cyb50-quant/research/dragon/v2/dragon_deep_oversold_confirmation_review.md

Dragon Deep Oversold Confirmation Review

Scope

• Research branch only.
• Default workbook-preserving baseline remains unchanged.
• Tested idea: replace immediate entry for weak deep_oversold subtypes with short-window QL confirmation.

Experiment Result

• Baseline: 107 trades, avg_return 2.52%, profit_factor 3.31, real BUY / SELL overlap 106 / 105
• confirm_weak_with_ql_1bar: 104 trades, avg_return 2.67%, profit_factor 3.56, real BUY / SELL overlap 101 / 102
• confirm_weak_with_ql_2bar: same top-level result as 1bar
• confirm_weak_with_ql_3bar: 103 trades, avg_return 2.73%, profit_factor 3.66, real BUY / SELL overlap 101 / 101

What Actually Changed

• All confirmation branches removed these baseline weak-subtype trades:
• 2022-03-11 -> 2022-03-15 positive_b1_rebound
• 2024-01-18 -> 2024-01-22 shallow_false_start

• 2024-01-23 -> 2024-01-29 shallow_false_start

• 1bar / 2bar also shifted two entries later rather than deleting them:

• 2022-03-16 -> 2022-03-25 became 2022-03-17 -> 2022-03-25

• 2024-08-29 -> 2024-09-09 became 2024-08-30 -> 2024-09-09

• 3bar was numerically best, but for an important reason:

• it no longer re-entered the 2022-03-16 positive-B1 chain at all

• it only kept the 2024-08-30 confirmed shallow-false-start chain

Trade-Level Interpretation

• 2022-03-11 had no QL confirmation before failure. This is a clean veto candidate.
• 2022-03-16 did get next-day QL, but delayed confirmation still produced a losing trade (2022-03-17 -> 2022-03-25, about -3.18%).
• 2024-01-18 and 2024-01-23 had no QL confirmation and both failed quickly. These are also clean veto candidates.
• 2024-08-29 did get next-day QL, but the delayed confirmed trade (2024-08-30 -> 2024-09-09) still lost about -2.66%.

Quant Judgment

• The confirmation branch improves top-level returns mostly by vetoing weak subtype trades, not by creating better confirmed entries.
• In other words, QL confirmation is acting more like a rejection filter than a genuine alpha confirmation for the weak deep_oversold bucket.
• The best-looking 3bar result is better mainly because it blocks more weak trades, not because a longer confirmation window is intrinsically superior.
• This reinforces the earlier conclusion from the subtype blocking experiments: the alpha-first direction for weak deep_oversold subtypes is closer to selective veto / removal than to delayed confirmation.

Next Decision

• For the workbook-preserving baseline: do nothing. These changes cost too much date alignment.
• For a future alpha-first branch: prioritize weak-subtype veto logic before building a more elaborate confirmation state machine.
• If confirmation is revisited later, it should be paired with an additional quality filter rather than using QL alone.