rank183 · decomposition（baseline / components / ablation）

生成时间：2026-04-08 11:33:27 北京时间。本页服务于“把 P2/P3 策略拆到可执行组件”的需求。

← 返回该 rank 主报告 ｜ P3/P2 总表

一、baseline / components / ablation 拆分

模块	定义	为什么需要
baseline	均值回复主baseline（pairs/fair-value）	提供主 alpha 假设与对照面。
increment	cbETH-ETH rolling fair-basis MR	该 rank 相比 baseline 的独立改动点。
role	challenger	挑战/替代 baseline 的候选路线，先追求可解释增量，再看成本后存活。
stage/status	P3 / bench	决定当前是继续推进、观测，还是归档/降级。

二、信号定义（具体口径）

策略类型：均值回复

1. 输入层（Input）：价格/成交/跨资产数据，要求明确频率与对齐时点。
2. 候选触发层（Candidate Trigger）：先给出“触发候选”的条件，不直接等同于下单。
3. 过滤/否决层（Filter/Veto）：说明哪些条件会阻止候选触发变成真实交易。
4. 执行层（Execution Signal）：给出最终下单定义（方向、仓位、并发约束、冷却时间）。
5. 退出层（Exit）：至少覆盖止损、止盈、超时退出与异常风控保护。
6. 成本层（Cost）：固定费率/滑点假设 + 敏感性结论（费率上升后是否仍成立）。

简化记忆法：候选信号 ≠ 真实执行信号；只有通过过滤与执行约束后，才是可归因交易。

三、ablation 建议

• 先测 baseline only（无增量、无额外过滤）。
• 再测 baseline + increment（只加该 rank 独立增量）。
• 再测 baseline + increment + filter/veto（看收益与回撤是否同步改善）。
• 最后做 cost stress（费率/滑点提高后的存活性）。

当前下一步：先解除交易接入阻塞（Coinbase），否则维持 shelve

四、现有证据摘录

证据摘要来源：reports/artifacts/rank183_pair_proxy_scan_20260401/README.md

# Rank 183 deep-dive proxy scan (2026-04-01)

## Goal
Answer three concrete user questions:
1. Why does Rank 183 use `CBETH spot + ETH perp` instead of two generic coins on the same exchange?
2. How stable is Rank 183 historically, month by month?
3. If we switch to generic crypto pairs like `BTC/ETH`, `ETH/SOL`, `ETH/XRP`, `SOL/XRP`, can simple correlation be turned into a tradable mean-reversion strategy?

## Key conclusion
- `Rank 183` is a **basis mean-reversion** strategy, not a generic cross-asset pair trade.
- The reason it uses `CBETH spot + ETH perp` is that there is a clearer economic anchor:
  - `CBETH` is a staked-ETH spot asset with a premium/discount structure,
  - `ETH perp` is the cheap hedge leg for broad ETH beta,
  - so the trade is betting on **basis / fair-value convergence**, not just "these two charts usually move together".
- By contrast, `BTC/ETH`, `ETH/SOL`, `ETH/XRP`, `SOL/XRP` are mostly **cross-asset correlation pairs**. They may co-move, but they do not automatically have a stable fair-value anchor strong enough for post-cost mean reversion.

## Rank 183 local honest sample
Using the current artifact:
- `reports/artifacts/quant_digests/cbeth_eth_basis_probe_20260326_0850_15m/trade_log.csv`

The local trade log only spans:
- start: `2026-02-01 18:45 UTC`
- end: `2026-03-26 08:45 UTC`

So the exact current paper-spec evidence is only about **2 months**, not 1 year / 3 years / 5 years.
Also, an exact `5y` backtest on the same instrument is not literally possible because `CBETH` itself is younger than 5 years.

### Monthly split in the local trade log
- `2026-02`
  - `642` trades
  - mean net `+26.69 bps/trade`
  - win rate `98.13%`
  - median hold `1` bar
- `2026-03`
  - `533` trades
  - mean net `+12.17 bps/trade`
  - win rate `80.30%`
  - median hold `2` bars

Interpretation:
- still positive in both months,
- but March is clearly weaker than February,
- so even inside the local honest window, this is **positive but already decaying**, not a flat monthly machine.

## Generic perp-pair proxy scan
To answer the "what if we replace the underlying" question, we ran a simple **beta-hedged rolling-z mean-reversion shell** on Binance 15m pe