Volatility Alert
这一页不预测涨跌，只回答一个问题：现在是不是"可能出事"的时候。This page does not forecast direction. It answers one question: is this a moment when something could break? 依据是一条在 58 个 ETF 上反复出现的规律——真实的趋势转折，系统性地发生在高波动的日子里。告警响起时，接下来两周内出现转折的概率约为平时的 1.46–1.47 倍；It rests on a regularity that recurs across 58 ETFs — real trend turning points happen systematically on high-volatility days. When an alert fires, the chance of a turning point in the next two weeks is roughly 1.46–1.47× the usual rate; 但它不告诉你往哪个方向转（实测：被告警覆盖的转折里，开启上涨段的占 38%，而全部转折是 42%）。But it does not tell you which way the turn goes (measured: of the turning points an alert covers, 38% begin an uptrend, against 42% across all turning points). 强度与其边界见 §3 与 §4。Strength and its limits are set out in §3 and §4.
两个指标各自与该标的自身历史比较，只用截至前一交易日的信息。Each indicator is compared against the ticker's own history, using only information through the prior trading day.
A 单日冲击：刚出现了一次对该标的而言异常大的单日波动。这是事件。A · Single-day shock: a single-day move unusually large for this ticker has just occurred. This is an event.
B 波动水平：近期整体波动处在该标的历史的高位。这是环境，可以连续数月为真。B · Volatility level: recent volatility sits high in this ticker's own history. This is the environment, and it can stay true for months.
告警状态 = 近 10 个交易日内 A 触发过（最近 3 个交易日内触发标「告警」，再往前标「关注中」）。取 10 日而非只看昨天，是为了与下方证据的口径一致：那里衡量的是“告警与转折相距 10 个交易日以内”，即告警提示的是一段值得关注的时期，而不是某一天的单点判断。B 单独处于高位时只标记为高波动期，作背景参考。Alert state = A fired within the last 10 trading days (firing in the last 3 trading days is labelled “Alert”, anything earlier “Watch”). The window is 10 days rather than yesterday alone so that it matches the scope of the evidence below, which measures “an alert within 10 trading days of a turning point” — the alert marks a period worth watching, not a single-day verdict. When only B is elevated the page marks a high-volatility stretch as background, nothing more.
| 标的Ticker | 类别Category | 状态State | 最近告警Last alert | A 单日冲击A · Single-day shock | B 波动水平B · Volatility level | 当前年化波动Annualized vol | 最新价Last | 近 5 日5-day |
|---|---|---|---|---|---|---|---|---|
| USO 原油Crude Oil | 商品Commodities | 关注中Watch | 3 日前 d ago | 66% | 93% | 57% | 126.60 | +7.3% |
| PDBC 综合商品Broad Commodities | 商品Commodities | 关注中Watch | 3 日前 d ago | 74% | 91% | 25% | 17.91 | +3.8% |
| SLV 白银Silver | 商品Commodities | 关注中Watch | 3 日前 d ago | 67% | 84% | 39% | 58.48 | +1.7% |
| XLE 能源Energy | 美股行业US Sectors | 关注中Watch | 3 日前 d ago | 3% | 60% | 24% | 61.91 | +7.7% |
| GLD 黄金Gold | 跨市场核心Cross-Asset Core | 关注中Watch | 4 日前 d ago | 85% | 88% | 24% | 401.48 | +0.8% |
| GDX 金矿Gold Miners | 商品Commodities | 关注中Watch | 4 日前 d ago | 81% | 83% | 48% | 89.97 | +0.1% |
| SHY 短端国债Short Treasuries | 债券信用Bonds & Credit | 关注中Watch | 4 日前 d ago | 88% | 78% | 2% | 82.00 | +0.1% |
| ARKK 创新主题Disruptive Innovation | 高波动成长High-Volatility Growth | 关注中Watch | 4 日前 d ago | 56% | 65% | 38% | 81.10 | +2.1% |
| SMH 半导体Semiconductors | 美股行业US Sectors | 关注中Watch | 7 日前 d ago | 37% | 82% | 43% | 587.82 | +0.9% |
| EEM 新兴市场Emerging Markets | 国际International | 关注中Watch | 7 日前 d ago | 22% | 74% | 25% | 66.61 | +1.5% |
| QQQ 纳斯达克100Nasdaq-100 | 跨市场核心Cross-Asset Core | 关注中Watch | 7 日前 d ago | 65% | 65% | 23% | 731.07 | +1.1% |
| SPY 标普500S&P 500 | 跨市场核心Cross-Asset Core | 关注中Watch | 7 日前 d ago | 60% | 45% | 12% | 776.34 | +0.4% |
| AGG 综合债Aggregate Bonds | 债券信用Bonds & Credit | 静默Quiet | — | 73% | 68% | 4% | 97.48 | -0.1% |
| XLP 日常消费Consumer Staples | 美股行业US Sectors | 静默Quiet | — | 81% | 63% | 13% | 86.09 | +1.1% |
| XLV 医疗Health Care | 美股行业US Sectors | 静默Quiet | — | 4% | 54% | 14% | 167.37 | +1.0% |
| 000001.SS 上证指数Shanghai Composite | 跨市场核心Cross-Asset Core | 静默Quiet | — | 27% | 53% | 17% | 3,926.97 | +0.7% |
| ^HSI 恒生指数Hang Seng Index | 跨市场核心Cross-Asset Core | 静默Quiet | — | 53% | 51% | 18% | 25,396.51 | -0.5% |
| LQD 投资级债Investment Grade | 债券信用Bonds & Credit | 静默Quiet | — | 72% | 51% | 6% | 106.12 | -0.4% |
| XLU 公用事业Utilities | 美股行业US Sectors | 静默Quiet | — | 39% | 47% | 14% | 44.31 | +1.6% |
| XBI 生物科技Biotech | 美股行业US Sectors | 静默Quiet | — | 66% | 43% | 25% | 157.41 | +0.0% |
| FXI 中国大盘China Large-Cap | 国际International | 静默Quiet | — | 53% | 31% | 19% | 34.89 | -3.5% |
| EFA 发达市场Developed Markets | 国际International | 静默Quiet | — | 16% | 23% | 11% | 108.64 | +0.1% |
| IWM 罗素2000Russell 2000 | 跨市场核心Cross-Asset Core | 静默Quiet | — | 18% | 23% | 15% | 305.09 | +1.2% |
| KRE 区域银行Regional Banks | 美股行业US Sectors | 静默Quiet | — | 27% | 17% | 16% | 77.93 | +2.3% |
| TLT 20年期美债20+ Year Treasuries | 跨市场核心Cross-Asset Core | 静默Quiet | — | 54% | 16% | 9% | 82.04 | -0.9% |
| EWZ 巴西Brazil | 国际International | 静默Quiet | — | 13% | 16% | 21% | 33.93 | -4.0% |
| HYG 高收益债High Yield | 债券信用Bonds & Credit | 静默Quiet | — | 51% | 13% | 3% | 79.71 | +0.1% |
| XLF 金融Financials | 美股行业US Sectors | 静默Quiet | — | 43% | 10% | 10% | 58.16 | +1.0% |
| BTC-USD 比特币Bitcoin | 跨市场核心Cross-Asset Core | 静默Quiet | — | 0% | 2% | 17% | 62,820.00 | -3.1% |
| ETH-USD 以太坊Ethereum | 跨市场核心Cross-Asset Core | 静默Quiet | — | 9% | 1% | 25% | 1,877.85 | -1.6% |
红色竖线 = 单日冲击告警（事件）；橙色底色带 = 高波动期（环境，可持续数月）。Red vertical lines = single-day shock alerts (events); orange bands = elevated-volatility periods (environment, can persist for months). 黑线为对数价格。The black line is log price.告警密集处往往落在价格结构改变的位置附近。Clusters of alerts tend to fall near points where the price structure changes.
| 指标Signal | 报警日占比Alert-day share | 覆盖转折Turn coverage | 同条件零假设Same-conditions null | 净增Net | 精确率Precision | 同条件零假设Same-conditions null | 提升中位Median lift | 分子段后的中位区间Median across sub-periods |
|---|---|---|---|---|---|---|---|---|
| B 波动水平B Volatility level | 13.7% | 41% | 28% | +14% | 34% | 23% | 1.47× | 1.40 – 1.80 |
| A 单日冲击A Single-day shock | 12.5% | 84% | 78% | +6% | 34% | 23% | 1.46× | 1.40 – 1.54 |
转折点由 PT-cycle 标注器在完整历史上回溯给出，作为真值（与「标识状态」页同一套定义）；命中定义为告警日与转折点相距 10 个交易日以内。各列用同一个分母（扩展窗分位有定义之后的交易日）。Turning points come from the PT-cycle labeler run over the full history and serve as ground truth (the same definition the State Identification page uses); a hit means an alert day falls within 10 trading days of a turning point. Every column uses the same denominator (trading days on which the expanding-window percentile is defined).
为什么覆盖率也要配零假设：报警日只占一成多，而一个转折的命中窗有21 天，光靠触发率就能盖住大半转折，与信号认不认得出转折无关。把报警序列整体循环平移（触发次数与波动聚集都保住，只打散响在哪里）重算 200 次，得到上表的「同条件零假设」列。读法：只有「净增」那一列是信号自己挣来的。Why coverage needs a null too: alert days are only a little over a tenth of all days, while each turning point has a 21-day hit window — the firing rate alone already covers most turning points, whether or not the signal locates them. Rotating the whole alert series (preserving both the number of firings and the volatility clustering, scrambling only where they land) 200 times gives the “same-conditions null” columns above. Read it this way: only the “net” column is what the signal itself earned.
A 的覆盖看着高（84%），但零假设已经拿到 78%，净增只有 +6%；B 的覆盖看着低（41%），净增却有 +14%。精确率一侧两者的零假设都在 23% 上下，与“随机挑同样多的日子”吻合：逐标的胜过各自零假设的，精确率是 58/58（A）与 57/58（B），覆盖率则是 43/58（A）与 54/58（B）。A's coverage looks high (84%), but the null already reaches 78%, leaving a net of just +6%; B's coverage looks low (41%) yet nets +14%. On precision both nulls sit around 23%, matching “pick the same number of days at random”: per ticker, 58/58 (A) and 57/58 (B) beat their own null on precision, versus 43/58 (A) and 54/58 (B) on coverage.
稳健性的口径：全样本上各标的的提升几乎全部大于 1，但这不是 58 次独立确认——这些标的日收益的有效独立样本数只有 4.1（相关阵特征值的参与比），且它们共享同一段 2016–2026 的宏观历史。把样本分成前后两半后，前半有 6 个标的的提升翻到 1 以下（最差 IEF 0.65、TIP 0.83），后半没有。因此本页不以“全部通过”作为强度依据，而以提升中位在每个时间子段都稳定为依据：逐年剔除任一年、以及前后半分样本（共 12 个子段），中位都落在上表最后一列的区间内。What robustness means here: on the full sample nearly every ticker's lift exceeds 1, but that is not 58 independent confirmations — these tickers' daily returns carry only 4.1 effective independent samples (the participation ratio of the correlation matrix's eigenvalues), and they share one 2016–2026 macro history. Split the sample in half and 6 tickers in the first half fall below 1 (worst: IEF 0.65, TIP 0.83); none do in the second. So this page does not treat “all tickers pass” as strength, and instead relies on the median lift staying stable in every time sub-period: dropping any single year, plus the two halves (12 sub-periods in all), the median stays inside the range in the last column above.
两处口径差异：① 宇宙有 62 个 ETF，实际参与统计的是 58 个——AGG、BND、SHY、VUG 在标注器的有效域门上被拒（声明节奏不可达），它们没有真值，因此也没有成绩，不能计入。② 本表算在 Alpaca SIP 日线（2016 起）上，而上方 §1 的实时表格走 yfinance（历史更长，SPY 自 1993 年起）。扩展窗分位依赖历史长度，两者在重叠期上仍有约 3.5%（A）/ 5.6%（B）的日子给出不同判定（58 个两源皆有的标的的中位；逐标的最大 7.4% / 12.7%）。本表是这条规律的证据，不是上方表格逐日输出的保证。Two differences in scope that must be stated: (1) the universe holds 62 ETFs but only 58 are scored — AGG, BND, SHY, VUG are rejected by the labeler's domain gate (the declared cadence is unreachable), so they carry no ground truth and therefore no score. (2) This table is computed on Alpaca SIP daily (from 2016), while the live table in §1 runs on yfinance (a longer history — SPY from 1993). The expanding percentile depends on history length, and over the overlapping period the two still disagree on roughly 3.5% (A) / 5.6% (B) of days (median over the 58 tickers carried by both sources; up to 7.4% / 12.7%). This table is evidence for the regularity, not a guarantee about the day-by-day output of the table above.
按「净增」排，B 反而更强（覆盖净增 +14% vs A 的 +6%，精确率 34% vs 34%）。取 A 不是因为它信息量更大，而是因为用途不同：Ranked by the net column, B is the stronger of the two (coverage net +14% against A's +6%; precision 34% against 34%). A is chosen not because it carries more information but because the two serve different purposes:
B 波动水平是环境，一响可以连着几个月。作为“现在该看了”的提示，它无法界定“什么时候开始看”，故只在页面上画成底色带作背景。B, the volatility level, is an environment: once it turns on it can stay on for months. As a prompt meaning “look now” it cannot mark when to start looking, so the page draws it only as a background band.
A 单日冲击是离散事件，天然对应“昨天出事了”，能落到具体某一天。且就实际会漏掉多少转折而言 A 更少（漏约一成，B 漏一半）——注意力分配这件事上，漏报的代价大于误报。A, the single-day shock, is a discrete event — it maps naturally onto “something happened yesterday” and lands on a specific day. And in terms of how many turning points you actually miss, A misses fewer (about a tenth, against half for B) — when allocating attention, a miss costs more than a false alarm.
代价：A 的报警日里约 66% 附近并没有转折；而它高达 84% 的覆盖率里，78% 是同触发率的零假设也能拿到的，不要把它读成“这个信号认得出 84% 的转折”。State the cost in full: about 66% of A's alert days have no turning point nearby; and of its headline 84% coverage, 78% is what a null with the same firing rate already achieves. Do not read it as “this signal recognises 84% of turning points”.
· 不预测方向。被告警覆盖的转折里开启上涨段的占 38%，全部转折是 42%——在方向上等于没说。It does not predict direction. 38% of the turning points an alert covers begin an uptrend, against 42% across all turning points — nothing about direction.
· 不是买点。把告警当进场信号、持有 20 个交易日，对同敞口买入持有的超额收益中位只有 +1.80%/年，且 51/58 个标的的 90% 区间包含 0。更关键的是对照组：改用「距 250 日高点回撤最深」这一完全不含波动信息的信号、取同样多的进场日，超额收益是 +2.20%/年 —— 不低反高。⇒ 高波动这层信息在进场上没有额外贡献。Not an entry signal. Treating an alert as an entry and holding 20 trading days, the median excess return over same-exposure buy-and-hold is only +1.80%/yr, and 51/58 tickers have a 90% interval containing 0. The control matters more: swapping in “deepest drawdown from the 250-day high” — a signal carrying no volatility information at all — over the same number of entry days yields +2.20%/yr, higher, not lower. So the volatility layer adds nothing on entry.
· 会漏。按 B 指标约一半的转折发生在告警之外，按 A 指标约一成。It misses. Roughly half the turning points fall outside a B alert, and about one in ten outside an A alert.
· 覆盖率不等于信息量。A 那 84% 的覆盖里，78% 是同触发率的零假设也能拿到的，净增 +6%（见 §3）。Coverage is not information. Of A's 84% coverage, 78% is already reached by a null with the same firing rate; the net is +6% (see §3).
· 标的数不等于证据量。面板里的标的高度相关，强度依据是提升中位在各时间子段上的稳定性，不是“多少个标的都成立”（见 §3）。The number of tickers is not the amount of evidence. The tickers are highly correlated; strength rests on the median lift holding across time sub-periods, not on “it works on N tickers” (see §3).
· 证据与实时表格不同源。§3 算在 Alpaca SIP 日线（2016 起），§1 走 yfinance（历史更长），重叠期上约 3.5%（A）/ 5.6%（B）的日子判定不同（见 §3）。The evidence and the live table use different data. §3 runs on Alpaca SIP daily (from 2016), §1 on yfinance (a longer history); over the overlap they disagree on roughly 3.5% (A) / 5.6% (B) of days (see §3).
· 不含仓位建议。No position guidance.本页不产生任何买卖或仓位指令。This page issues no buy, sell, or sizing instructions.
· 加密的口径差异。Crypto follows a different convention.BTC/ETH 为 7×24 交易，年化按各标的实测 bar 频率换算；BTC/ETH trade 24/7, so annualization uses each ticker's observed bar frequency; 「10 个交易日」对它们相当于 10 个自然日，对 ETF 相当于两周。"10 trading days" means 10 calendar days for them, against roughly two weeks for the ETFs.