cyb50-quant/research/dragon/v2/memory/2026-04-11.md

2026-04-11

• User asked to continue autonomously after the lean Phase 2 result.
• Follow-up work stayed narrow and only targeted the remaining promising lane:
• mid_zone_very_weak_b1

• goal: test whether the current problem is early exit treatment after followthrough reentry

• Implemented new OpenSpec change:

• openspec/changes/followthrough-mid-exit-shadow-probe/

• Added:

• dragon_glued_followthrough_exit.py

• dragon_followthrough_mid_exit_review.py

• tests/test_followthrough_mid_exit_probe.py

• openspec/changes/followthrough-mid-exit-shadow-probe/acceptance-summary.md

• Strategy change is shadow-only:

• new config gates in dragon_strategy_config.py are default-off

• new branch config alpha_first_glued_followthrough_mid_exit_probe_config() lives in dragon_branch_configs.py

• dragon_strategy.py now has a branch-gated hold for:

• entry code entry_glued_followthrough_reentry

• kdj_sell=True

• ql_sell=False

• early holding window

• mild positive repair state

• this prevents immediate knife_take_profit_2_glued only in the new shadow branch

• Added the new branch to dragon_daily_signal_pipeline.py so it is visible in daily and forward observation outputs.

• Historical result from the new branch:

• original mid probe:

• 2020-12-01 -> 2020-12-09

• sell reason knife_take_profit_2_glued

• return -0.24%

• new mid exit probe:

• 2020-12-01 -> 2021-02-19

• sell reason high_regime_confirmed_exit:kdj_sell

• holding 80 days

• return +26.65%

• Trade-path difference versus the old mid probe:

• removed:

• 2020-12-01 -> 2020-12-09 followthrough reentry loss

• 2020-12-15 -> 2021-02-19 later glued re-buy

• added:

• one merged trend trade 2020-12-01 -> 2021-02-19

• interpretation:

• the repaired trend was real

• the old branch was cutting it too early and then buying back later

• Review artifact conclusions from dragon_followthrough_mid_exit_review.md:

• same_close compounded return:

• base 1424.12%

• mid_probe 1420.41%

• mid_exit_probe 1460.02%

• next_open compounded return:

• base 1295.54%

• mid_probe 1288.52%

• mid_exit_probe 1332.91%

• this makes alpha_first_glued_followthrough_mid_exit_probe the only currently promising shadow extension worth monitoring

• but promotion is still not justified because the evidence is still one repaired-trend path

• Validation:

• py -3 -m py_compile dragon_glued_followthrough_exit.py dragon_followthrough_mid_exit_review.py tests/test_followthrough_mid_exit_probe.py dragon_daily_signal_pipeline.py

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

• py -3 dragon_followthrough_mid_exit_review.py

• py -3 -m unittest discover -s tests -v -> 26 tests passed

• py -3 dragon_daily_signal_pipeline.py passed

• py -3 dragon_forward_observation_pipeline.py passed

• Also fixed a weekend-sensitive test failure unrelated to the strategy logic:

• updated tests/test_data_fetcher_tday_cache.py to use an effective business day when testing snapshot-merge writeback

• this keeps the full suite green even when current date falls on a weekend