ashare-sop-l3-stock-resonance

name	ashare-sop-l3-stock-resonance
description	L3 stock resonance & risk control engine for the A-share SOP. Use to score individual stocks against buy/sell rules, apply theme state gates, and output 候选观察 or 风控观察 alerts with cooling mechanism.
version	1.1.0
author	Hermes Agent
license	MIT
metadata	{"hermes":{"tags":["a-share","sop","stock-scoring","risk-control","l3","buy-sell-rules"],"related_skills":["a-share-theme-structure-sop","ashare-sop-l2-theme-engine","ashare-sop-l4-execution-review"]}}

A-Share SOP L3 — Stock Resonance & Risk Control Engine

Overview

This is the L3 tier of the A-share theme structure SOP engine (v1.7+). Within the theme context from L2, it screens individual stocks using explicit buy/sell scoring rules with numerical thresholds, applies theme-state gates, and outputs "候选观察" or "风控观察" alerts.

Precondition: L2 state must NOT be 退潮期 (all candidates blocked). If L2=分歧/分化期, core pool requires buy_score ≥ 3.

Load this skill when screening individual stocks or building intraday alerts — with L2 context loaded.

Engine Position

L1 (宏观环境) → L2 (题材/阶段/池) → L3 (本模块: 个股打分/风控) → feeds → L4 (执行复盘)

Stock Pool Hierarchy (from L2)

L3 operates on 3 layers per active theme:

Pool	Description
核心股池 (Core)	Top 3 by turnover, highest consecutive boards, or market-recognized leaders
分支/补涨股池 (Branch/Laggard)	Medium turnover; technical pattern shows "回踩趋势" or "密集区突破"
风险风向标 (Risk Anchor)	High-volatility names strongly correlated with the theme, or names with huge prior gains

Theme State Gating

L2 State	Core Pool	Branch/Laggard Pool
退潮期	清空所有候选, 强制风控观察	Same — no buy candidates
分歧/分化期	buy_score ≥ 3	Normal rules
启动/发酵期 or 高潮/主升期	Normal rules	Normal rules

Buy Candidate Rules (buy_score)

Each satisfied rule increments buy_score by 1:

#	Rule	Condition
1	突破20日高点且放量	current_price > 20D_high × 1.002 AND est_turnover / 5D_avg_turnover ≥ 1.25
2	重新站上5日线并放量	current_price > MA5 AND prev_close < prev_MA5 AND est_turnover / 5D_avg_turnover ≥ 1.10
3	价格位于5/10日线上方且走强	current_price > MA5 > MA10 AND intraday_change ≥ +0.8%
4	异常放量上行	est_turnover / 20D_avg_turnover ≥ 1.8 AND intraday_change > 0
5	刷新20日高点	current_price ≥ 20D_high

Output Condition

Must satisfy BOTH:

1. buy_score ≥ 2
2. buy_score ≥ sell_score + 1 (more buy reasons than sell reasons)

→ Output: 候选观察

Risk Control Rules (sell_score)

Each satisfied rule increments sell_score by 1:

#	Rule	Condition
1	跌破10日低点且放量	current_price < 10D_low × 0.998 AND est_turnover / 5D_avg_turnover ≥ 1.20
2	跌破5日线且跌幅扩大	current_price < MA5 AND intraday_change ≤ -1.5%
3	相对昨收回撤超3%	intraday_change ≤ -3.0%
4	异常放量下行	est_turnover / 20D_avg_turnover ≥ 1.8 AND intraday_change < -1.0%

Output Condition

sell_score ≥ 2 → Output: 风控观察

Cooling Mechanism

• Per stock, same direction (buy candidate / risk alert): minimum 20-minute cooldown between alerts.
• This prevents signal flooding during volatile consolidation periods.
• Apply at the notification layer (intraday monitor, Discord alerts, cron outputs).

Data Sources for L3 Computation

# K-line derived indicators (MA5, MA10, MA20, 20D high/low, 10D low)
python3 /Users/gordon8018/.hermes/scripts/sop_enrich.py --codes sh600584,sz300750

# Turnover / volume estimates
npx -y westock-data-skillhub@1.0.3 asfund <codes>

# 120-day K-line for long-lookback indicators (via remote TDengine)
python3 /Users/gordon8018/.hermes/scripts/td_kline_fetcher.py --codes <codes> --days 120

Implementation Status

The buy_score/sell_score rules are validated against 2026 H1 TDengine K-line data. See references/backtest-2026h1-findings.md for full results.

• ✅ Factor scoring: factor_scorer.py v1.2 computes a different 5-factor scorecard (板块宽度/个股承接/量能健康/席位/资金/风险扣分) from KDJ/WR/BOLL, ATR, MainNetFlow, 龙虎榜 seats, 北向资金, 250D position
• ✅ K-line indicators: sop_enrich.py computes MA5/MA10/MA20, 20D high/low, 10D low from TDengine K-line data
• ✅ L3 buy_score/sell_score backtest: scripts/backtest_l3_rules.py validates all 9 rules against 2026 H1 daily K-line (6,839 stocks, 119 trading days). Signals show statistically significant positive excess returns across all horizons (D1–D20, t > 4.9). Critical findings documented below.
• ✅ Cooling: Must be enforced at the intraday monitor / notification layer

When building the L3 scorer:

• Use sop_enrich.py for MA5/MA10/MA20 and 20D/10D high/low
• Use westock-data asfund or tick-based estimates for turnover ratios
• Use TDengine stock.kline for the 120-day lookback window
• Integrate with existing cron pipeline: consume L2 hypothesis_queue → score → append to pool output

Backtest Validation (2026 H1)

Full backtest script: scripts/backtest_l3_rules.py. Key findings from 2026-01–2026-07 (63 signal days):

Metric	D1	D3	D5	D10
Signal mean	+0.10%	+0.41%	+0.42%	+1.22%
Signal median	-0.35%	-0.32%	-0.59%	-0.20%
Win rate	45.3%	47.0%	46.4%	48.9%
Universe mean	-0.04%	-0.03%	-0.16%	-0.48%
Excess (α)	+0.14%	+0.44%	+0.58%	+1.71%
t vs random	+4.95	+8.59	+8.37	+16.27

buy_score gradient (validated)

buy_score	N	D1	D3	D5
2	18,454	+0.06%	+0.34%	+0.37%
3	3,249	+0.22%	+0.66%	+0.59%
4	216	+1.40%	+2.08%	+2.21%

Rule effectiveness (which rules actually fire)

Rule	Universe trigger rate	In signals
R1: 突破20D高且放量	0.00% 🔴	0.0%
R2: 重站5日线且放量	6.49%	43.9%
R3: 5/10线上方走强	15.36%	89.7%
R4: 异常放量上行	5.04%	66.4%
R5: 刷新20D高点	1.04%	16.7%

Limit-up dependency (critical)

Limit-up stocks (day change ≥ 9.5%) account for 22.3% of signals but drive nearly all alpha:

• Limit-up signals: D1=+1.03%, D10=+2.92%
• Non-limit-up signals: D1=-0.16%, D10=+0.74%

Without the limit-up tail, L3 signals have no edge at D1/D3. The strategy functions as a limit-up continuation screener in practice.

Standard Output Format

## L3 个股打分
| 标的 | 角色 | buy_score | sell_score | 净分 | 输出 |
|------|------|:---------:|:----------:|:----:|------|
| XX股 | 核心 | 3 | 0 | +3 | 候选观察 ✅ |
| YY股 | 分支 | 2 | 1 | +1 | 候选观察 ✅ |
| ZZ股 | 核心 | 1 | 2 | -1 | — |
| WW股 | 风险 | 0 | 3 | -3 | 风控观察 ⚠️ |

## L3 输出汇总
- 候选观察：X 只
- 风控观察：Y 只
- 冷却中：Z 只

Common Pitfalls

1. Not gating by L2 state: Never output buy candidates when L2=退潮期, regardless of individual scores.
2. Not applying cooling: A volatile stock in consolidation can produce rapid alternating buy/sell signals. 20-min cooldown is mandatory. Backtest confirms: days with >500 signals showed consistently poor D1 returns (6/8 negative).
3. Computing MA values intraday: MA5/MA10 use daily closes. For intraday rules, estimate current MA with today's current price replacing today's close.
4. Using stale turnover estimates: "预计全天成交额" should use intraday tick volume extrapolation. Don't use yesterday's full-day turnover as today's estimate.
5. Missing the buy_score ≥ sell_score + 1 condition: A stock with buy_score=2 and sell_score=2 should NOT output 候选观察 — the signals cancel out.
6. R1 (突破20D高且放量) is DEAD: The condition close > 20D_high × 1.002 AND volume ≥ 1.25×5D_avg fired ZERO times in 2026 H1 across 6,839 stocks. In A-shares with ±10% daily limits, exceeding 20D_high by 0.2% while maintaining heavy volume is structurally impossible outside of limit-up gaps. Consider replacing with close ≥ 20D_high × 0.995 or removing R1 entirely.
7. Limit-up dependency: 22.3% of signals are limit-up stocks (day change ≥ 9.5%), and they drive nearly all alpha. Non-limit-up signals have D1 = -0.16% → essentially noise. This means the L3 rules function as a limit-up continuation screener in practice — not as a general momentum/breakout detector. Do not interpret L3 signals as standalone buy candidates outside the limit-up context.
8. Highly right-skewed returns: Mean returns are positive but median returns are negative at every horizon. The strategy relies on a few large winners (~10% of signals) to offset many small losers. This makes stop-loss discipline absolutely critical — without it, the strategy will bleed out.
9. Signal concentration is a red flag: Days with >500 signals (top 12% of days by signal count) showed poor D1 returns (6/8 negative). High signal count = indiscriminate triggering. Prefer days with focused, high-quality signals.
10. Market regime matters less than expected: Signals on down-market days actually performed better at D1/D3 (mean-reversion), while up-market signals performed better at D10+ (trend continuation). L1 gating to "多头趋势 only" may hurt short-term returns.