PHASE 2 / TIME-SPLIT & ROLLING ROBUSTNESS
This page is a pressure test after the rule is frozen. It does not claim a fully untouched OOS result, because the historical sample has already been observed during research.
这一页是封板后的压力测试，不宣称为完全未触碰 OOS，因为研究过程中已经观察过历史样本。
核心策略页已经说明 top12_keep37 为什么被封板。这里不再继续找更好的参数，而是换一个问题：最终规则离开全样本汇总之后，在不同时间切片、年度环境和滚动窗口里是否仍然能被解释？
The core strategy page already explains why top12_keep37 was frozen. This page does not continue parameter search. It asks a different question: after leaving the full-sample summary, does the final rule remain explainable across time splits, calendar regimes and rolling windows?
这个检查不会把策略升级成 live-proven。它只提高研究可信度：如果弱区间集中在哪里、回撤压力来自哪里、后期样本和早期样本差异多大都能被看见，策略就更适合进入 pre-live，而不是继续在历史样本里调参。
This check does not turn the strategy into live proof. It improves research credibility: if weak periods, drawdown pressure and the difference between early and later samples are visible, the candidate is better prepared for pre-live validation rather than more historical tuning.
| Item | Setting | 中文说明 |
|---|---|---|
| Frozen candidate | top12_keep37 |
目标持仓 12 只，已有持仓若仍在前 37 名内则允许保留。 |
| Data source | Latest focused validation daily return file | 使用封板后的 daily_returns_key_variants.csv 重新计算。 |
| Research convention | Weekly rebalance, T+2 style execution, 15 bps one-way cost | 沿用核心研究口径，不为这页重新调参。 |
| Claim boundary | Time-split robustness, not untouched OOS | 这是时间切片稳健性检查，不是严格意义的全新样本外。 |
Figure 1 / 图 1. Time-split robustness
| Segment | CAGR | MDD | Calmar | How to read it |
|---|---|---|---|---|
| 2017-2020 early sample | 21.13% | -12.48% | 1.693 | Early sample is stronger and should not be treated as the whole story. |
| 2021-2026 later sample | 16.77% | -18.76% | 0.894 | Later sample is weaker but still positive under the frozen rule. |
| 2021-2025 time split | 15.86% | -18.76% | 0.846 | Useful time-split check, but not a fully untouched OOS claim. |
| 2024 credit-shock year | 16.85% | -11.10% | 1.519 | Connects back to the credit-risk and low-price-bond repricing discussion. |
这个结果的价值不在于每一段都漂亮，而在于后期样本没有被隐藏。2017-2020 更顺，2021 以后更难，2024 又包含信用冲击和低价债重估。最终规则在这些切片里仍然保持正收益轮廓，但后期回撤压力明显更真实。
The value is not that every segment looks perfect. It is that the weaker later sample is visible. 2017-2020 was easier; the post-2021 sample was harder; 2024 included credit shock and low-price-bond repricing. The final rule still keeps a positive profile, but later-period drawdown pressure is clearly more realistic.