Task #5 IC Pack v1
1. Executive Summary
Task #5 is now in its current locked candidate form:
- **v1.2 candidate = v1.1 production + public-fund combined sponsorship bonus**
- FULL annualized **14.76%**
- FULL vol **19.97%**
- FULL Sharpe **0.739**
- FULL max drawdown **-22.34%**
- Average exposure **28.47%**
- Trades **405**
This is the best frontier point found so far under the locked risk framework. It improves both Sharpe and drawdown versus v1.1.
Canonical references:
- Decision log: output/streaming_v0/task5-decision-log.md
- Candidate summary: output/streaming_v0/v1_2_candidate/summary.md
- Stress report: output/streaming_v0/stress_tests/task5_stress_summary.md
2. Strategy Thesis
The strategy combines CANSLIM-style fundamental acceleration with Minervini-style trend discipline. The reason is capital efficiency: A-share beta alone has not historically delivered the required 15-25% return / MDD <= 25% profile, while purely discretionary "good company" selection is not repeatable enough for a production sleeve.
Core idea:
- only trade during favorable market regimes;
- only enter stocks with strong relative strength;
- only keep risk through ATR / trend / fundamental exits;
- use sponsorship as a small ranking bonus, not a hard filter.
Why CANSLIM + Minervini
CANSLIM provides the stock-selection spine: current earnings acceleration, annual earnings quality, leadership, sponsorship, and market direction.
Minervini SEPA provides the entry and risk spine: Stage 2 trend, relative strength, specific entry point discipline, and concentrated but controlled exposure.
Together they form a systematic stock-picking framework rather than a loose factor basket. This is better suited to a concentrated A-share leader portfolio than broad factor exposure, and more measurable than pure discretionary growth investing.
A-share Localization
The US source frameworks cannot be copied mechanically. Four localizations matter:
| Dimension | US original | A-share v1.2 | Rationale |
|---|---|---|---|
| Stop loss | -7% to -8% hard stop | Three-axis exit: trend, ATR, fundamentals | A-share leaders often have larger normal pullbacks; hard stops cut winners too early. |
| Trend template | Strict all-condition template | Calibrated trend template and RPS dual window | Keeps leadership quality without starving the universe. |
| M-filter | Market direction signal | Strict systemic shutoff | A-share beta dominates single-name alpha in weak regimes. |
| Sponsorship | 13F / institutional ownership | Public-fund quarterly holdings bonus | Public-fund holdings are slower but closer to durable sponsorship; LHB was too noisy. |
v1.2 Definition
v1.2 is the first candidate where all accepted improvements are incremental and data-backed:
- v1 foundation: CANSLIM + Minervini + M-filter + three-axis exit + friction/cap realism;
- v1.1: add B3 dual-window RPS (RPS250 > 80 and RPS125 > 80);
- v1.2: add public-fund combined sponsorship bonus using market-value and holding-count scores.
Rejected overlays are explicitly archived in the decision log: B1/B2/B4/B5, O'Neil S1-S6, and LHB sponsorship.
Role In Portfolio
Task #5 is a stock-picking alpha sleeve, not an asset-allocation framework.
At portfolio level, it should complement Task #4:
- Task #5 seeks concentrated A-share leader alpha and naturally goes to cash in weak regimes.
- Task #4 is the all-weather beta/risk-premium sleeve and should carry defensive assets when equity regimes are poor.
- Task #8 / Task #10 should eventually decide cross-sleeve cash and risk allocation; v1 keeps each strategy's cash management separate.
3. Backtest Results
Progression
| Version | FULL annualized | FULL Sharpe | FULL MDD | Avg exposure |
|---|---|---|---|---|
| v0 baseline | 18.00% | 0.880 | -21.78% | 29.64% |
| v1 production | 13.69% | 0.678 | -24.41% | 29.29% |
| v1.1 production | 14.34% | 0.716 | -23.42% | 28.24% |
| v1.2 candidate | 14.76% | 0.739 | -22.34% | 28.47% |
OOS / Stress
- OOS-A stays at 0% by design because the strict M-filter blocks weak regimes.
- Stress windows 2015 / 2018 / 2024 all show 0 exposure and 0 MDD in this candidate because the regime filter stays off.
- The 2015 window is explicitly caveated: stock-level candidate history does not extend back far enough to show individual stock exit logs for that period.
Rule-level verdicts
- Adopt: B3 RPS250 + RPS125 dual-window.
- Reject: B1, B2, B4, B5.
- Reject: O'Neil sell S1-S6.
- Reject: LHB sponsorship bonus/filter.
- Accept as candidate: public-fund combined sponsorship bonus.
4. Risk Narrative
The risk system has three layers:
1. **Single-name three-axis exit**
- ATR stop
- trend break
- fundamental disproof
2. **Profit protection**
- breakeven after +20%
- MA50 trail after +50%
- MA100 trail after +100%
3. **Portfolio regime control**
- strict HS300 M-filter
- industry cap <= 30%
- winner trim at 15%
This risk stack is calibrated for A-share leader behavior rather than copied from US microstructure rules.
The strict M-filter is the core protection layer. The system is supposed to stay out when the market is bad.
Three-axis Exit Rationale
The exit system uses three independent reasons to leave a position:
1. **Trend break**: MA50 / MA200 / trend-template deterioration.
2. **Volatility stop**: initial stop at entry_price - 2.5 * ATR20, then profit-protection trailing.
3. **Fundamental disproof**: EPS yoy deterioration or similar fundamental invalidation.
This replaces the US-style hard -7% to -8% stop. The A/B tests showed that O'Neil sell overlays mostly either did not trigger or sold winners too early. The current framework lets volatility-adjusted trend and fundamental evidence decide exits.
Industry Cap
The 30% Shenwan L1 industry cap remains because sensitivity testing showed it adds concentration control with little alpha cost. It is not the same as the failed Task #4 sleeve cap experiment; here it was tested inside the Task #5 stock-picking system and improved deployability.
M-filter And Exposure
Average exposure of 28.47% is a feature, not a defect. It is the result of regime filtering.
- Bullish windows can carry much higher exposure.
- Bearish or unstable windows can go close to cash.
- 2018 and 2024 stress windows show 0 exposure / 0 MDD because the M-filter was already off.
Cash during weak regimes is part of the strategy's risk control. It should not be interpreted as idle capital until the portfolio-level allocation layer decides how to redeploy it.
Target Positioning
| Risk target | Target | v1.2 result | Status |
|---|---|---|---|
| Annualized return | 15-25% | 14.76% | Near lower bound |
| Max drawdown | <= 25% | -22.34% | Pass |
| Sharpe | > 0.5 implicit | 0.739 | Pass |
| Stress windows | Defensive behavior | 0 exposure / 0 MDD | Pass with 2015 caveat |
The return is slightly below the nominal target lower bound, but this is the cost of strict regime protection. Given the drawdown improvement and stress behavior, this is acceptable for an observation-stage candidate.
5. Decision Asks
1. Approve **v1.2 candidate** as the Task #5 final working candidate.
2. Approve an observation allocation.
- Suggested starting size: RMB 1.5-3.0m, subject to final Task #8 portfolio sizing.
- Scale trigger: 3-month live/paper deviation within acceptable band; 6-month pass before full scale.
3. Approve monitoring cadence.
- Daily: exits, limit events, abnormal drawdown, execution issues.
- Weekly: candidate scan, RPS ranking, concentration review.
- Monthly: M-filter, industry cap, performance/risk review.
- Quarterly: public-fund sponsorship refresh, IC review, decision-log update.
4. Keep Task #4 and Task #5 cash/risk management separate for v1.
- Task #5 cash from M-filter shutoff stays inside Task #5 for now.
- Cross-sleeve redeployment should wait for Task #8 / Task #10 portfolio-level design.
6. Iteration Plan
Potential v1.3+ ideas, only if they improve the frontier:
- refine public-fund sponsorship:
- distinguish new entrants from continuing holders;
- test fund-manager-level weighting versus fund-level aggregation;
- combine research/roadshow signal with holdings signal if reliable data is available;
- integrate northbound sponsorship more carefully;
- add sector momentum confirmation if it improves the frontier;
- test ATR-stop sensitivity inside the existing three-axis exit;
- refine slippage and liquidity models using live/paper execution deviations.
Do not add overlays that do not clear the drawdown line or that over-filter the universe.
Explicitly out of scope for Task #5 v1.3:
- high-frequency trading;
- pair trading / market-neutral overlays;
- large discretionary overlays;
- hard individual-stock stop loss copied from US rules;
- Task #4-specific items such as ERC, bond curve monitoring, or dynamic risk allocation.
7. Appendix
Reference files:
- [Task #5 decision log](./task5-decision-log.md)
- [v1.2 candidate summary](./v1_2_candidate/summary.md)
- [Stress test summary](./stress_tests/task5_stress_summary.md)
- [Public fund sponsorship A/B](./fund_sponsorship_ab/task5_fund_sponsorship_ab_summary.md)
- [LHB sponsorship A/B](./sponsorship_ab/task5_v0plus_sensitivity_summary.md)
- [Per-rule A/B](./per_rule_ab/task5_per_rule_ab_summary.md)