cyb50-quant/research/dragon/v2/memory/2026-04-08.md

2026-04-08

• Ran dragon_daily_signal_pipeline.py on request date 2026-04-08.
• Latest available market bar resolved to 2026-04-07.
• Latest 399673 snapshot on 2026-04-07:
• close 3325.910
• a1 = -0.0182
• b1 = -0.1651
• c1 = 44.35
• KDJ buy/sell = False/False
• QL buy/sell = False/False
• No branch fired any new event on the latest bar.
• workbook_preserving, alpha_first_selective_veto, and alpha_first_glued_refined_hot_cap all remain flat.
• Latest real event for all three branches is still 2026-02-13 SELL knife_take_profit_2_glued.

• User-facing answer for this run: no buy trigger on the latest available bar.

• Also ran dragon_forward_observation_pipeline.py to persist the daily check.

• Forward-observation result on 2026-04-07:

• signal-change count 0

• divergence level none

• warnings 0

• hard breaches 0

• missing data metrics 0

• Later on 2026-04-08, upgraded the live-signal path to support intraday realtime triggering.

• Implementation:

• dragon/data_fetcher_v2.py now supports realtime snapshot fetch and conditional current-day snapshot append.

• dragon_indicators.py now supports fetch_daily_data(include_intraday_snapshot=True) and exposes fetch metadata.

• dragon_daily_signal_pipeline.py now uses intraday snapshot mode by default for same-day checks and reports:

• data_mode

• historical_latest_bar_date

• snapshot_timestamp

• explicit note that current market price is used as today's provisional close

• Verified live run after the upgrade:

• latest evaluated bar became 2026-04-08 via realtime snapshot

• historical latest official bar remained 2026-04-07

• snapshot timestamp example: 2026-04-08T14:47:31

• live price snapshot close around 3524.990

• latest markers on all three branches:

• KDJ buy=True

• QL buy=True

• but no real BUY event fired

• local decision check on the last bar shows all three branches were blocked by:

• buy_block_glued_high_weak_rebound

• Verified dragon_forward_observation_pipeline.py still runs successfully after the schema change.

• Analyzed buy_block_glued_high_weak_rebound on 2026-04-08.

• Current rule location:

• dragon_strategy.py _buy_filter_glued_high_weak_rebound(...)

• official daily-bar replay through 2026-04-07 shows the same historical block count on all three tracked branches:

• full history block count 17

• release-window (2016-01-01 -> 2025-12-31) block count 15

• subtype split in the release window:

• high_zone_weak_b1: 11

• mid_zone_very_weak_b1: 2

• ql_rebound_weak_followthrough: 2

• release-window block dates:

• 2016-04-29, 2017-09-27, 2018-04-18, 2019-04-24, 2019-10-10, 2020-08-18, 2020-08-24, 2020-11-30, 2021-07-29, 2021-12-09, 2022-08-01, 2024-06-04, 2024-06-17, 2025-10-21, 2025-11-13

• including the intraday realtime snapshot on 2026-04-08, RC1 block count becomes 18 total; the current live block belongs to the ql_rebound_weak_followthrough clause.

• Effect check by disabling only this filter while keeping the rest of the strategy unchanged:

• workbook_preserving: trades 106 -> 121, avg_return 2.57% -> 2.13%, profit_factor 3.37 -> 2.90

• alpha_first_selective_veto: trades 102 -> 117, avg_return 2.86% -> 2.36%, profit_factor 4.04 -> 3.38

• alpha_first_glued_refined_hot_cap / RC1: trades 91 -> 106, avg_return 3.42% -> 2.80%, profit_factor 5.11 -> 4.08

• In all three branches the filter removal adds 15 trades and removes 0.

• Added-trade quality is weak:

• avg_return about -0.99%

• win_rate about 26.7%

• Therefore current judgment: this filter is materially useful and should be treated as a real quality filter, not noise.

• Added detailed blocked-trade artifacts:

• dragon_glued_high_weak_rebound_block_review.md

• dragon_glued_high_weak_rebound_blocked_trades.csv

• RC1 blocked-trade detail sorted by realized return shows the worst candidates were:

• 2020-08-18 -> 2020-08-19 -3.91%

• 2019-04-24 -> 2019-04-25 -3.42%

• 2018-04-18 -> 2018-04-20 -3.10%

• best completed blocked candidates were:

• 2020-11-30 -> 2020-12-09 +2.44%

• 2017-09-27 -> 2017-10-16 +1.60%

• 2020-08-24 -> 2020-08-25 +0.90%

• one extra blocked date 2025-11-13 is present as a candidate date but its corresponding exit falls outside the release-window trade summary, so it has no completed in-window trade row.

• Checked the original reconstruction branch (workbook_preserving) on the intraday snapshot later on 2026-04-08.

• Snapshot example:

• evaluated timestamp 2026-04-08T15:56:41

• close 3532.798

• a1 = -0.0073

• b1 = -0.0777

• c1 = 47.86

• KDJ buy = True

• QL buy = True

• original branch decision was still:

• BLOCK buy_block_glued_high_weak_rebound

• User later requested the 2026 RC1 buy/sell points.

• Extracted real-trade RC1 points for calendar year 2026:

• 2026-01-05 BUY dual_gold_resonance_buy @ 3488.230

• 2026-01-20 SELL ql_mid_zone_take_profit @ 3419.621

• 2026-01-27 BUY glued_buy @ 3476.169

• 2026-02-02 SELL knife_take_profit_2_glued @ 3432.873

• 2026-02-09 BUY glued_buy @ 3493.591

• 2026-02-13 SELL knife_take_profit_2_glued @ 3423.045

• As of the intraday snapshot on 2026-04-08, RC1 remains flat and no new real event has been added after 2026-02-13.

• Later analysis of why no intraday buy fired on 2026-04-08:

• the live snapshot did create raw buy markers (KDJ buy=True, QL buy=True) and satisfied glued-state entry context (abs(a1) < 0.02)

• but the final buy decision was blocked by buy_block_glued_high_weak_rebound

• exact subtype was ql_rebound_weak_followthrough

• snapshot values around 2026-04-08T14:48:51:

• a1 = -0.00779

• b1 = -0.08157

• c1 = 47.62

• this matched the veto window:

• 35 < c1 < 55

• b1 < -0.06

• a1 > -0.013

• ql_buy = True

• the move was therefore interpreted as a fast intraday rebound with still-weak followthrough rather than a confirmed real buy

• Follow-up judgment on whether this may miss a real rally or reflect overfitting:

• yes, any veto rule can miss an individual future rally

• but current evidence does not support calling this specific filter an obvious overfit blocker

• disabling buy_block_glued_high_weak_rebound adds 15 release-window trades and weakens all three tracked branches

• RC1 added-trade set stats if allowed:

• win rate about 26.7%

• avg return about -0.99%

• best completed blocked trade only about +2.44%

• today's exact subtype ql_rebound_weak_followthrough has only two historical daily examples and both were losers (-2.20%, -1.04%)

• nuanced conclusion:

• strategy-level overfitting risk remains medium-to-high overall because of rule-tree complexity

• but today's specific veto is more likely a useful quality filter than a clear case of overfit signal suppression

• remaining uncertainty is concentrated in the sparse ql_rebound_weak_followthrough subclause because its sample is still small

• Direct advisory judgment for the user:

• if forced to decide on the live 2026-04-08 snapshot, the correct call is still do not buy yet

• do not weaken the global buy_block_glued_high_weak_rebound rule because of this one signal

• safer future research path, if needed, is a subtype-specific second-chance release after stronger followthrough repair, not removal of the veto itself

• Later nuance after the user suggested this may be the start of a big trend:

• that concern is legitimate; broad neighborhood analysis around today's indicator shape is not obviously bearish

• among the 12 nearest historical RC1 buy-shape neighbors to the 2026-04-08 snapshot, avg return was about +2.09%, median about +1.67%, win rate about 66.7%, with 3 trades above +5% and 1 above +10%

• however, the exact vetoed subclause neighborhood (ql_rebound_weak_followthrough) still has only 2 historical blocked daily examples and both were losers

• operational interpretation:

• there is a non-trivial chance today is an early stage of a larger move

• but current evidence still does not justify overriding RC1 on this bar alone

• if followthrough continues and either B1 repairs above the current veto threshold or the setup migrates out of this weak-followthrough window, the bull case should be re-evaluated quickly rather than dismissed

• Later projected tomorrow's B1 using the latest available intraday snapshot around 2026-04-08T18:25:09 with close = 3532.798.

• Using the current Y2/Y3 recursion and assuming no new 38-day extreme beyond the projected close:

• if tomorrow closes flat near 3532.798, projected B1 is about -0.0229

• if tomorrow closes -1%, projected B1 is about -0.0421

• if tomorrow closes -2%, projected B1 is about -0.0614

• approximate threshold for B1 >= -0.06 is tomorrow close around 3464.7, about -1.93% vs the latest snapshot close

• practical takeaway:

• yes, under the current snapshot baseline, “tomorrow not falling” should place B1 above -0.06

• but that only means the current ql_rebound_weak_followthrough veto likely stops applying

• it does not guarantee a real buy tomorrow because KDJ buy / QL buy are cross-event flags and may not re-fire on the next bar

• Later reviewed the external recommendation file dragon_v2_rc1_restructuring_recommendations_for_codex_CN.md and wrote a local implementation proposal dragon_v2_rc1_restructuring_proposal_cn.md.

• Main judgment on that external recommendation:

• directionally correct on freezing RC1, splitting core / secondary / bridge, downgrading predictive logic, and adding golden tests

• but too aggressive as a direct big-bang refactor for the current repo

• key verified repo facts used in the proposal:

• dragon_strategy.py has 1297 lines

• StrategyContext has 50 fields

• _buy_decision() about 125 lines

• _sell_decision() about 303 lines

• StrategyConfig has 102 fields

• current repo still has no tests/ directory

• reason strings are already a de facto interface across strategy, daily pipeline, path trace, and html reporting

• proposal conclusion:

• do not directly rewrite RC1 or remove reason strings

• first add structured decision metadata with legacy reason compatibility

• first add golden baselines and layer attribution

• only then isolate weak families (deep_oversold, predictive, post_sell_rebound) into separate modules on a research path

• Executed the first implementation pass for proposal Phase 0-2 without changing RC1 live behavior.

• Added governance and scaffolding modules:

• dragon_rc1_frozen_policy.md

• dragon_reason_types.py

• dragon_decision_types.py

• dragon_rule_catalog.py

• Added executable scripts:

• dragon_rc1_golden_baseline.py

• dragon_rule_layer_attribution.py

• Generated new baseline artifact:

• dragon_rc1_golden_manifest.json

• generated_at 2026-04-08T23:43:07

• evaluation window 2016-01-01 -> 2025-12-31

• RC1 trade_count 91, event_count 272, win_rate 52.75%, avg_return 3.42%

• indicator source used: dragon_indicator_snapshot_full.csv

• Generated layer-attribution report:

• dragon_rule_layer_attribution.md

• key entry-layer read:

• core/glued remains dominant (50 trades, avg_return 4.92%, PF 7.15)

• secondary/deep_oversold remains weak (avg_return -0.19%, PF 0.80)

• bridge/predictive_break count still low (1 entry trade in-window)

• key exit-layer read:

• secondary/post_sell_rebound exits show weak quality (avg_return -2.00%, PF 0.02)

• core/high_regime exits remain strong (avg_return 11.63%, PF 107.38)

• mapping audit is now clean in the generated markdown (no unknown reason mapping)

• Important repo note:

• generated .csv artifacts from these scripts are covered by .gitignore (*.csv) and are not tracked by default

• Later answered the user's question about what profitable indicator profiles look like in RC1.

• Main conclusion:

• the best-performing trades are not primarily deep-oversold knife catches

• profitable trades are concentrated in glued_buy, plus smaller contributions from early_crash_probe_buy / oversold_recovery_buy

• holding duration matters strongly:

• 00-05d avg return about -2.03%

• 06-10d avg return about -0.81%

• 21-40d avg return about +6.33%

• 41d+ avg return about +24.13%

• top-profit trades (>=10%) mostly look like:

• A1 near zero / mild glue rather than deeply negative

• B1 around zero to mildly positive more often than the full sample

• C1 mainly in the 20-60 middle zone rather than extreme <20 oversold

• KDJ buy is almost always present

• QL buy is helpful but not mandatory

• strongest actionable read:

• large profits mostly come from entries that can transition into long trend holds and exit through regime/trend management (prewarning_reduction_exit, crash_protection_exit, high-regime exits), not from short rebound scalps