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

Dragon Alpha-First Branch Report

Branches

• Evaluation window: 2016-01-01 to 2025-12-31.
• workbook_preserving: official formal baseline, preserves workbook structure as much as possible.
• alpha_first_selective_veto: research branch using the current best narrow deep-oversold veto package.

Headline Comparison

• workbook_preserving: trades 106, avg_return 2.57%, profit_factor 3.37, real BUY / SELL 105/105
• alpha_first_selective_veto: trades 102, avg_return 2.86%, profit_factor 4.04, real BUY / SELL 101/101

Short-Holding Impact

• 00-05d avg_return: workbook -2.03% vs alpha-first -1.72%
• 06-10d avg_return: workbook -0.81% vs alpha-first -0.69%

Walk-Forward Comparison

• Anchored expanding: workbook positive 8/9, avg test return 3.68%; alpha-first positive 8/9, avg test return 4.00%
• Rolling 3Y: workbook positive 7/7, avg test return 4.78%; alpha-first positive 7/7, avg test return 5.19%

Trade-Diff Summary

• trades removed from alpha-first vs workbook: 4
• trades added in alpha-first vs workbook: 0
• Key removed deep-oversold trades are the narrow pathological subset identified in Track A, not the full weak-subtype family.

Governance

• Keep workbook_preserving as the official reconstruction baseline.
• Keep alpha_first_selective_veto as the leading performance-oriented research branch.
• Do not merge alpha-first veto rules back into the official baseline unless the objective explicitly changes from workbook preservation to alpha-first optimization.

Quant Judgment

• Stage 3 is complete once both baselines are explicitly separated and reproducible.
• The workbook-preserving baseline remains the authoritative reconstruction target.
• The alpha-first branch now has a concrete candidate baseline with better trade quality and better short-holding behavior, at the cost of expected workbook alignment loss.
• Future work should choose one branch explicitly before optimizing further; the main unresolved technical decision is governance, not missing analysis.