cyb50-quant/research/dragon/v2/memory/2026-04-10.md

2026-04-10

• User asked: 统计下4.8号那天是什么信号过滤了开单.
• Verified from daily_reports/dragon_daily_signal_report_2026-04-08.md and branch status CSV:
• 2026-04-08 had KDJ buy=True, QL buy=True, a1=-0.0073, b1=-0.0777, c1=47.86.
• No buy event fired (events_today_count=0) on all three branches.
• The blocking rule is buy_block_glued_high_weak_rebound, exact subclause ql_rebound_weak_followthrough.
• Matching condition set:
• ql_buy=True
• 35 < c1 < 55 -> 47.86
• b1 < -0.06 -> -0.0777

• a1 > -0.013 -> -0.0073

• User asked historical return stats for buy_block_glued_high_weak_rebound.

• Based on dragon_glued_high_weak_rebound_blocked_trades.csv current snapshot:

• total blocked rows 15, completed added trades 14, one pending/unclosed row (2025-11-13).

• completed sample stats:

• win rate 28.57% (4/14)

• avg return -0.98%

• median return -0.94%

• best/worst +2.44% / -3.91%

• avg holding 6.93 days

• compounded return if all allowed -13.04%

• subtype stats:

• high_zone_weak_b1: 10 trades, win rate 20%, avg -1.38%

• mid_zone_very_weak_b1: 2 trades, win rate 100%, avg +1.67%

• ql_rebound_weak_followthrough: 2 trades, win rate 0%, avg -1.62%

• User asked to further split by market regime and list each trade detail.

• Added regime classification at buy date using close/MA20/MA60 from dragon_indicator_snapshot_full.csv:

• uptrend: close >= MA20 >= MA60

• downtrend: close <= MA20 <= MA60

• range: otherwise

• Generated files:

• dragon_glued_high_weak_rebound_regime_stats.csv

• dragon_glued_high_weak_rebound_year_regime_stats.csv

• dragon_glued_high_weak_rebound_trade_details_with_regime.csv

• Completed-trade regime stats (14 samples):

• downtrend 1 trade, win rate 0%, avg -2.20%

• uptrend 3 trades, win rate 0%, avg -0.94%

• range 10 trades, win rate 40%, avg -0.87%

• User asked to list trades with the same block reason as 2026-04-08 (ql_rebound_weak_followthrough under buy_block_glued_high_weak_rebound).

• Matching historical blocked trades:

• 2024-06-04 -> 2024-06-06, return -1.04%, hold 2d, sell reason early_failed_rebound_exit.

• 2024-06-17 -> 2024-06-20, return -2.20%, hold 3d, sell reason knife_take_profit_2_glued.

• Summary of this subtype (2 completed samples): win rate 0%, avg return -1.62%.

• Reconfirmed on user follow-up: historical blocked trades matching 2026-04-08 same reason remain exactly two rows in dragon_glued_high_weak_rebound_blocked_trades.csv (2024-06-04 and 2024-06-17).

• Clarified count confusion with user:

• large count refers to top-level rule buy_block_glued_high_weak_rebound (historically 15 in release window; 17 full-history on 2026-04-08 analysis; 18 including 2026-04-08 realtime snapshot),

• while 2026-04-08 exact matching subclause ql_rebound_weak_followthrough has only 2 historical rows (2024-06-04, 2024-06-17).

• User asked for conceptual difference between the two signals.

• Clarification:

• buy_block_glued_high_weak_rebound is the parent veto rule (OR of three clauses).

• ql_rebound_weak_followthrough is only one of the three clauses (requires ql_buy=True, 35<c1<55, b1<-0.06, a1>-0.013), so naturally has much fewer historical hits.

• User asked to list all child veto clauses under the parent rule.

• Confirmed from strategy code: exactly three OR clauses:

• high_zone_weak_b1

• mid_zone_very_weak_b1

• ql_rebound_weak_followthrough

• User requested all interception details for buy_block_glued_high_weak_rebound.

• Re-listed full historical blocked rows from dragon_glued_high_weak_rebound_blocked_trades.csv:

• total 15, completed 14, pending 1 (2025-11-13).

• subtype counts: high_zone_weak_b1=11, mid_zone_very_weak_b1=2, ql_rebound_weak_followthrough=2.

• User requested packaging the full strategy for GPT-Pro consultation.

• Created new consultation bundle:

• folder: dragon_v2_consult_2026-04-10_rc1/

• zip: dragon_v2_consult_2026-04-10_rc1.zip

• Bundle includes:

• updated core source (layered engine, runtime decoupling, governance, daily/forward pipelines)

• parameter snapshots and manifests

• key reports (ablation/sensitivity/stability/rollout/weak-rebound block review)

• recent daily reports (2026-04-08, 2026-04-09)

• blocked-trade detail datasets including regime breakdown

• OpenSpec change packs for recent restructuring/governance work

• Added bundle docs:

• README_CN.md

• consult_context_CN.md

• chatgpt_pro_prompt_CN.md

• User asked to study detailed answers under gpt-pro/.

• Reviewed package:

• gpt-pro/README_CN.md

• gpt-pro/CODEX_STRICT_EXECUTION_PROMPT_CN.md

• gpt-pro/PATCH_PLAN_FILE_BY_FILE_CN.md

• gpt-pro/dragon_glued_followthrough_codex_pack_2026-04-10/IMPLEMENTATION_SPEC_CN.md

• .../PHASED_DELIVERY_AND_ACCEPTANCE_CN.md

• .../RESEARCH_NOTES_AND_EXPECTED_BEHAVIOR_CN.md

• .../preview_results/preview_validation.json

• Core proposal recognized:

• Keep buy_block_glued_high_weak_rebound hard block in default RC1.

• Add shadow branch alpha_first_glued_followthrough_probe with pending/confirmation scaffold.

• Only allow subtype mid_zone_very_weak_b1 into pending in phase-1; keep ql_rebound_weak_followthrough hard-blocked.

• Important compatibility note:

• GPT-Pro preview baseline hashes/counts (events=296, trades=98, hashes 85ca.../7af4...) do not match this repo's current RC1 golden (events=272, trades=91, hashes 8965.../1298...).

• Their patch paths assume source/ layout, while current repo uses root-level python files.

• Conclusion: strategy direction is reasonable as shadow-research scaffold, but must be adapted to current repo baseline/tests before any landing.

• User approved: execute full GPT-Pro recommendation in OpenSpec mode.

• Implemented new OpenSpec change:

• openspec/changes/glued-followthrough-shadow-reentry/

• includes .openspec.yaml, proposal.md, design.md, tasks.md, two spec files, and acceptance-summary.md.

• Code implementation completed (current repo layout adaptation):

• new module dragon_glued_followthrough_confirmation.py

• added glued followthrough config flags in dragon_strategy_config.py (default conservative/off)

• added branch config alpha_first_glued_followthrough_probe_config() in dragon_branch_configs.py

• strategy wiring in dragon_strategy.py:

• added glued pending state fields

• queue/clear helpers

• pending lifecycle update + invalidation

• pending confirmation decision path

• block-path queue hook under subtype-gated branch config

• clear pending on real buy/sell

• added rule catalog mapping in dragon_rule_catalog.py for glued_followthrough_reentry_buy:*

• added shadow branch to daily pipeline in dragon_daily_signal_pipeline.py

• Added test:

• tests/test_glued_followthrough_shadow_reentry.py

• Validation results:

• py -3 -m py_compile ... pass

• py -3 -m unittest discover -s tests -v pass (22 tests)

• py -3 dragon_daily_signal_pipeline.py pass

• py -3 dragon_forward_observation_pipeline.py pass

• Compatibility outcomes:

• release-window RC1 hashes unchanged (events 8965d1...a331, trades 1298be...cb97)

• probe branch release-window adds exactly 1 extra trade:

• 2020-12-01 -> 2020-12-09

• glued_followthrough_reentry_buy:confirmed_mid_zone_very_weak_b1

• return -0.0024350630226196435

• User asked what GPT-Pro suggestions remain not implemented.

• Cross-check result:

• Phase 1 mandatory scope is implemented.

• Remaining not-done items are GPT-Pro Phase 2 research suggestions:

• add block-after-followthrough observation metrics

• run independent shadow toggle experiment for high_zone_weak_b1

• run execution timing study (same-close vs next-open) for followthrough reentry path

• evaluate entry-specific exit treatment / exemption for followthrough reentry

• User clarified the core target explicitly:

• the strategy is meant to make money by following trends.

• the biggest real problem is how to rejoin the trend quickly after a false veto.

• future optimization ordering should focus on:

• false-veto detection

• timely followthrough reentry

• trend profit capture

• User further clarified execution style preference:

• do not expand this into a large research project.

• keep future work lean, regime-focused, and practical for a small team / ~1M capital setup.

• User pointed out that GPT-Pro had already stated this direction explicitly.

• Reconfirmed from GPT-Pro docs:

• IMPLEMENTATION_SPEC_CN.md states block subtype should enter pending and attempt delayed reentry within 1~3 bars after stronger confirmation.

• RESEARCH_NOTES_AND_EXPECTED_BEHAVIOR_CN.md prioritizes execution timing first and entry-specific exit treatment second.

• After re-reading GPT-Pro docs, one implementation detail was found missing and then fixed:

• added full-snapshot local hash pinning to tests/test_glued_followthrough_shadow_reentry.py

• full baseline pinned: events 296, trades 98, hashes 5636adc7... / 1d419aaf...

• User then made the priority explicit again:

• set false-veto recovery as the first priority and continue developing autonomously without mid-task interruptions.

• Implemented a lean Phase 2 OpenSpec change:

• openspec/changes/followthrough-lean-profit-loop/

• Added:

• dragon_followthrough_profit_loop_review.py

• tests/test_followthrough_profit_loop_review.py

• openspec/changes/followthrough-lean-profit-loop/acceptance-summary.md

• Ran:

• py -3 -m py_compile dragon_followthrough_profit_loop_review.py tests/test_followthrough_profit_loop_review.py

• py -3 -m unittest tests.test_followthrough_profit_loop_review -v

• py -3 dragon_followthrough_profit_loop_review.py

• py -3 -m unittest discover -s tests -v

• Final validation: full suite 25 tests passed.

• Lean Phase 2 findings:

• mid_zone_very_weak_b1:

• blocked_full 2

• shadow_confirmed 1

• confirm_like_3bar 50%

• avg_next20_max +7.04%

• probe trade 2020-11-30 -> 2020-12-01 -> 2020-12-09

• same_close -0.24%

• next_open -0.50%

• post_exit_max_10b +8.35%

• interpretation: the only subtype still worth a next-step experiment, and the bottleneck now looks more like execution/exit handling than whether any followthrough exists at all

• high_zone_weak_b1:

• blocked_full 13

• shadow_confirmed 2

• same_close avg -3.52%

• next_open avg -5.03%

• interpretation: do not widen this family; followthrough confirmation here still loses

• ql_rebound_weak_followthrough:

• blocked_full 3 (includes current full-history latest case)

• shadow_confirmed 0

• completed historical windows: confirm_like_3bar 0%, next3_sell_cross_rate 100%

• interpretation: hard block remains justified

• Latest live-relevant blocked case review:

• 2026-04-08 subtype is still ql_rebound_weak_followthrough

• by 2026-04-10 only 2/3 followthrough bars have elapsed

• ql_reconfirm_count 0

• sell_cross_count 0

• max_close_return_so_far +3.56%

• still no confirm-like delayed reentry signal under current mid-zone/q l confirmation logic