Risk Forecast
前两页认出的东西，能拿来做什么？What can you actually do with what the first two pages identify?这一页给出实测后唯一站得住的答案：This page gives the only answer that survived testing: 能说清今天的风险有多大，不能说该买还是该卖。It can tell you how much risk there is today; it cannot tell you whether to buy or sell. 下表给出每个标的明天收盘的预测区间——它不预测方向，只预测幅度。The table below gives a prediction interval for each ticker's next close — it forecasts magnitude, not direction.
95% 区间意为历史上有 95% 的交易日，次日收盘落在这个范围内。A 95% interval means the next close fell inside this range on 95% of historical trading days. 区间宽度随当前波动伸缩——波动高时自动放宽，平静时收窄。Its width tracks current volatility — wider when volatility is elevated, narrower when markets are quiet. 仓位上限一列是按"能承受组合单日回撤 1%"反推的（算法见 §2）。The position cap column is backed out from a tolerated single-day portfolio drawdown of 1% (arithmetic in §2).
| 标的Ticker | 类别Category | 最新价Last | 波动分位Vol percentile | 95% 区间95% interval | 95% 下行downside | 仓位上限Position cap | 99% 下行99% downside | 样本外破位 95/99Out-of-sample breach 95/99 |
|---|---|---|---|---|---|---|---|---|
| SPY 标普500S&P 500 | 跨市场核心Cross-Asset Core | 776.34 | 46% | 763.37 – 789.53 | −1.7% | 60% | −2.4% | 5.0% / 1.4% |
| QQQ 纳斯达克100Nasdaq-100 | 跨市场核心Cross-Asset Core | 731.07 | 65% | 709.26 – 753.55 | −3.0% | 34% | −3.9% | 5.3% / 1.8% |
| IWM 罗素2000Russell 2000 | 跨市场核心Cross-Asset Core | 305.09 | 20% | 299.43 – 310.86 | −1.9% | 54% | −2.4% | 5.1% / 1.9% |
| TLT 20年期美债20+ Year Treasuries | 跨市场核心Cross-Asset Core | 82.04 | 17% | 81.09 – 83.00 | −1.2% | 86% | −1.6% | 4.9% / 1.4% |
| GLD 黄金Gold | 跨市场核心Cross-Asset Core | 401.48 | 86% | 388.72 – 414.66 | −3.2% | 31% | −4.7% | 4.9% / 0.8% |
| 000001.SS 上证指数Shanghai Composite | 跨市场核心Cross-Asset Core | 3,926.97 | 47% | 3,833.31 – 4,022.91 | −2.4% | 42% | −3.5% | 4.7% / 1.3% |
| ^HSI 恒生指数Hang Seng Index | 跨市场核心Cross-Asset Core | 25,396.51 | 41% | 24,812.78 – 25,993.97 | −2.3% | 44% | −3.4% | 4.7% / 0.9% |
| BTC-USD 比特币Bitcoin | 跨市场核心Cross-Asset Core | 62,820.00 | 2% | 61,164.69 – 64,520.11 | −2.6% | 38% | −4.9% | 4.9% / 0.7% |
| ETH-USD 以太坊Ethereum | 跨市场核心Cross-Asset Core | 1,877.85 | 1% | 1,808.03 – 1,950.36 | −3.7% | 27% | −6.0% | 5.1% / 0.8% |
| SMH 半导体Semiconductors | 美股行业US Sectors | 587.82 | 85% | 554.89 – 622.70 | −5.6% | 18% | −8.0% | 6.0% / 1.2% |
| XLF 金融Financials | 美股行业US Sectors | 58.16 | 11% | 57.26 – 59.07 | −1.5% | 65% | −2.1% | 4.2% / 1.6% |
| XLE 能源Energy | 美股行业US Sectors | 61.91 | 63% | 59.93 – 63.96 | −3.2% | 31% | −4.3% | 5.3% / 1.4% |
| XLV 医疗Health Care | 美股行业US Sectors | 167.37 | 64% | 163.81 – 171.01 | −2.1% | 47% | −3.2% | 5.2% / 1.0% |
| XLU 公用事业Utilities | 美股行业US Sectors | 44.31 | 39% | 43.50 – 45.14 | −1.8% | 54% | −2.6% | 4.4% / 1.2% |
| XLP 日常消费Consumer Staples | 美股行业US Sectors | 86.09 | 71% | 84.42 – 87.80 | −1.9% | 51% | −2.9% | 5.4% / 1.0% |
| KRE 区域银行Regional Banks | 美股行业US Sectors | 77.93 | 15% | 76.34 – 79.56 | −2.0% | 49% | −3.2% | 5.2% / 1.1% |
| XBI 生物科技Biotech | 美股行业US Sectors | 157.41 | 46% | 152.25 – 162.74 | −3.3% | 31% | −4.7% | 5.4% / 0.6% |
| EEM 新兴市场Emerging Markets | 国际International | 66.61 | 80% | 64.29 – 69.01 | −3.5% | 29% | −5.0% | 5.3% / 1.1% |
| EFA 发达市场Developed Markets | 国际International | 108.64 | 37% | 106.74 – 110.57 | −1.7% | 57% | −2.4% | 5.3% / 1.4% |
| FXI 中国大盘China Large-Cap | 国际International | 34.89 | 21% | 34.04 – 35.76 | −2.4% | 41% | −3.3% | 4.7% / 1.3% |
| EWZ 巴西Brazil | 国际International | 33.93 | 18% | 32.98 – 34.91 | −2.8% | 36% | −4.0% | 5.7% / 1.2% |
| HYG 高收益债High Yield | 债券信用Bonds & Credit | 79.71 | 9% | 79.38 – 80.04 | −0.4% | 100% | −0.7% | 4.9% / 0.8% |
| LQD 投资级债Investment Grade | 债券信用Bonds & Credit | 106.12 | 50% | 105.36 – 106.88 | −0.7% | 100% | −1.0% | 5.4% / 1.0% |
| AGG 综合债Aggregate Bonds | 债券信用Bonds & Credit | 97.48 | 63% | 96.97 – 97.99 | −0.5% | 100% | −0.8% | 5.5% / 1.0% |
| SHY 短端国债Short Treasuries | 债券信用Bonds & Credit | 82.00 | 74% | 81.83 – 82.18 | −0.2% | 100% | −0.3% | 5.6% / 1.5% |
| SLV 白银Silver | 商品Commodities | 58.48 | 82% | 55.55 – 61.56 | −5.0% | 20% | −8.0% | 5.5% / 0.9% |
| USO 原油Crude Oil | 商品Commodities | 126.60 | 93% | 117.72 – 136.15 | −7.0% | 14% | −10.2% | 5.3% / 1.2% |
| GDX 金矿Gold Miners | 商品Commodities | 89.97 | 84% | 84.31 – 96.01 | −6.3% | 16% | −8.6% | 5.3% / 1.5% |
| PDBC 综合商品Broad Commodities | 商品Commodities | 17.91 | 90% | 17.31 – 18.53 | −3.3% | 30% | −4.7% | 3.8% / 1.0% |
| ARKK 创新主题Disruptive Innovation | 高波动成长High-Volatility Growth | 81.10 | 66% | 77.00 – 85.42 | −5.1% | 20% | −7.1% | 3.6% / 0.9% |
区间只说幅度、不说方向，所以它能回答的是"多大"，不是"往哪"。The interval speaks to magnitude, not direction, so it answers "how far", never "which way". 下面三件事只需要幅度。The three uses below need magnitude alone.
止损放进 95% 区间之内，会被日常波动打掉，与判断对错无关。要让止损只在“确实不对”时触发，它必须放在区间之外。A stop placed inside the 95% interval gets taken out by ordinary noise, regardless of whether the view was right. For a stop to fire only when something has genuinely gone wrong, it has to sit outside the interval.
今日 USO：95% 下沿 117.72（−7.0%）、99% 下沿 113.63（−10.2%）。Today, USO: the 95% lower edge is 117.72 (−7.0%), the 99% edge 113.63 (−10.2%).
| 止损挂在Stop placed at | 实测被打掉的频率Measured hit frequency |
|---|---|
| 95% 下沿lower edge | 每 19 个交易日一次（≈ 0.9 个月）once every 19 trading days (≈ 0.9 months) |
| 99% 下沿lower edge | 每 67 个交易日一次（≈ 3.2 个月）once every 67 trading days (≈ 3.2 months) |
频率由留出段上直接数命中得到（30 个标的的中位）：只计下沿被触及，且以盘中最低价为准——区间按收盘定，而止损是被日内插针打掉的。The frequencies come from counting actual hits over the hold-out segment (median across 30 tickers): only the lower edge counts, and it counts when the intraday low reaches it — the interval is set on closes, but a stop is taken out during the session.
先定你能承受的组合单日回撤，再除以该标的的 95% 下行幅度，Fix the single-day portfolio drawdown you can tolerate, then divide it by this ticker's 95% downside, 得到不突破这个容忍度的最大权重。and you have the largest weight that stays within that tolerance.
仓位上限 = 可承受单日回撤 ÷ 该标的 95% 下行position cap = tolerated single-day drawdown ÷ the ticker's 95% downside
按 1% 算：USO 下行 −7.0% ⇒ 上限 14%；SHY 下行 −0.2% ⇒ 上限 100%（已封顶）。两者的 95% 下行相差 35 倍——同样一条容忍度，落到仓位上天差地别。At 1%: USO has a −7.0% downside ⇒ cap 14%; SHY has −0.2% ⇒ cap 100% (capped). Their 95% downsides differ by 35× — one tolerance, wildly different position sizes.
把持仓权重乘以各自的 95% 下行再相加，得到组合层的单日 95% 亏损上界Multiply each holding's weight by its 95% downside and sum the results: an upper bound on the portfolio's single-day 95% loss （忽略分散化，故是保守估计）。(diversification is ignored, so the estimate is conservative).
组合单日 95% 亏损 ≈ Σ 权重 × 该标的 95% 下行portfolio single-day 95% loss ≈ Σ weight × that ticker's 95% downside
这个数会随市场波动自动变化——同一组权重，在平静期与动荡期的风险不是一回事。The number updates on its own as volatility changes — the same set of weights carries very different risk in a quiet period than in a turbulent one. 表格每日更新，重算一次即可。The table updates daily; just recompute.
三件事都只用"幅度"，不用"方向"。它们不构成买卖建议——只是把同一个预测区间换成三种可直接使用的形式。All three use magnitude only, never direction. They are not buy or sell advice — just the same prediction interval expressed three ways you can act on.
判据只有一个：把收益除以预测的波动之后，还剩多少"波动聚集"？There is a single test: after dividing returns by the forecast volatility, how much volatility clustering is left? 剩得越少，说明预测越吃住了真实的波动结构。The less that remains, the more of the real volatility structure the forecast has captured.下表比的是三种做法。The table compares three approaches.
| 标准化方式Normalization | |z| 自相关 lag1|z| autocorr lag1 | lag5 | lag21 | 说明Notes |
|---|---|---|---|---|
| 原始收益 |r|Raw returns |r| | +0.265 | +0.258 | +0.162 | 未做任何归一化no normalisation at all |
| 除以无条件波动后Divided by unconditional volatility | +0.259 | +0.251 | +0.157 | 最平凡的替代：长期历史波动the most trivial alternative: long-run historical volatility |
| 除以冻结的常数后Divided by a frozen constant | +0.265 | +0.258 | +0.162 | 把波动冻成一个数；除以常数不改变相关系数，故与首行必然相同volatility frozen to a single number; dividing by a constant cannot change a correlation, so this row must equal the first |
| 除以条件波动后Divided by conditional volatility | +0.012 | +0.013 | -0.006 | 本页用的预报the forecast this page uses |
|z| 自相关衡量"大波动是否扎堆"。数值接近 0 表示波动聚集已被吸收。30 个标的的中位数，四行算在同一批日子上。逐标的看，条件波动的 lag1 比无条件更接近 0 的有 30/30 个。The autocorrelation of |z| measures whether large moves cluster; a value near 0 means the clustering has been absorbed. Medians across 30 tickers, with all four rows computed on the same set of days. Per ticker, conditional volatility lands closer to 0 at lag 1 than unconditional on 30/30 of them.
最平凡的替代——直接用历史波动——对波动聚集的吸收率是零：The most trivial alternative — plain historical volatility — absorbs none of the clustering:
归一化之后的自相关与原始收益几乎一模一样。after normalizing, the autocorrelation is all but identical to that of raw returns.条件波动率把它吸干净了。
The conditional volatility forecast removes it entirely.
这是本项目全部测试里，唯一一处"平凡替代做不到、模型能做到"的地方。Across every test in this project, this is the only place where the trivial alternative fails and the model does not.
上表的第三行说明：光看自相关，冻结版与原始收益无法区分。真正把两者分开的是区间到底准不准——同一条流水线，只把条件波动换成一个常数（仅用前段估计），重算破位率：The third row above shows that on autocorrelation alone the frozen version is indistinguishable from raw returns. What actually separates them is whether the interval is calibrated — same pipeline, conditional volatility swapped for a single constant (estimated on the earlier segment only), breach rates recomputed:
| 用哪种波动定区间Which volatility sets the interval | 样本外破位率（名义 5%）Out-of-sample breach rate (nominal 5%) | （名义 1%）(nominal 1%) | 更贴近名义值的标的数Tickers closer to nominal |
|---|---|---|---|
| 冻结成一个常数Frozen to a constant | 2.50% | 0.53% | 1/30 · 4/30 |
| 条件波动（本页）Conditional volatility (this page) | 5.22% | 1.13% | 29/30 · 26/30 |
冻结版把区间画得系统性偏宽（破位率远低于名义值），也就是把风险说小了。区间校准这件事上，随时间变化的那部分是必需的。
同一个"波动可预测"，用来定区间成立，用来提高风险调整后收益不成立（见 §4）——两者判据不同。The frozen version draws intervals that are systematically too wide (breach rates far below nominal), which understates the risk. For calibrating an interval, the time-varying part is necessary.
The same fact that volatility is predictable holds up for setting intervals and fails for improving risk-adjusted return (see §4) — the two are judged on different criteria.
· 它给你要的风险，不给你更好的收益。按波动倒数配权、按波动调整止损宽度、波动率目标化、风险预算配权、波动率入场滤网，五种「拿波动去承担风险」的用法都测过，全部败给各自最平凡的替代（前两项：对等权 Sharpe 0.96 vs 1.14；对固定百分比止损 ΔSharpe −0.01，区间跨零。这组数出自本页之外的回测层，一次性测于 2026-08-09，不随本页每日重算）。①② 声称的只是把单日下行控制在你说的幅度内，验收标准是 §3 的破位率，不是收益。It gives you the risk you asked for, not better returns. Inverse-volatility weighting, volatility-scaled stops, volatility targeting, risk-budget weighting and volatility entry filters — all five ways of bearing risk with a volatility forecast were tested, and every one lost to its own most trivial alternative (first two: Sharpe 0.96 against 1.14 for equal weight; ΔSharpe −0.01 against a fixed-percentage stop, interval spanning zero. These figures come from the backtest layer outside this page, measured once on 2026-08-09; they are not recomputed daily). ① and ② claim only to keep single-day downside within the size you named; the acceptance test is the breach rate in §3, not return.
· 不预测方向。No directional forecast.区间是对称的；模型对涨跌没有任何判断。The interval is symmetric; the model takes no view on up versus down.
· 只在日频证实。Validated at daily frequency only.20 个交易日的不重叠区间、以及组合层的区间预报，Non-overlapping 20-day intervals, and interval forecasts at the portfolio level, 实测过度保守（破位率显著低于名义），尚未证实。come out clearly conservative (breach rates well below nominal) and remain unconfirmed.
· 区间会被突破。The interval will be breached.名义 95% 意味着每 20 个交易日里预计有 1 天落在区间外，Nominal 95% means roughly 1 day in every 20 falls outside the interval, 且这些破位日本身仍有轻微聚集。and the breach days themselves still cluster slightly.
· 不含仓位建议。No position advice.本页不产生任何买卖或仓位指令。This page issues no trade or sizing instructions.