ETF Gard · Methodology

方法论：我们怎么算、怎么防作弊 Methodology: how we compute, and how we prevent cheating

这页把方法讲清楚到「一个懂行的怀疑者可以自己复核」的程度。流程公开，参数与内部工艺不公开——怎么做公开，调到多准保密。

This page explains our method in enough detail that a sceptical practitioner can re-derive it independently. The process is public; the tuned parameters and internal tooling are not: how we do it is open, how precisely we tuned it is not.

规则版本 Rules v0.4.1 规则哈希 Hash fb72f74ec46f7d11 更新 Updated 2026-10-05 语言 ZH / EN

1 · 我们做什么、不做什么 1 · What we do, and what we never do

我们做三件事：记录历史、计算统计、公开方法。所有输出都是历史统计与结构性风险说明，动作永远留给用户。

We do three things: keep records, compute historical statistics, and publish our method. Every output is a historical statistic or a structural risk disclosure. Decisions always remain with the reader.

我们不做：推荐买卖、建议买入或卖出、给目标价、承诺收益、发出任何动作指令（含止损、逃命类表述）、提供一键跟单、发布策略信号、预测抄底或必涨稳赚。

We never: recommend securities, suggest buying or selling, set a target price, promise outcomes, issue action instructions, offer copy trading, publish trading signals, or predict market bottoms.

2 · 数据从哪来 2 · Where the data comes from

美股 ETF 日线行情（含复权、分红处理） Daily US-listed ETF prices, with adjustment and dividend handling
宏观序列（收益率曲线、美元、实际利率、CPI、信用利差）— 来自公开的中央银行／官方统计接口 Macro series (yield curve, US dollar, real rates, CPI, credit spreads) from public official statistics interfaces
波动率指数 VIX — 来自交易所公开发布的历史文件 VIX volatility index — from the exchange's public historical files
事件日历（议息会议、通胀与就业数据公布日） Event calendar (policy meetings, inflation and employment release dates)

每条事实记录带两个时间戳：event_time（数据对应的市场时间）与 available_at（该数据真正可得的时间）。所有查询按 available_at 过滤。

Every fact row carries two timestamps: event_time (the market time the data refers to) and available_at (when the data actually became available). All queries filter on available_at.

为什么这是生死线：不按可得时间过滤，就会在历史回测里用上「当时还看不到的数据」——那会让任何方法看起来都很好，却毫无意义。 Why this is the line between meaningful and meaningless: without point-in-time filtering, a backtest silently consumes data that did not exist yet at decision time. That flatters every method, and proves nothing.

We do not redistribute raw price series — our data agreements do not permit it.

3 · 我们算哪些因子（18 个） 3 · Which factors we compute (18)

ETF 是资产包，针对个股的财务报表类因子在 ETF 层面天然失效，所以我们只用「量价 + 宏观状态 + 事件 + 结构」四组。因子只列出名称与方向含义，权重与参数不公开。

An ETF is a basket, so stock-level financial-statement factors do not survive at the fund level. We therefore use four groups: price/volume, macro state, events, and structure. Names and direction only — weights and parameters are not published.

组别 Group	因子 Factors（名称与方向 name & direction）
量价组 Price / volume	动量 12-1（剔除最近一个月的中期趋势）· 动量 3-1 · 短期反转 5 日 · 波动率 60 日年化 · 下行波动率与最大回撤 · 90 日日均成交额 · 换手与份额趋势 · 200 日均线偏离 Momentum 12-1 (excluding the most recent month) · Momentum 3-1 · 5-day short-term reversal · 60-day annualised volatility · downside volatility and max drawdown · 90-day average dollar volume · turnover and share-count trend · deviation from the 200-day moving average
宏观组 Macro	收益率曲线斜率（10 年 − 2 年）· 美元指数动量 · 实际利率变化 · CPI 同比动量 · VIX 水平及变化 · 高收益信用利差 Yield-curve slope (10Y − 2Y) · US dollar momentum · change in real rates · CPI year-over-year momentum · VIX level and change · high-yield credit spread
事件组 Events	议息会议日效应 · 通胀／就业数据公布日效应 Policy-meeting-day effect · CPI and payrolls release-day effect
结构组 Structure	资产类别分组（组内相对排名的基准）· 上市年龄（由历史长度推算）与流动性代理（90 日成交额、换手与份额趋势） Asset-class grouping (the basis for within-group relative ranking) · listing age (derived from history length) and a liquidity proxy (90-day dollar volume, turnover and share-count trend)

因子池是活的：新因子必须先通过预注册与检验（见第 5 节）才可能进入，被淘汰的因子会留下记录。 The factor pool is alive: a new factor must pass pre-registration and validation (section 5) before it can enter, and retired factors leave a record.

4 · 怎么裁决 4 · How we adjudicate

我们不看「谁回测最好看」。我们用滚动的前进式竞技场（walk-forward）来判一个因子是否算数：

训练 36 个月 → 验证 12 个月 → 测试 3 个月 36 months training → 12 months validation → 3 months test
权重在验证段结束后冻结，测试段的结果不用于回头调整 Weights are frozen at the end of validation; test-period results are never used to adjust them
每季度整体向前滚动一次，用新数据重算 The whole window rolls forward every quarter and is recomputed on new data

评分形式（λ 为换手惩罚系数，其取值属内部参数）：

S(W) = t( Top 组 − Bottom 组的月度超额收益 ) − λ × 换手率
S(W) = t( monthly excess return of Top minus Bottom ) − λ × turnover

We do not rank factors by how good a backtest looks in isolation. A factor only counts if it survives a rolling walk-forward arena with a frozen, never-retuned test window.

5 · 怎么防作弊（四道工序） 5 · How we prevent cheating (four gates)

因子预注册档案制：每个因子在写代码之前先立档案（定义、方向、数据源、预期、失效条件）。没有档案的因子代码不允许合入——这一条由机器闸门强制，不靠自觉。 Factor pre-registration: every factor gets a filed record (definition, direction, data source, expectation, falsification condition) before any code is written. Code for an unfiled factor cannot be merged — enforced by an automated gate, not by good intentions.
参数集中且版本化：所有窗口、阈值、禁词集中在版本化配置里，每次变更留下提交记录；散落在代码中的魔法数字会被机器拦下。 Central, versioned parameters: every window, threshold and restricted term lives in a versioned configuration file; each change leaves a commit record. Magic numbers scattered in code are blocked by machine.
合规文本机器校验：所有对外文本（包括你正在读的这一页）必须通过禁词规则扫描才能合入。 Machine-checked compliance copy: every outward-facing text — including this page — must pass a restricted-term scan before it can be merged.
输出前的攻击性复核：报告与推送在发出前，由独立的复核流程专门尝试「攻破」它——找夸大、暗示、越界表述。提示词与阈值不公开。 Adversarial review before publication: before a report or alert goes out, an independent review pass tries to break it — hunting for overstatement, implication, and boundary-crossing language. Prompts and thresholds are not published.

6 · 怎么查账 6 · How you can audit us

承诺：对任意历史时点 T，你可以用公开数据源独立重算，再与我们的结果比对。我们提供的是转换后的统计值与哈希指纹——不提供原始行情序列（数据协议限制）。

Our commitment: for any historical date T you can recompute independently from public sources and compare against our published result. What we publish are transformed statistics and their hash fingerprints — not raw price series, which our data agreements forbid us to redistribute.

重算接口 /verify 随 M1 上线；届时本页会更新说明与用法。当前版本的承诺只到「方法与口径公开」这一层，我们不假装接口已经存在。 The /verify endpoint ships with M1; this page will be updated with usage details then. Until that exists, our commitment stops at "the method and definitions are public" — we do not pretend the endpoint is already live.

6.1 · 关注列表里看得到什么 6.1 · What a watchlist entry shows

你在关注列表里可以放进任意 ETF 代码，不只是我们池子里的。但每只标的在页面上带的身份标签不同，看到的字段也不同——这一点我们写在页面上，而不是等你发现：

池内：风险分层、告警证据（这只标的多久出一次事）、历史关照（跌成什么样、多久回来）、你所在位置的历史同类结局——全给。
观察席 / 临界 / 长尾：只给基础属性与同资产类别的池化历史，并标注样本量。「临界」还会告诉你还差多少进池。
不可添加：杠杆与反向产品，搜索时即拦截。它们不是「没排上」，是不适合当长期持有品。

两条我们主动说清楚的自我限制：① 池外标的不打风险分层标签——分层区间是用池内成员标定的，外推不成立；② 池外标的不给单只的历史事件统计——样本太薄，给了是假精确。宁可写「样本不足」，也不给一个看起来很精确的数。

另外，关注与盯守是两件事：关注看得到什么，与是否收到主动提醒，是分开的。

You may add any ETF ticker to a watchlist, not only the ones inside our screened universe. Each entry is labelled with its own status, and the fields it shows depend on that status. Two limits stated up front: we never apply pool-calibrated risk tiers to out-of-pool funds (the calibration does not transfer), and we never show single-fund event statistics for them (the samples are too thin — we would rather say "insufficient sample" than print a falsely precise number). Watching and alerting are separate things.

7 · 版本与免责 7 · Version and disclaimer

规则版本 v0.4.1 · 更新日期 2026-10-05。规则每次变更都会更新本页版本号、更新日期与规则哈希，变更内容可追溯。

Rules version v0.4.1 · updated 2026-10-05. Every change to the rules updates the version number, the date, and the rules hash on this page, so any change can be traced.

规则摘要哈希（SHA-256，覆盖数据口径、合规禁词与全部闸门定义共 5 段规则本体；不含版本号、更新日期与部署配置）：

RULES-HASH:00343063c85aed5413d1b59f636f9a961b98d4db78e858f4bb4338a8fe192ec6

Rules digest (SHA-256) over the five rule sections that define data semantics, restricted terms, and every internal gate — excluding the version label, the date, and deployment configuration. Any change to those rules changes this hash; a machine check in our build pipeline fails if this page and the live rules disagree.