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

Dragon Formal Research Baseline

Scope

• Universe: 399673 only.
• Objective: preserve workbook real-trade alignment while upgrading the strategy into a researchable, testable, parameter-aware baseline.
• Current baseline type: workbook-preserving baseline.

Locked Baseline Metrics

• real BUY overlap: 106/106
• real SELL overlap: 105/105
• aux BUY overlap: 1/1
• aux SELL overlap: 19/21
• strategy trades: 107
• win_rate: 44.86%
• avg_return: 2.52%
• median_return: -0.60%
• profit_factor: 3.31

Baseline Config Snapshot

• Snapshot file: dragon_baseline_config_snapshot.json.
• Rule switches default to the current aligned strategy baseline; any future research branch should fork from this snapshot rather than editing against memory.

Core Alpha Families

• glued_buy: trades 61, avg_return 3.71%, win_rate 45.90%
• early_crash_probe_buy: trades 6, avg_return 4.62%, win_rate 66.67%
• oversold_recovery_buy: trades 4, avg_return 3.35%, win_rate 50.00%

Structural Support Families

• dual_gold_resonance_buy: trades 14, avg_return 0.84%, win_rate 42.86%
• deep_oversold_rebound_buy:classic_oversold: trades 5, avg_return 0.04%, win_rate 60.00%

Frozen Bridge Rules

• predictive_b1_break_exit: bridge-style split-chain exit; loosening worsens results, tightening breaks workbook alignment.
• predictive_error_reentry_buy: part of the same bridge chain; should be evaluated together with the predictive-break exit, not as an isolated entry.
• Any internal hold gates added only to preserve workbook-aligned split paths should remain frozen unless the objective explicitly changes away from workbook preservation.

Redundant Or Label-Only Families

• non_glued_positive_expansion_buy: now absorbed by dual_gold_resonance_buy on the same in-sample dates; treat as redundant label, not independent alpha.
• Auxiliary same-side post-exit sell compression: keep as hygiene logic, not as a primary optimization frontier.

Active Research Families

• deep_oversold_rebound_buy:positive_b1_rebound: trades 2, avg_return -2.98%, win_rate 0.00%
• post_washout_kdj_reentry_buy: trades 1, avg_return -2.86%, win_rate 0.00%
• deep_oversold_rebound_buy:shallow_false_start: trades 3, avg_return -2.63%, win_rate 33.33%
• deep_oversold_rebound_buy:deep_capitulation: trades 1, avg_return -2.05%, win_rate 0.00%
• deep_oversold_rebound_buy:mixed_oversold: trades 3, avg_return -1.80%, win_rate 33.33%
• oversold_reversal_after_ql_buy: trades 1, avg_return -0.77%, win_rate 0.00%
• post_sell_rebound_buy: trades 4, avg_return 1.84%, win_rate 25.00%

Threshold Classification

• Fragile parameters: change them only inside explicit experiment branches and always rerun full alignment diagnostics.
• predictive_b1_break_short_b1_max: avg_return_range 0.17%, min real BUY 106, min real SELL 104
• deep_oversold_entry_c1_max: avg_return_range 0.10%, min real BUY 104, min real SELL 103
• deep_oversold_entry_b1_min: avg_return_range 0.09%, min real BUY 102, min real SELL 101
• predictive_b1_break_long_b1_max: avg_return_range 0.07%, min real BUY 105, min real SELL 104
• Relatively robust parameters: acceptable first candidates for future controlled sweeps.
• aux_sell_high_zone_kdj_only_block_c1: avg_return_range 0.00%, profit_factor_range 0.00
• post_exit_confirmation_window_days: avg_return_range 0.00%, profit_factor_range 0.00
• glued_high_weak_rebound_high_c1: avg_return_range 0.09%, profit_factor_range 0.14
• glued_high_weak_rebound_high_b1: avg_return_range 0.16%, profit_factor_range 0.21

Temporal Stability

• Anchored expanding windows: positive out-of-sample years 8/10.
• Rolling 3Y windows: positive out-of-sample years 7/8.
• This validation holds the strategy fixed; it is a time-stability audit, not a refit-based optimizer.
• Strong family persistence candidates:
• early_crash_probe_buy: years_active 5, positive_years 3, negative_years 2, avg_yearly_avg_return 5.59%
• glued_buy: years_active 10, positive_years 7, negative_years 3, avg_yearly_avg_return 4.41%
• oversold_recovery_buy: years_active 4, positive_years 2, negative_years 2, avg_yearly_avg_return 3.35%
• Weak family persistence candidates:
• deep_oversold_rebound_buy: years_active 5, positive_years 1, negative_years 4, avg_yearly_avg_return -1.35%
• dual_gold_resonance_buy: years_active 7, positive_years 2, negative_years 5, avg_yearly_avg_return 0.16%
• post_sell_rebound_buy: years_active 3, positive_years 1, negative_years 2, avg_yearly_avg_return 0.52%

Operating Rules For Future Research

• Do not trade off 106/106 and 105/105 alignment silently. Any alignment loss must be treated as a branch with an explicit objective change.
• Do not blind-tune predictive-break thresholds. That family is frozen under the current baseline objective.
• Do not optimize the auxiliary layer first. The main leverage is now in weak entry-family redesign and short-holding loss control.
• New ideas should first be tested as local attribution experiments, then full-sample reruns, then temporal-stability checks.

Next Research Track

• Track A: redesign remaining deep_oversold_rebound_buy weak subtypes with delayed confirmation or fallback routing, not blunt deletion.
• Track B: explicitly target short holding buckets 00-05d and 06-10d, which remain the main quality drag.
• Track C: separate a future alpha-first research branch from this workbook-preserving baseline if the goal later changes from reconstruction to pure performance.