Startsida Utforska Hjälp
Logga in
openclaw
/
cyb50-quant
1
0
Fork 0
Filer Ärenden 0 Pull-förfrågningar 0 Wiki
Träd: fa4f4147ea
Grenar Taggar
cyb50-quant
/
research
/
dragon
/
v2
/
memory
/
2026-04-10.md

2026-04-10.md 11 KB
Historik Rå

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