cyb50-quant/research/dragon/v2/memory/2026-04-04.md

2026-04-04

• 用户请求：分析 dragon/v2 目录下文件，总结交易数据。
• 目录内实际只有一个 Excel 文件：龙泉回测20260109.data.xlsx。
• 工作簿包含两个 sheet：
• 龙泉：交易节点、阶段估值、年度收益汇总
• 焚诀(规则)：策略规则文字说明
• 关键统计结论：
• 初始资金 55,450
• 2025-12-31 已实现资金 910,785.96
• 2016-2025 累计收益约 +1542.54%，资金约 16.43x
• 对应 CAGR 约 32.30%
• 用户确认表内 21 个未重新开仓就再次出现的卖点只是看空信号标记，不是真实交易卖点
• 另有 1 个持仓中再次买入记录，当前先归类为持仓中辅助看多信号
• 已输出 true_trade_events.csv、true_trades.csv、auxiliary_signals.csv、trade_split_summary.md
• 按真实交易口径，可配对 105 笔完整交易，胜率约 44.8%，均值收益约 +3.15%
• 后续分析结论：
• 短持有交易（0-10 天）整体显著亏损，11 天以上才开始明显转正
• KDJ/QL双金叉、刀口舔血2、带明确信号标签的买点质量明显高于未标注买点
• 空仓后的重复卖点看空标记有一定信息量，到下一次真实买点前平均还能多跌约 4.1%，但只有约 60% 有效，更适合做仓位/等待过滤，不适合作为硬性 veto
• 用户补充确认：
• C1 与 C4 为同一指标的不同名称
• QL凤凰线 的 B/S 是上穿/下穿的计算结果
• 本轮重构只针对 399673
• 持仓中再次出现的 BUY 视为辅助看多信号
• 已完成第一阶段代码实现并生成：
• dragon_workbook.py
• dragon_indicators.py
• dragon_state_machine.py
• dragon_validate.py
• dragon_indicator_snapshot.csv
• dragon_workbook_layers.csv
• dragon_signal_alignment.csv
• dragon_validation.md
• 当前验证结果：
• 明确标记口径下，KDJ 对齐 197/197 的买点、196/197 的死叉；唯一死叉错位日为 2018-05-23
• QL 的 B/S 标记对齐 112/112
• 已实现 dragon_strategy.py 与 dragon_backtest.py，完成第一版可执行规则树与基线回测
• 当前基线贴合度：
• 真实买点：工作簿 106，策略 97，重合 78
• 真实卖点：工作簿 105，策略 96，重合 64
• 辅助买点：工作簿 1，策略 48，重合 0
• 辅助卖点：工作簿 21，策略 195，重合 11
• 当前基线交易统计：96 笔，胜率约 42.71%，平均收益约 +2.23%，中位数约 -0.61%
• 当前最明显的问题：
• 买入规则过度偏向 glued_buy，共 77 笔
• 卖出规则过度偏向 knife_take_profit_2_glued，共 70 笔
• 说明 见好就收1/2、高位大行情退出、预警减仓点、股灾兜底等高位退出规则仍未充分编码
• 后续推进：
• 已补入 crash_protection_exit，2025-10-14 已成功命中真实卖点
• 已生成 dragon_event_gaps.csv 与 dragon_event_gaps.md，用于系统化分析工作簿未命中/额外命中事件
• 当前关键偏差集中在：
• 2019-03-25 预警减仓点
• 2020-03-09 / 2020-07-24 / 2020-08-07 的大行情高位退出
• 2021-06-16 / 2021-07-16 的“大行情第一次不卖、第二次确认再卖”逻辑

• 2022-07-08 / 2022-07-11 的预警减仓与正式卖点拆分

• Later refinement on dragon_strategy.py:

• added explicit high-regime exit handling before generic glued exits

• changed B1 negative count from daily accumulation to sell-signal-node accumulation

• blocked premature ql-only high-zone exits such as 2021-06-03

• recovered workbook-style key events: 2019-03-25, 2020-03-09, 2020-07-24, 2021-06-16, 2021-06-21, 2021-07-16, 2022-07-08, 2022-07-11, 2022-07-15

• current fit after rerun:

• real BUY overlap 78/106

• real SELL overlap 70/105

• aux BUY overlap 0/1

• aux SELL overlap 15/21

• strategy trades 97, win rate 44.33%, avg return 2.42%, median return -0.59%

• Later refinements after more rule work:

• added ql_high_zone_take_profit / ql_mid_zone_take_profit

• added deep_oversold_rebound_buy and oversold_recovery_buy

• added super_hot_trend_hold to preserve 2025-10-14 crash-protection exit

• latest verified fit:

• real BUY overlap 96/106

• real SELL overlap 88/105

• missing real BUY down to 10

• missing real SELL down to 17

• strategy trades 118, win rate 44.07%, avg return 2.04%, median return -0.61%

• important recovered dates in this stage:

• buys: 2022-02-09, 2022-02-15, 2022-03-11, 2022-03-16, 2022-04-27, 2022-09-26, 2022-11-01, 2023-06-12

• sells: 2020-05-22, 2021-11-18, 2024-03-22, 2022-02-11, 2022-02-22, 2022-03-15, 2022-03-25, 2022-05-24, 2022-09-30, 2022-11-10, 2023-06-21, 2025-10-14

• Final edge-pattern refinements in this session:

• fixed same-bar buy/sell execution bug in run()

• added oversold_reversal_after_ql_buy

• added oversold_rebound_take_profit

• added predictive_b1_break_exit and predictive_error_reentry_buy

• latest verified fit:

• real BUY overlap 98/106

• real SELL overlap 91/105

• missing real BUY down to 8

• missing real SELL down to 14

• remaining misses are concentrated in early 2016-2019 samples

• recovered late-sample dates:

• 2023-08-28, 2023-09-07, 2023-10-12, 2024-11-28, 2024-11-29

• Final pass today closed all remaining real-trade gaps in dragon/v2.

• Root causes were three overly broad ql-only exit paths and one non-standard post-washout reentry day.

• Implemented:

• narrow low_zone_wait_kdj_confirmation to 14<=c1<30

• add glued_mid_zone_wait_kdj_confirmation

• add oversold_low_zone_wait_kdj_confirmation

• add post_washout_kdj_reentry_buy

• allow _buy_decision() evaluation while flat even on non-buy-signal days, so workbook-style special reentry can be captured

• final verified fit after serial rerun:

• real BUY overlap 106/106

• real SELL overlap 105/105

• aux BUY overlap 1/1

• aux SELL overlap 15/21

• strategy trades 130, win rate 42.31%, avg return 1.84%, median return -0.77%

• Later quality optimization on the same day:

• added buy_block_glued_high_weak_rebound with c1 > 68 and b1 < -0.08 to cut false glued entries in high weak rebound states

• compressed flat-state auxiliary sell output via _should_emit_aux_sell()

• aux sell now only emits for:

• repeated bearish confirmation within 10 days after a real sell

• or high-zone renewed weakness (c1 > 80)

• or kdj_sell with c1 > 60 and b1 < -0.05

• corrected dragon_backtest.py to trim strategy evaluation to workbook min/max dates, not just min date

• latest verified fit after these changes:

• real BUY overlap 106/106

• real SELL overlap 105/105

• extra real BUY 14

• extra real SELL 14

• aux BUY overlap 1/1

• aux SELL overlap 18/21

• strategy aux SELL 94 vs prior 181

• strategy trades 120, win rate 44.17%, avg return 2.10%, median return -0.59%

• Continued late-session optimization cut remaining extra real trades from 14+14 down to 5+5 without breaking full real-trade alignment.

• Current verified fit:

• real BUY overlap 106/106

• real SELL overlap 105/105

• extra real BUY 5

• extra real SELL 5

• aux BUY overlap 1/1

• aux SELL overlap 18/21

• strategy aux SELL 89

• strategy trades 111, win rate 45.05%, avg return 2.32%, median return -0.60%

• Remaining extra real BUY dates:

• 2020-10-28, 2023-02-13, 2023-05-10, 2023-10-26, 2024-11-27

• Remaining extra real SELL dates:

• 2020-10-23, 2022-12-27, 2023-05-19, 2023-11-10, 2024-11-22

• Important lesson: some apparent extra buys are state-bridge entries that preserve later workbook-aligned sells, so pruning must always be verified on downstream sell alignment, not just on local buy precision.

• Phase 1 attribution outputs created:

• dragon_residual_attribution.py

• dragon_residual_trade_attribution.csv

• dragon_residual_trade_review.md

• Current residual recommendation split:

• delete candidates: 2020-10-23, 2022-12-27, 2023-05-10, 2023-05-19

• keep bridge: 2020-10-28

• keep alpha: 2023-10-26, 2023-11-10

• observe bridge: 2023-02-13, 2024-11-27

• observe: 2024-11-22

• Phase 2 residual cleanup was then executed directly in dragon_strategy.py.

• Implemented three narrow filters:

• block the weak oversold_reversal_after_ql_buy pattern around 23<c1<26, b1>-0.12, a1>-0.035

• narrow early_failed_rebound_exit from any b1<0 to mild weakness -0.05<b1<0

• add a ql-only follow-through hold before knife_take_profit_2_glued for the narrow 2020-10-23 pattern

• A follow-up hot-zone hold mid_hot_wait_ql_confirmation was added after the first rerun to restore the workbook-aligned 2020-11-11 sell that had shifted to 2020-11-10

• Latest verified fit after serial rerun:

• real BUY overlap 106/106

• real SELL overlap 105/105

• extra real BUY 2: 2023-10-26, 2024-11-27

• extra real SELL 2: 2023-11-10, 2024-11-22

• aux BUY overlap 1/1

• aux SELL overlap 18/21, strategy aux SELL 85

• strategy trades 108, win rate 44.44%, avg return 2.49%, median return -0.61%

• Net result: all four prior DELETE_CANDIDATE residual rows were removed, and the remaining residual set is now only KEEP_ALPHA plus OBSERVE/OBSERVE_BRIDGE.

• Phase 3 residual cleanup then targeted the last bridge chain 2024-11-22 -> 2024-11-27 -> 2024-11-28 -> 2024-11-29.

• Full-history uniqueness checks showed:

• the long-held ql-only fade pattern around 2024-11-22 was unique

• the long-held predictive break pattern around 2024-11-28 was also unique

• Implemented:

• long_glued_wait_predictive_break to hold through the premature 2024-11-22 ql-only exit

• a long-held extension of predictive_b1_break_exit so the same position exits on 2024-11-28

• Latest verified fit after serial rerun:

• real BUY overlap 106/106

• real SELL overlap 105/105

• extra real BUY 1: 2023-10-26

• extra real SELL 1: 2023-11-10

• aux BUY overlap 1/1, strategy aux BUY 48

• aux SELL overlap 18/21, strategy aux SELL 85

• strategy trades 107, win rate 44.86%, avg return 2.52%, median return -0.60%

• 2024-11-27 is no longer a real extra buy; it is now a holding-period auxiliary bullish signal.

• Residual review now contains only one KEEP_ALPHA pair: 2023-10-26 / 2023-11-10.

• Phase 4 auxiliary-signal optimization then started on 2026-04-04.

• Added audit tooling and outputs:

• dragon_aux_signal_audit.py

• dragon_aux_signal_audit.csv

• dragon_aux_signal_review.md

• Aux BUY optimization:

• added strength-gated holding reconfirmation logic plus cooldown in dragon_strategy.py

• reduced strategy aux BUY from 48 to 19 while preserving workbook overlap 1/1

• Aux SELL optimization:

• initially tested a broader redesign, then rolled back to the more robust original trunk

• kept only duplicate-cooldown suppression plus a narrow state_crash_followthrough path

• recovered workbook aux sell 2025-10-17

• also suppressed premature high-zone kdj-only aux sells so workbook dates 2019-04-11 and 2021-07-26 now align

• Latest verified fit after serial rerun:

• real BUY overlap 106/106

• real SELL overlap 105/105

• aux BUY overlap 1/1, strategy aux BUY 19

• aux SELL overlap 19/21, strategy aux SELL 82

• strategy trades 107, win rate 44.86%, avg return 2.52%, median return -0.60%

• Current bottleneck from audit:

• the main remaining unmatched auxiliary cluster is SELL / post_exit_confirmation with 53 unmatched rows

• future compression should be based on grouping by real-sell cycle, not only simple date cooldowns

• User approved sequence: finish Stage 2 first, then move into Stage 3 robustness validation.

• Stage 2 structural outputs were completed and verified:

• dragon_trade_path_trace.csv

• dragon_rule_taxonomy.md

• dragon_aux_sell_cycle_audit.csv

• dragon_aux_sell_cycle_summary.csv

• dragon_aux_sell_cycle_review.md

• dragon_trade_path_trace.py was corrected so Layer 4 links both in-position aux BUY and post-exit aux SELL until the next real BUY, matching the agreed auxiliary-layer definition.

• A narrow cycle-aware aux-sell dedupe was then added to dragon_strategy.py:

• within post_exit_confirmation, only the first same-side signal (ql or kdj) is emitted in a flat cycle

• this keeps workbook-style ql -> kdj double confirmation cycles but removes same-side repeat noise

• Latest verified fit after rerun:

• real BUY overlap 106/106

• real SELL overlap 105/105

• extra real BUY 1: 2023-10-26

• extra real SELL 1: 2023-11-10

• aux BUY overlap 1/1, strategy aux BUY 19

• aux SELL overlap 19/21, strategy aux SELL 75

• strategy events 307, strategy trades 107

• win rate 44.86%, avg return 2.52%, median return -0.60%

• Cycle compression effect:

• repeated post_exit_confirmation cycles 10 -> 3

• zero-anchor repeated cycles 8 -> 1

• remaining protected double-confirmation cycles are 2022-07-08 -> 2022-09-26 and 2018-12-06 -> 2019-01-18

• remaining zero-anchor repeated candidate is 2023-12-04 -> 2024-01-18

• Stage 3 was then started with a first executable robustness pack.

• Added dragon_robustness_report.py.

• Generated:

• dragon_trade_quality.csv

• dragon_trade_group_summary.csv

• dragon_yearly_performance.csv

• dragon_rule_contribution_entry.csv

• dragon_rule_contribution_exit.csv

• dragon_rule_stability.csv

• dragon_robustness_report.md

• First results:

• baseline 107 trades, avg MFE 6.66%, avg MAE -2.66%, avg exit followthrough 5d -2.95%

• short holding buckets remain the main weakness:

• 00-05d win rate 0%, avg return -2.03%

• 06-10d avg return -0.81%

• long holding buckets contain most alpha:

• 21-40d avg return 6.33%

• 41d+ win rate 100%, avg return 24.13%

• later sample is stronger than earlier sample:

• 2016-2020 avg return 1.95%, profit factor 2.83

• 2021-2025 avg return 3.15%, profit factor 3.81

• current attribution focus for next optimization round:

• entry side: deep_oversold_rebound_buy

• exit side: knife_take_profit_2_glued

• report explicitly notes the next engineering gap: parameterize hard-coded thresholds before formal perturbation and true leave-one-rule-out reruns

• Stage 3 continued and delivered the next three modules in this same session.

• Module A:

• added dragon_strategy_config.py

• parameterized dragon_strategy.py so DragonRuleEngine(config=...) can run experiments without changing default baseline behavior

• default config was verified to preserve:

• real BUY overlap 106/106

• real SELL overlap 105/105

• aux BUY overlap 1/1

• aux SELL overlap 19/21, strategy aux SELL 75

• strategy trades 107, avg return 2.52%

• Module B:

• added dragon_rule_ablation.py

• generated dragon_rule_ablation.csv and dragon_rule_ablation.md

• rerun ablation findings:

• glued_buy is still the backbone; disabling it collapses overlap to BUY 46, SELL 52

• deep_oversold_rebound_buy is the clearest weak entry family:

• disabling it raises avg return by about +0.54%

• but breaks overlap to BUY 93, SELL 92

• predictive_b1_break_exit and some oversold rebound/reentry rules look mildly weak, but current evidence is not enough to cut them directly because alignment still degrades

• disabling knife_take_profit_2_glued caused no top-level metric change in rerun form, suggesting replacement by alternate same-cycle exits

• Module C:

• added dragon_threshold_perturbation.py

• generated:

• dragon_threshold_perturbation.csv

• dragon_threshold_perturbation.md

• dragon_threshold_sensitivity_summary.csv

• first perturbation pack conclusions:

• fragile parameters:

• predictive_b1_break_short_b1_max

• predictive_b1_break_long_b1_max

• deep_oversold_entry_c1_max

• deep_oversold_entry_b1_min

• robust parameters:

• post_exit_confirmation_window_days

• aux_sell_high_zone_kdj_only_block_c1

• glued_high_weak_rebound_high_c1

• glued_high_weak_rebound_high_b1

• Module D:

• added dragon_stability_report.py

• generated dragon_stage3_stability_report.md

• current quantitative judgment:

• short holding buckets remain the main drag

• glued_buy remains core alpha structure

• optimization leverage is now concentrated in weak oversold entry logic and fragile predictive-break exit thresholds

• auxiliary sell layer is no longer the main optimization frontier

• Deep-oversold focus work was then completed in the same session.

• Added localized audit artifacts:

• dragon_deep_oversold_audit.py

• dragon_deep_oversold_audit.csv

• dragon_deep_oversold_subtype_summary.csv

• dragon_deep_oversold_review.md

• Audit result:

• all 18 deep-oversold trades are workbook-aligned

• weakest subtypes are positive_b1_rebound and shallow_false_start

• relatively better subtype is classic_oversold

• Strategy research granularity was improved without changing behavior:

• deep_oversold_rebound_buy now emits subtype-tagged reasons

• a family-aware _entry_reason_is(...) helper was added so subtype labels do not break downstream sell logic

• Verified after rerun:

• real BUY overlap 106/106

• real SELL overlap 105/105

• aux BUY overlap 1/1

• aux SELL overlap 19/21, strategy aux SELL 75

• strategy trades 107, avg return 2.52%

• Practical conclusion:

• next optimization should target subtype-aware redesign of positive_b1_rebound and shallow_false_start, not the full deep-oversold family

• One more subtype-aware pass was then executed.

• Default safe optimization applied:

• shallow-false-start entries with ql_buy are now routed to same-day fallback rules instead of remaining inside the deep-oversold family

• very light positive-B1 rebound with a1 > -0.02 is also routed to same-day fallback

• Rerouted dates:

• 2018-01-02 -> glued_buy

• 2018-10-19 -> dual_gold_resonance_buy

• 2023-06-02 -> dual_gold_resonance_buy

• 2024-09-11 -> glued_buy

• Verified after rerun:

• real BUY overlap 106/106

• real SELL overlap 105/105

• aux BUY overlap 1/1

• aux SELL overlap 19/21, strategy aux SELL 75

• strategy trades 107, avg return 2.52%

• deep-oversold family count reduced from 18 to 14

• Added focused experiment outputs:

• dragon_deep_oversold_experiments.py

• dragon_deep_oversold_experiments.csv

• dragon_deep_oversold_experiments.md

• Focused experiment results show the remaining weak subtype blocking is not acceptable under current objective:

• block remaining positive-B1 rebound -> BUY 104, SELL 103

• block remaining shallow-false-start without ql -> BUY 103, SELL 102

• block both -> BUY 101, SELL 100

• Conclusion: safe cleanup is mostly exhausted; any further deep-oversold optimization now requires accepting alignment loss or inventing more complex replacement logic

• Continued Stage 3 optimization then removed one redundant weak entry label from the effective in-sample decision tree.

• In dragon_strategy.py, non_glued_positive_expansion_buy was narrowed to non-dual_gold cases only.

• This pushed the two historical in-sample occurrences (2016-07-13, 2020-04-30) onto the later dual_gold_resonance_buy fallback branch on the same dates.

• Verified after serial rerun:

• real BUY overlap 106/106

• real SELL overlap 105/105

• aux BUY overlap 1/1

• aux SELL overlap 19/21, strategy aux SELL 75

• strategy trades 107, avg return 2.52%, profit factor 3.31

• Structural implication:

• dual_gold_resonance_buy count increased from 11 to 14

• non_glued_positive_expansion_buy now has no remaining in-sample trades and should be treated as a redundant label, not an independent alpha family

• dragon_stage3_stability_report.md was updated to record this judgment.

• Stage 3 was then formally closed into a research baseline package instead of continuing blind residual tuning.

• Added dragon_walk_forward_validation.py.

• Generated:

• dragon_walk_forward_summary.csv

• dragon_walk_forward_family_year.csv

• dragon_walk_forward_family_stability.csv

• dragon_walk_forward_report.md

• Walk-forward conclusions on the fixed current baseline:

• anchored expanding windows positive out-of-sample years 8/10

• rolling 3Y windows positive out-of-sample years 7/8

• strongest persistent family remains glued_buy; early_crash_probe_buy also remains strong but smaller-sample

• weakest family persistence remains deep_oversold_rebound_buy; dual_gold_resonance_buy and post_sell_rebound_buy are secondary / weaker persistence families

• Added dragon_research_baseline.py.

• Generated:

• dragon_baseline_config_snapshot.json

• dragon_formal_research_baseline.md

• Formal baseline classification now fixed as:

• core alpha families: glued_buy, early_crash_probe_buy, oversold_recovery_buy

• structural support families: dual_gold_resonance_buy, deep_oversold_rebound_buy:classic_oversold

• frozen bridge rules: predictive_b1_break_exit, predictive_error_reentry_buy, plus related internal hold gates that only preserve workbook split paths

• redundant label: non_glued_positive_expansion_buy

• active research families: remaining weak deep-oversold subtypes, post_washout_kdj_reentry_buy, oversold_reversal_after_ql_buy, post_sell_rebound_buy

• Current stage judgment:

• the workbook-preserving reconstruction baseline is now formalized

• next work should target either:

• Track A: subtype-aware redesign of weak deep-oversold entries

• Track B: short-holding drag reduction in 00-05d / 06-10d

• not further predictive-break tuning or aux-layer micro-optimization

• Continued Track A work then implemented a switchable deep-oversold confirmation branch in dragon_strategy.py and dragon_strategy_config.py, with default behavior still unchanged.

• New config controls:

• deep_oversold_confirm_weak_with_ql

• deep_oversold_confirm_window_bars

• Added pending-state support for weak deep-oversold subtype setups so alpha-first experiments can delay entry until later QL confirmation without touching the formal baseline.

• Added dragon_deep_oversold_confirmation_experiments.py.

• Generated:

• dragon_deep_oversold_confirmation_experiments.csv

• dragon_deep_oversold_confirmation_event_changes.csv

• dragon_deep_oversold_confirmation_experiments.md

• dragon_deep_oversold_confirmation_review.md

• Key result:

• QL confirmation for weak deep-oversold subtypes improves alpha-first branch metrics somewhat (avg_return up to about 2.73%, profit_factor up to about 3.66)

• but this improvement comes mainly from vetoing weak trades, not from turning them into strong confirmed winners

• 2022-03-11, 2024-01-18, 2024-01-23 are clean veto wins

• delayed confirmed entries such as 2022-03-17 and 2024-08-30 still remain losing trades

• Quant implication:

• for future alpha-first research, weak deep-oversold redesign should prioritize selective veto / removal before adding more confirmation complexity

• for the workbook-preserving baseline, no change should be applied from this branch

• Track A then moved from delayed-confirmation to narrow selective-veto experiments.

• Added more alpha-first experiment hooks in dragon_strategy_config.py / dragon_strategy.py:

• deep_oversold_selective_positive_b1_c1_max

• deep_oversold_selective_shallow_c1_min

• deep_oversold_selective_shallow_b1_min

• deep_oversold_selective_mixed_c1_max

• deep_oversold_selective_mixed_require_no_ql

• Added dragon_deep_oversold_selective_veto_experiments.py.

• Generated:

• dragon_deep_oversold_selective_veto_experiments.csv

• dragon_deep_oversold_selective_veto_event_changes.csv

• dragon_deep_oversold_selective_veto_experiments.md

• dragon_deep_oversold_selective_veto_review.md

• Key result:

• best branch is selective_veto_plus_mixed_c1_lt_10_2_no_ql

• metrics: 103 trades, avg_return about 2.81%, profit_factor about 3.96

• this is better than broad blocking of all remaining weak subtypes (102 trades, avg_return about 2.78%, profit_factor about 3.76)

• and also loses fewer aligned dates

• Best branch removes four clearly weak deep-oversold trades:

• 2022-02-09

• 2022-03-11

• 2024-01-18

• 2024-01-23

• while preserving the more borderline chains:

• 2022-03-16

• 2024-08-29

• Current quant judgment:

• Track A priority is now clear:

• narrow pathological-pattern veto > delayed QL confirmation > broad subtype deletion

• workbook-preserving baseline still unchanged

• alpha-first branch best current candidate is the selective-veto package, not the confirmation package

• Stage 3 was then formally closed with branch-level governance and completion artifacts.

• Added dragon_alpha_first_baseline.py.

• Generated:

• dragon_alpha_first_baseline.md

• dragon_alpha_first_config_snapshot.json

• dragon_alpha_first_branch_summary.csv

• dragon_alpha_first_branch_comparison.csv

• dragon_alpha_first_branch_holding_buckets.csv

• dragon_alpha_first_branch_walk_forward.csv

• dragon_alpha_first_branch_trade_diff.csv

• Alpha-first candidate baseline fixed as alpha_first_selective_veto with:

• trades 103

• avg_return about 2.81%

• profit_factor about 3.96

• real BUY / SELL overlap 102 / 101

• short-holding buckets improved versus workbook baseline:

• 00-05d about -2.03% -> -1.72%

• 06-10d about -0.81% -> -0.69%

• walk-forward also improved modestly while keeping the same positive-year counts:

• anchored 8/10 positive years, average test return improved

• rolling 3Y 7/8 positive years, average test return improved

• trade diff vs workbook baseline is simple and interpretable:

• removed only 4 pathological deep-oversold trades

• added 0 new trades

• Added final closure document:

• dragon_stage3_completion.md

• Final stage judgment:

• Stage 1 complete

• Stage 2 complete

• Stage 3 complete

• Current governance:

• workbook_preserving is the authoritative reconstruction baseline

• alpha_first_selective_veto is the leading performance-oriented research branch

• future work must explicitly choose the branch before continuing optimization

• Follow-up short-holding governance review completed after Stage 3 closure.

• Confirmed dragon_short_holding_master_review.md already existed and was valid.

• Added alpha_first_glued_selective_veto in dragon_branch_configs.py plus formal review script dragon_glued_alpha_candidate.py.

• Generated:

• dragon_glued_alpha_candidate_summary.csv

• dragon_glued_alpha_candidate_comparison.csv

• dragon_glued_alpha_candidate_holding_buckets.csv

• dragon_glued_alpha_candidate_walk_forward.csv

• dragon_glued_alpha_candidate_trade_diff_vs_alpha.csv

• dragon_glued_alpha_candidate_trade_diff_vs_workbook.csv

• dragon_glued_alpha_candidate_config_snapshot.json

• dragon_glued_alpha_candidate_review.md

• Candidate result:

• trades 92

• avg_return 3.35%

• profit_factor 4.95

• real BUY / SELL overlap 90/89

• short buckets improved further:

• 00-05d -1.67%

• 06-10d -0.58%

• walk-forward improved too:

• anchored 9/10 positive, avg test return 3.86%

• rolling 3Y 7/8 positive, avg test return 4.78%

• Governance conclusion:

• do not replace current formal alpha-first baseline yet

• keep alpha_first_selective_veto as official alpha-first branch

• keep alpha_first_glued_selective_veto as next research branch because overlap loss is much larger (102/101 -> 90/89)

• Trade-diff detail:

• vs alpha-first, glued candidate removes 12 trades and adds 1

• all 12 removals are glued_buy

• added trade is fallback reroute 2021-11-22 -> 2021-11-30, dual_gold_resonance_buy -> small_positive_a1_declining:kdj_sell

• Continued immediately with glued-branch residual attribution instead of blind retuning.

• Added:

• dragon_glued_veto_attribution.py

• dragon_glued_veto_attribution.csv

• dragon_glued_veto_bucket_summary.csv

• dragon_glued_veto_review.md

• Attribution result on the 12 removed trades vs alpha_first_selective_veto:

• removed-set avg_return -1.60%

• win_rate 8.33%

• profit_factor 0.03

• bucket split:

• low_weak_range: 4 trades, all losing, avg_return -2.35%, avg_holding 3.0, clean promotion candidate

• hot_positive_b1: 8 trades, 7 keep-veto + 1 observe-veto, avg_return -1.23%, avg_holding 6.1

• only borderline sample is:

• 2021-11-05 -> 2021-11-18, glued_buy -> ql_mid_zone_take_profit, return +0.69%

• this suggested narrowing hot veto instead of rolling back glued veto globally

• Added configurable hot-cap support:

• glued_selective_hot_c1_max in dragon_strategy_config.py

• corresponding condition update in dragon_strategy.py

• Added focused refinement pack:

• dragon_glued_refine_experiments.py

• dragon_glued_refine_experiments.csv

• dragon_glued_refine_experiments.md

• dragon_glued_refined_candidate_config.json

• dragon_glued_refined_trade_diff_vs_full.csv

• added branch helper alpha_first_glued_refined_hot_cap_config() in dragon_branch_configs.py

• Refined glued candidate result:

• glued_veto_hot_cap75_and_low

• trades 92

• avg_return 3.36%

• profit_factor 4.98

• real BUY / SELL overlap 91/90

• This is better than the prior full glued candidate on both performance and overlap.

• Minimal diff vs prior full glued branch:

• add back 2021-11-05 -> 2021-11-18, glued_buy -> ql_mid_zone_take_profit

• remove fallback reroute 2021-11-22 -> 2021-11-30, dual_gold_resonance_buy -> small_positive_a1_declining:kdj_sell

• Updated stage judgment:

• alpha_first_selective_veto still remains the formal alpha-first baseline

• next research branch should now be the refined glued candidate, not the old full glued candidate

• Completed the formal upgrade-review pack for the refined glued branch.

• Added:

• dragon_glued_refined_branch_review.py

• dragon_glued_refined_branch_summary.csv

• dragon_glued_refined_branch_comparison.csv

• dragon_glued_refined_branch_walk_forward.csv

• dragon_glued_refined_branch_trade_diff.csv

• dragon_glued_refined_holding_breakdown.csv

• dragon_glued_refined_yearly_breakdown.csv

• dragon_glued_refined_family_breakdown.csv

• dragon_glued_refined_regime_breakdown.csv

• dragon_glued_refined_branch_config_snapshot.json

• dragon_glued_refined_branch_review.md

• Result for alpha_first_glued_refined_hot_cap:

• trades 92

• avg_return 3.36%

• profit_factor 4.98

• avg_MFE 7.57%

• avg_MAE -2.57%

• real BUY / SELL overlap 91/90

• walk-forward:

• anchored 9/10, avg test return 3.87%

• rolling 3Y 7/8, avg test return 4.79%

• relative to current alpha_first_selective_veto:

• avg_return +0.55%

• profit_factor +1.02

• overlap -11/-11

• removed trades 11, added trades 0

• Governance conclusion from the review:

• quality and walk-forward gate now pass

• but promotion still should not be silent because the overlap loss remains material

• recommended status remains:

• keep alpha_first_selective_veto as formal alpha-first baseline

• keep alpha_first_glued_refined_hot_cap as the leading next alpha-first candidate

• Also completed refined removed-trade attribution.

• Added:

• dragon_glued_refined_removed_trade_attribution.py

• dragon_glued_refined_removed_trade_attribution.csv

• dragon_glued_refined_removed_trade_review.md

• Removed-set result vs current alpha-first:

• removed trades 11

• avg_return -1.81%

• win_rate 0%

• profit_factor 0.00

• recommendation mix:

• KEEP_REMOVAL 11

• OBSERVE_REMOVAL 0

• OVER_REMOVAL 0

• The refined branch now removes only losing short-holding glued trades from the current alpha-first baseline; no profitable sample remains in the removed set.

• Completed final alpha-branch governance package.

• Added:

• dragon_alpha_branch_governance.py

• dragon_alpha_branch_governance_matrix.csv

• dragon_alpha_branch_governance_decision.json

• dragon_alpha_branch_governance.md

• Final gate results:

• headline_quality_gate = PASS

• stability_gate = PASS

• removal_quality_gate = PASS

• alignment_cost_gate = FAIL

• explicit deltas of refined candidate vs current formal alpha:

• avg_return +0.55%

• profit_factor +1.02

• BUY overlap -11

• SELL overlap -11

• removed trades 11, added trades 0

• Final governance decision:

• DUAL_TRACK_GOVERNANCE

• Meaning:

• alpha_first_selective_veto stays as the formal alpha-first baseline

• alpha_first_glued_refined_hot_cap becomes the governed leading candidate, but not auto-promoted because overlap loss exceeds the automatic-promotion tolerance

• User then explicitly supported the stronger alpha-first view and asked for the next research direction and detailed plan under that stance.

• Planning implication:

• next roadmap should treat alpha_first_glued_refined_hot_cap as the main forward candidate and alpha_first_selective_veto as the benchmark control, rather than continuing workbook-style fine-tuning.

• Executed the first full formalization-validation pack under the stronger alpha-first direction.

• Added:

• dragon_research_direction_update.md

• dragon_forward_branch_roles.md

• dragon_cost_stress_test.py

• dragon_cost_stress_test.csv

• dragon_cost_stress_test.md

• dragon_glued_refined_sensitivity.py

• dragon_glued_refined_sensitivity.csv

• dragon_glued_refined_sensitivity_summary.csv

• dragon_glued_refined_sensitivity.md

• dragon_glued_refined_year_regime_review.py

• dragon_glued_refined_year_regime_review.csv

• dragon_glued_refined_regime_review.csv

• dragon_glued_refined_holding_review.csv

• dragon_glued_refined_family_review.csv

• dragon_glued_refined_year_regime_review.md

• dragon_equity_curve_review.py

• dragon_equity_curve_review.csv

• dragon_drawdown_review.csv

• dragon_monthly_return_review.csv

• dragon_yearly_return_review.csv

• dragon_equity_curve_review.md

• Core validation conclusions:

• cost stress:

• refined remains ahead at every tested cost level

• at 20 bps/side, current alpha CAGR 23.55%, refined CAGR 26.62%

• local sensitivity:

• neighborhood cases 81

• avg_return range 3.23% -> 3.42%

• PF range 4.66 -> 5.11

• robust-nearby cases 52

• year/regime:

• improvements concentrate in short-holding cleanup and glued_buy

• medium and long buckets are preserved

• equity curve:

• refined is best on growth and drawdown quality together

• CAGR 31.32%

• max DD -12.79%

• Calmar 2.45

• Updated stage judgment:

• the refined alpha candidate now looks quantitatively robust across:

• cost pressure

• local threshold perturbation

• year/regime consistency

• drawdown quality

• the only remaining blocker is governance preference on workbook divergence, not lack of quantitative evidence

• Added dragon_alpha_promotion_decision.md.

• This document converts the prior "governed candidate" status into an explicit recommendation tree:

• keep alpha_first_selective_veto if workbook continuity is still a hard governance constraint

• promote alpha_first_glued_refined_hot_cap if the objective is explicitly robust alpha first

• otherwise keep dual-track governance but use refined as the forward research default and selective-veto as the benchmark control

• Practical conclusion after this step:

• quant evidence now points clearly to refined as the better strategy branch

• non-promotion is now a governance choice, not an evidence gap

• Added dragon_forward_research_plan.md so the next-stage roadmap is fixed in one file for later verification.

• The plan formalizes the next work around alpha_first_glued_refined_hot_cap and splits it into:

• branch freeze

• alpha attribution

• execution-aware robustness

• formal governance

• final strategy memo

• Immediate execution focus remains Phase 2: explain exactly where refined alpha comes from.

• Completed Phase 2 refined attribution work.

• Added:

• dragon_refined_alpha_attribution.py

• dragon_refined_family_decomposition.csv

• dragon_refined_alpha_attribution.csv

• dragon_refined_removed_trade_recheck.csv

• dragon_refined_edge_review.md

• Main conclusions:

• refined alpha still comes mainly from glued_buy

• the branch wins by deleting weak short glued trades rather than adding complex paths

• removed trades vs control remain 11, all classified KEEP_REMOVAL

• removed-set avg_return -1.81%, win_rate 0%

• Completed execution-aware robustness validation.

• Added:

• dragon_refined_execution_validation.py

• dragon_refined_execution_stress.csv

• dragon_refined_latency_review.csv

• dragon_refined_risk_cluster_review.csv

• dragon_refined_stability_review.md

• Main conclusions:

• refined still leads under next-open execution: avg_return 3.31% vs 2.76%, PF 4.73 vs 3.78

• refined still leads under next-close execution: avg_return 2.44% vs 1.98%, PF 2.78 vs 2.37

• refined still leads under next-open + 20 bps/side: CAGR 25.51% vs 22.40%, PF 3.64 vs 2.92

• refined also improves risk clusters under next-open:

• max loss streak 8 vs 10

• worst 5-trade sum -7.96% vs -10.11%

• max drawdown -13.19% vs -17.84%

• Completed formal strategy packaging.

• Added:

• dragon_formal_strategy_governance.md

• dragon_parameter_governance.md

• dragon_strategy_monitoring_template.csv

• dragon_formal_strategy_memo.md

• This closes the staged roadmap from Phase 1 through Phase 5 in the current session.

• Workspace-level conclusion is now explicit:

• alpha_first_glued_refined_hot_cap is the best branch quantitatively

• if not promoted, that is now a governance decision rather than a missing-research issue