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

Dragon Refined Stability Review

Scope

• branches: alpha_first_selective_veto vs alpha_first_glued_refined_hot_cap
• execution models: same_close, next_open, next_close
• costs: 0, 5, 10, 20 bps/side

Latency Review

• same_close control vs refined: avg_return 2.86% -> 3.42%, PF 4.04 -> 5.11
• next_open control vs refined: avg_return 2.76% -> 3.31%, PF 3.78 -> 4.73
• next_close control vs refined: avg_return 1.98% -> 2.44%, PF 2.37 -> 2.78

Cost + Next-Open Stress

• next_open + 20 bps/side control CAGR 22.41% vs refined 25.51%
• next_open + 20 bps/side control PF 2.92 vs refined 3.64
• next_open + 20 bps/side control max DD -28.86% vs refined -19.08%

Risk Cluster Review

• next_open control max loss streak 10 vs refined 8
• next_open control worst 5-trade sum -10.11% vs refined -7.96%
• next_open control short-loss share 90.40% vs refined 87.99%
• next_open control worst loss family glued_buy vs refined glued_buy

Judgment

• If refined still leads after next-bar execution and cost drag, its edge is less likely to be a same-bar backtest artifact.
• If refined also keeps loss clustering and drawdown no worse than control, the branch is moving closer to a deployable research baseline.