cyb50-quant/research/dragon/v2/dragon_alpha_promotion_decision.md

Dragon Alpha Promotion Decision

Scope

• Evaluation window: 2016-01-01 to 2025-12-31
• Reference branches:
  • workbook_preserving: reconstruction reference
  • alpha_first_selective_veto: current formal alpha control
  • alpha_first_glued_refined_hot_cap: leading higher-alpha candidate

Current Fact Pattern

• alpha_first_selective_veto

  • trades 102
  • avg_return 2.86%
  • profit_factor 4.04
  • CAGR 28.70%
  • real BUY / SELL overlap 101/101

• alpha_first_glued_refined_hot_cap

  • trades 91
  • avg_return 3.42%
  • profit_factor 5.11
  • CAGR 31.32%
  • real BUY / SELL overlap 90/90

• refined minus current alpha

  • avg_return +0.56%
  • profit_factor +1.07
  • CAGR +2.62%
  • overlap cost -11 BUY / -11 SELL

What Has Already Been Proven

• Headline quality: PASS

  • higher avg trade
  • higher win rate
  • higher profit factor
  • better short-holding cleanup

• Stability: PASS

  • cost stress remains superior through 20 bps/side
  • local neighborhood sensitivity stays strong across 81 nearby cases
  • anchored walk-forward improves from 8/10 positive years to 9/10
  • rolling 3Y walk-forward is not worse and average test return is higher

• Removal quality: PASS

  • removed trades vs current alpha: 11
  • removed-set avg_return -1.81%
  • removed-set win rate 0%
  • removed-set review: 11 should stay removed, 0 over-removals

• Drawdown quality: improved

  • max drawdown -14.84% -> -12.79%
  • Calmar 1.93 -> 2.45

What Is Not Proven False

• The refined branch is not a knife-edge threshold artifact.
• The refined branch does not depend on deleting profitable samples.
• The refined branch does not lose its edge once reasonable costs are applied.
• The refined branch does not improve only in one tiny sample slice; the gain is concentrated in intended glued short-holding cleanup while medium/long buckets are preserved.

Real Remaining Blocker

• The only blocker is governance preference, not quant evidence.
• Automatic promotion was blocked because overlap loss exceeded the preset tolerance of 8 extra BUYs and 8 extra SELLs.
• In other words:
  • quant answer: refined is better
  • governance answer: refined is more different

Decision Tree

Option A: Keep alpha_first_selective_veto As Formal Alpha

Use this if the institution still values workbook continuity as a hard constraint.

• Pros

  • smaller divergence from workbook-style reconstructed history
  • easier to explain as a conservative formal branch
  • lower organizational change cost

• Cons

  • knowingly keeps 11 losing short-holding glued trades
  • gives up better CAGR, PF, win rate, and drawdown quality
  • keeps the weaker branch as the official alpha reference

Option B: Promote alpha_first_glued_refined_hot_cap To Formal Alpha

Use this if the objective is now explicitly "robust alpha first, workbook resemblance second".

• Pros

  • best headline return and compounding
  • best cost-stress behavior
  • best drawdown profile
  • better walk-forward behavior
  • removed set is fully explainable and low quality

• Cons

  • formal overlap drops from 101/101 to 90/90
  • governance narrative must acknowledge deliberate workbook divergence

Option C: Keep Dual Track

Use this if governance is not yet ready to bless the stronger branch as official, but research should clearly move around the stronger branch.

• Operational meaning
  • alpha_first_selective_veto stays frozen as control
  • alpha_first_glued_refined_hot_cap becomes the default forward research branch
  • all future new ideas are judged first against refined, then checked against control

Recommendation

• Governance-neutral recommendation: keep DUAL_TRACK_GOVERNANCE
• Alpha-first recommendation: promote alpha_first_glued_refined_hot_cap

Reason:

• If the question is "which branch is the better quant strategy", the answer is already alpha_first_glued_refined_hot_cap.
• If the question is "which branch is safer to keep as an official bridge from workbook reconstruction", the answer remains alpha_first_selective_veto.

Given the current research direction has already shifted toward stronger alpha, the practical operating stance should be:

• forward research default: alpha_first_glued_refined_hot_cap
• benchmark control: alpha_first_selective_veto
• reconstruction reference: workbook_preserving

Promotion Trigger Rule

Promote refined immediately if any one of the following is accepted explicitly:

• workbook overlap is a cost metric, not a hard gate
• overlap tolerance is widened from 8/8 to at least 11/11
• the official objective becomes "maximize robust alpha on 399673 under explainable rules"

Do not block promotion on any of the following, because they have already been tested:

• transaction cost pressure
• local threshold fragility
• weak walk-forward behavior
• profitable-trade over-removal
• worse drawdown quality

Execution Rule Going Forward

• Do not resume workbook-style micro-alignment tuning as the main path.
• New experiments should target:
  • whether refined can be improved further without damaging its current robustness
  • whether any new rule beats refined on quality and stability, not on workbook resemblance alone
• If a future branch beats refined, compare it primarily against refined, not against workbook.