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

Dragon Forward Research Plan

Core Objective

Promote alpha_first_glued_refined_hot_cap from a strong research candidate into a formal quant strategy track that is:

• explainable
• robust
• governable
• suitable for continued research and practical validation

This is no longer a workbook-reconstruction project. The main question is no longer "does it look like the workbook?" but "is this now the best explainable and robust strategy branch for 399673?"

Branch Roles

• workbook_preserving

  • role: reconstruction reference
  • use: explain workbook behavior, run regression checks, preserve reverse-engineered baseline
  • not the main optimization target

• alpha_first_selective_veto

  • role: benchmark control branch
  • use: compare current official alpha behavior against stronger branches
  • keep frozen unless governance explicitly changes

• alpha_first_glued_refined_hot_cap

  • role: forward research default
  • use: main strategy candidate for future formalization
  • primary target for all next-stage analysis

Working Principles

• Do not return to workbook-style micro-alignment as the main path.
• Do not use black-box tuning.
• Every meaningful new change must pass:
  • local attribution
  • full-sample validation
  • robustness review
• Priorities remain:
  • robustness first
  • reduce useless trades second
  • preserve large-trend profits third
  • annualized return fourth

Phase 1: Freeze Research Objective

Goal

Lock branch usage so later work does not drift back into mixed objectives.

Rules

• workbook_preserving is only for reconstruction reference.
• alpha_first_selective_veto is only for control comparison.
• alpha_first_glued_refined_hot_cap is the default forward branch.
• New headline judgments should be refined vs control, not refined vs workbook.

Status

• completed

Main Reference

• dragon_alpha_promotion_decision.md

Phase 2: Refined-Branch Alpha Attribution

Goal

Explain exactly where refined alpha comes from and why the branch is better.

Questions To Answer

• Which entry families truly create alpha?
• Which exit families actually improve trade quality?
• Are deleted trades genuinely low quality?
• Are preserved winning trades structurally consistent?
• Is the edge coming from better entries, better exits, or better interaction between both?

Work Items

• decompose all entry families, especially glued_buy
• decompose exit families and their realized trade outcomes
• fully recheck the 11 removed trades versus control
• profile the retained winning trades in refined
• run entry/exit interaction attribution
• classify rule families into:
  • core alpha families
  • support families
  • bridge / governance families
  • redundant families

Planned Outputs

• dragon_refined_alpha_attribution.csv
• dragon_refined_family_decomposition.csv
• dragon_refined_removed_trade_recheck.csv
• dragon_refined_edge_review.md

Success Criteria

• main edge source is explainable
• removed set remains structurally weak
• contribution is not dominated by one accidental sample pocket
• family roles are clearly separated

Phase 3: Execution-Aware Robustness Validation

Goal

Test whether refined remains strong under more realistic trading constraints.

Questions To Answer

• Does refined stay ahead after more realistic costs and slippage?
• Does delayed execution materially damage the edge?
• Is risk concentrated in a small set of years, regimes, or rule families?
• Is drawdown improvement real rather than cosmetic?
• Does the branch remain stable under small local perturbations?

Work Items

• cost and slippage stress
• execution-lag / next-bar execution review
• risk-cluster review
• regime / year / family / holding-bucket stability review
• local parameter-neighborhood checks around refined only

Planned Outputs

• dragon_refined_execution_stress.csv
• dragon_refined_latency_review.csv
• dragon_refined_risk_cluster_review.csv
• dragon_refined_stability_review.md

Success Criteria

• refined still outperforms control under realistic stress
• no major hidden concentration risk appears
• improvements survive local perturbation
• short-holding cleanup remains valid under execution drag

Phase 4: Formal Strategy Governance

Goal

Convert refined from a research branch into a governed strategy asset.

Questions To Answer

• What formally triggers promotion to official strategy status?
• Which parameters are frozen and which may still be researched?
• Which KPIs should be monitored continuously?
• What conditions should trigger downgrade, rollback, or forced review?

Work Items

• define promotion rules
• define parameter governance rules
• define monitoring KPIs
• define failure and downgrade triggers
• define branch management protocol

Planned Outputs

• dragon_formal_strategy_governance.md
• dragon_parameter_governance.md
• dragon_strategy_monitoring_template.csv

Success Criteria

• strategy upgrades are rule-based rather than opinion-based
• future experiments cannot silently pollute the formal branch
• monitoring and downgrade paths are explicit

Phase 5: Final Strategy Memo

Goal

Produce a concise formal memo describing the strategy as a durable research and operating asset.

Contents

• official objective
• edge source
• known weak points
• validated strengths
• practical operating rules
• what still requires future validation
• final branch recommendation

Planned Output

• dragon_formal_strategy_memo.md

Success Criteria

• future reviews can rely on one memo instead of scattered research notes
• branch choice, edge source, and governance rules are all explicit

Current Stage Target

The immediate target is not to compress a few more historical dates and not to blindly chase annualized return.

The immediate target is:

prove that alpha_first_glued_refined_hot_cap is a strategy branch with:

• clear edge source
• justified trade deletions
• stable behavior under realistic execution constraints
• explicit governance and monitoring rules

Immediate Next Step

Start with Phase 2:

1. family-level decomposition
2. removed-trade recheck
3. entry/exit interaction attribution
4. consolidated edge review

Once this is complete, the branch will move from "better result" to "explained result".