方法与回测Methodology & backtest

在你相信任何信号之前，应该先问：它检验过吗？结果如何？这页如实公布我们的回测——包括不好看的部分。Before trusting any signal you should ask: was it tested, and what did the test say? This page publishes our backtest as-is — including the unflattering parts.

我们怎么测How we tested

30 只美股大盘股，逐时点前向检验：每个历史时点只用「当时能看到」的日线算因子分，再对照之后 5 / 21 / 30 个交易日相对大盘（SPY）的真实收益，共约 4,800 组样本。没有未来数据，没有事后挑选。30 large-cap US stocks, walk-forward: at each historical point the factors are computed only from data visible at that time, then compared against the actual return vs. the market (SPY) over the following 5 / 21 / 30 trading days — roughly 4,800 samples. No look-ahead, no cherry-picking.

各因子有没有预测力Does each factor predict?

信息系数（IC）衡量因子分与未来相对收益的相关性：0 是没有预测力，量化行业里 +0.05 已算有用的因子。21 个交易日窗口：The information coefficient (IC) measures how a factor's score correlates with future relative returns: 0 means no predictive power; in the quant industry +0.05 is already a useful factor. Over 21 trading days:

「超买超卖」（均值回归）的 IC 是负的——它不但没帮忙，还在帮倒忙。所以我们把它的权重降为零、从信号里剔除了。一个只想卖订阅的产品不会告诉你这些；我们宁可把引擎变小，也不装作每个因子都灵。Mean reversion's IC came out negative — it wasn't just useless, it was actively harmful. So we zeroed its weight and removed it from the signal. A product that only wants your subscription wouldn't tell you this; we'd rather shrink the engine than pretend every factor works.

窗口越长，优势越明显The edge grows with the window

方向准确率的随机基线是 50%。51–53% 看起来不起眼——这正是重点：真实的量化优势就是小而持续的，不是「十拿九稳」。这也是为什么 App 里 5 天窗的把握度天生更低：短窗噪音大，我们照实标出来，而不是把三个窗口都吹成一样准。The random baseline for direction accuracy is 50%. 51–53% looks unimpressive — and that's exactly the point: a real quantitative edge is small and persistent, not "nine times out of ten." It's also why the 5-day window shows lower confidence in the app: short windows are noisier, and we label that honestly instead of pretending all three windows are equally reliable.

这份回测的局限What this backtest can't tell you

只测了大盘股；没有计入交易成本；权重由同一段历史校准，未来的市场环境可能不同。过往表现不预示未来。正因如此，我们同时在跑公开的实盘战绩：每条信号到期后对照真实价格核对，按 5 / 21 / 30 天窗逐一填充——回测是起点，实盘核对才是持续的检验。Large caps only; trading costs not modeled; the weights were calibrated on the same history, and future regimes may differ. Past performance does not predict future results. Which is why we also run a public live track record: every signal is checked against real prices at expiry, filling in window by window — the backtest is the starting point, the live record is the ongoing test.

回测数据截至 2026-07-29 · 信息聚合与分析展示，非投资建议。Backtest as of 2026-07-29 · information & analysis, not investment advice.