cyb50-quant/research/dragon/v2/memory/2026-04-07.md

2026-04-07

• Added dragon_excel_export.py to export v2 strategy results into an Excel workbook styled off the original 龙泉回测20260109.data.xlsx.
• Export output file: dragon_v2_compare_export.xlsx.
• Export workbook keeps the original sheets and adds three v2 comparison sheets:
• 龙泉_v2_workbook_preserving
• 龙泉_v2_alpha
• 龙泉_v2_RC1
• Data source uses freshly fetched 399673 daily history through the latest available market bar and recomputes indicators plus branch events locally.
• Export sheet layout intentionally stays close to the original workbook:
• original header rows retained
• KDJ / QL signal columns retained
• real / auxiliary strategy operations written into the same operation-direction column
• in-position floating PnL rows and yearly summary rows are also generated

• Current generated workbook run used latest available bar 2026-04-03.

• User then requested an overfitting-risk inspection on dragon v2.

• Current judgment:

• not a trivial black-box overfit, because walk-forward / local sensitivity / execution-latency stress all remain directionally supportive

• but still carries medium-to-high overfitting risk because the strategy tree is complex, the parameter surface is wide, and multiple weak families remain sample-dependent

• key hard risk buckets:

• deep_oversold_* family

• predictive_b1_break_* bridge logic

• broad historical iterative tuning path dependence around workbook alignment

• serious code-based overfitting review completed on 2026-04-07.

• verdict: not a fake strategy, but still medium-to-high overfitting risk due to high rule-tree complexity, 102-field config surface, and several low-sample specialized families.

• strongest stable core: glued_buy and its refined selective filter family.

• main fragile families: deepoversold* and predictive_b1break* bridge logic; also one-off reentry/exit rules should be treated as provisional rather than core alpha.

• Added dragon_overfit_trade_map.py and generated dragon_overfit_trade_map.html.

• This new HTML visualizes full historical trades against the price curve, supports branch switching, and explicitly highlights trades judged as medium/high overfitting risk.

• Current first-pass counts from the generated report:

• RC1: total trades 94, high-risk 10, medium-risk 12

• alpha_first_selective_veto: total trades 105, high-risk 10, medium-risk 14

• workbook_preserving: total trades 109, high-risk 8, medium-risk 20

• Built an external consultation bundle for dragon/v2:

• folder: dragon_v2_consult_2026-04-07/

• zip: dragon_v2_consult_2026-04-07.zip

• bundle contents:

• Chinese bottleneck/risk memo

• suggested ChatGPT Pro prompt

• core source files

• branch / RC1 parameter snapshots

• key evidence reports (rule_ablation, threshold_perturbation, walk_forward, deep_oversold, predictive_break)

• minimal supporting data and workbook exports

• Main consulting judgment written into the bundle:

• strongest stable core remains glued_buy

• biggest architecture bottleneck is the large order-sensitive rule tree in dragon_strategy.py

• biggest quant fragility still concentrates in deep_oversold_* and predictive_b1_break_*