosure selection remains substantial.
Historical coverage
The following percentages pool records within each year. They do not measure investability or returns.
| Year | Formations | Stock-date records | Five-metric common | PE available | PB available | PB ex G/I available |
|---|---|---|---|---|---|---|
| 2016 | 12 | 31,697 | 33.0% | 65.3% | 99.5% | 53.3% |
| 2017 | 12 | 36,755 | 42.3% | 86.8% | 99.7% | 53.4% |
| 2018 | 12 | 40,514 | 43.8% | 87.4% | 98.9% | 54.4% |
| 2019 | 12 | 43,599 | 41.1% | 82.4% | 98.5% | 55.6% |
| 2020 | 12 | 46,850 | 36.7% | 78.6% | 98.0% | 52.7% |
| 2021 | 12 | 52,614 | 36.0% | 80.8% | 99.1% | 50.2% |
| 2022 | 12 | 58,244 | 34.5% | 78.9% | 99.2% | 48.7% |
| 2023 | 12 | 62,503 | 32.5% | 72.6% | 99.3% | 47.8% |
| 2024 | 12 | 64,128 | 32.3% | 71.8% | 99.3% | 47.6% |
| 2025 | 12 | 64,876 | 31.4% | 69.9% | 99.4% | 47.4% |
| 2026 through October 16 | 10 | 55,168 | 31.2% | 68.9% | 99.0% | 47.0% |
Early 2016 also has a TTM warm-up gap because retained statement coverage starts in 2015 and lacks the preceding annual leg. This must not be interpreted as a loss-making-company count.
Availability and retained-version checks
Historical rows are reconstructed from the October 18 retained version. The later of ann_date and f_ann_date must be strictly before the formation day; only standard consolidated report_type=1 is admitted. TTM uses current YTD plus previous full year minus matching prior YTD. Missing legs remain unavailable. Reports older than 550 calendar days are rejected separately for each table. Observed timestamps must exist and not lie in the future.
This reconstruction can retain subsequently revised historical reports. It is not certified historical point-in-time data. A second variant additionally requires the actual source observation timestamp no later than 18:00 Shanghai on the formation day, using retained snapshots where admissible.
| Formation / retained version | Result at the same quoted prices |
|---|---|
| August 31 / August 29 | Rejected: on-disk manifest chain fails; no reconstructed replacement counted as observed |
| September 28 / September 26 | 84 balance-report-period differences; PB differs numerically for 25 mutually valid names, adjusted PB for 15; availability also differs |
| October 16 / October 16 | 11 balance-report-period differences; no numeric differences among mutually valid values, but availability differs |
September’s observed five-way sample contains 1,738 names; October’s contains 1,743. Same-metric rank correlations on the mutually common sample are 1.0. That does not validate excluded names or establish an untouched out-of-sample test.
Of 29 retained date directories, four lack complete top-level seals and eight fail the on-disk metadata chain. A separate read-only forensic check reproduced all 32 original child-manifest hashes by reversing only the historical path migration; all 32 underlying file hashes and parent content aggregates match. This supports unchanged file content, while the original stored chain still fails. No manifests, seals or timestamps were rewritten, and the strict loader still rejects those versions. Metadata inventory alone does not certify all archive data files; the admitted input files used here were hash-verified.
Accounting checks and remaining limits
Independent checks against the official unaudited 2026 first-half reports for Focus Media and China Yangtze Power matched all six checked fields to the cent: parent profit, deducted parent profit, parent equity, minority equity, goodwill and intangibles. The announcement dates also match. This is a two-company, one-period check, not market-wide or historical certification.
The deductions subtract consolidated goodwill/intangibles from parent equity without a complete parent/minority allocation. They remain proxies. Positive minority equity appears in 3,854 latest names; preferred-equity information is positive for 21 and unknown for 5,401. Numerators use A-price-implied total market capitalization, not a certified economic value for multiple share classes. No historical ST, liquidity or execution filter is imposed, and the universe comes from each date’s quote records rather than today’s instrument list. None of these ranks is a trading recommendation.
Next useful work is to repair financial-template statement coverage and obtain historically retained report versions, then verify parent/minority and preferred-share scope before specifying a return experiment. The analysis source is src/market_research/valuation_panel.py; the reproducible entry point is scripts/analyze_valuation_panel.py. Public output contains reviewed aggregates only; supplier rows remain in owner storage.
Reproduction
Supply owner asset paths explicitly as FINANCIAL_VINTAGES and DAILY_BASIC_ASSET, and an external RESEARCH_OUTPUT root. The output directory must not exist. Match the published input hashes before reproducing; a changed daily asset alias constitutes a new experiment.
uv run --locked --extra duckdb --extra dev python scripts/analyze_valuation_panel.py \
--vintages "$FINANCIAL_VINTAGES" \
--daily-basic "$DAILY_BASIC_ASSET" \
--reference-vintage vintage=20261018 --start 20160101 --end 20261016 \
--output "$RESEARCH_OUTPUT/valuation-panel-20261018"