okx-codex-trader/reports/exploration-round-summary.md

Exploration Round Summary

Scope: index of this round's new research reports under reports/eth-exploration and reports/long-short-fusion. Research only; no live path or order path changed.

Summary

direction	status	conclusion
staged entry	observe	Risk-shaping benefit exists, but not a general upgrade over next-open entry.
relative momentum	reject	Standalone and explicit regime-gated variants failed the full usable filter.
crash_follow	observe	Small overlay improves return, but no stable fusion improvement across return, Calmar, and drawdown.
calendar carry	include in portfolio search	Low-weight anomaly leg is justified; validated upper bound is 0.125.
bearish failure confirmation	reject	Recent windows are strong, but narrow validation showed no usable overlay or filter.
false breakout reversal	reject	Drawdown is low, but trade coverage is too sparse for a reliable strategy.
wick rejection	reject	Low correlation, but returns are negative across required windows.

Staged Entry

Conclusion: do not replace the current next-open entry. Staged entry lowers drawdown and raises profit factor, but the main zero-offset candidate gives up compounded return; pure offset variants skip too many trades.

Status: observe.

Key metrics:

• Baseline market_next_open: total return 338.84%, annualized 26.49%, max drawdown 10.89%, win rate 71.84%, PF 2.35, trades 277.
• Best period-breadth staged variant staged_o0.0000-0.0015-0.0030_v4: total return 290.70%, annualized 24.17%, max drawdown 10.16%, win rate 79.42%, PF 2.69, trades 277.
• Regime validation: return improved in 3/11 eligible slices; median return delta -1.92%, median drawdown delta -1.17%, median PF delta +0.67.
• Bull 90d slice improved return by +11.85%, but bear 90d slice underperformed by -29.92%.

Next step: keep staged entry as a separate read-only/live-intent variant only. Do not route default execution through it.

Relative Momentum

Conclusion: reject as a standalone direction. No candidate passed positive full/3y/1y/6m/3m filtering. The later 4H-lb84 regime-gate validation also failed to produce a logically closed usable strategy.

Status: reject.

Key metrics:

• Initial selected candidates had positive full returns, but recent horizons broke: top row full return 62.23%, max drawdown 18.58%, PF 1.49, but 3y -3.93% and 6m -0.91%.
• Some lb84 4H variants had better recent behavior, but full sample was weak or negative before gating.
• Best gated row eth_bear90_low_vol: full return 64.47%, max drawdown 28.01%, PF 2.93, trades 35, but min year -7.83%, min market segment -3.04%, min vol segment -12.54%.
• No regime-gated candidate passed the stated usable filter.

Next step: stop this branch unless a new premise is introduced. Do not add it to combination search.

Crash Follow

Conclusion: the crash-follow leg has useful convexity, but only as a constrained research overlay. It improves raw return at small weights, yet representative fusion overlay testing found no combination that improved return, Calmar, and drawdown across all required horizons.

Status: observe.

Key metrics:

• Conservative baseline plus crash_follow at 5%: full return 64.41%, max drawdown 7.55%, Calmar 1.08; delta return +6.94%, delta drawdown +0.27%.
• At 10%: full return 72.52%, max drawdown 9.27%, Calmar 0.97; return improves but drawdown expands.
• Crash-follow component alone: full return 101.14%, max drawdown 44.90%; 3y return 43.00%, 1y return 40.03%, 6m return 18.89%, 3m return 4.46%.
• Fusion overlay search: no stable improvement candidates; even 2% overlay improved returns but failed Calmar/drawdown stability on full horizon.

Next step: do not enter the next fusion main search now. If revisited, cap candidate weights at tiny overlay levels and require full-horizon Calmar not to deteriorate.

Calendar Carry

Conclusion: include in portfolio search as a small low-weight calendar anomaly leg. It is all-window-positive and improves representative fusion/conservative portfolio drawdown/Calmar at low overlay weights, but parameter neighborhood and 2024 stability are not clean enough for standalone promotion.

Status: include in portfolio search.

Key metrics:

• Selected rule eth-1h-long-h14-weekend-hold8-volcalm: full return 63.89%, annualized 8.18%, max drawdown 12.25%, win rate 52.75%, PF 1.36, trades 345.
• Required horizons: 3y +7.23%, 1y +3.80%, 6m +5.04%, 3m +0.36%.
• Year stability: 2024 was negative at -4.77%; positive month rate only 33.33% in 2024.
• Overlay check: 2.5%-10% weights improved full-period Calmar versus baseline, with 10% full return 53.77%, max drawdown 6.28%, Calmar 1.13.
• Fusion-leg validation: weights 0.025-0.125 improved full return, full annualized return, full max drawdown, and full Calmar versus the representative fusion baseline.
• At weight 0.125: full return 449.68% vs 414.33%, annualized 31.16% vs 29.78%, max drawdown 9.42% vs 9.72%, Calmar 3.31 vs 3.06.
• Degradation check: only 2024 was worse than baseline by year; 31/77 months were worse, with worst monthly delta -1.57%.
• Main fusion search with calendar leg: selected candidates concentrate at calendar weight 0.100-0.125. Best selected row fusion-cal-lr_riskoff00-l1.20-brs0.08-eth4hs0.00-btc4hs0.04-eg0.12-bg0.00-cal0.125 has full return 447.74%, annualized 31.09%, max drawdown 8.61%, Calmar 3.61.
• Main fusion caveat: top rows improve full/6m/3m versus no-calendar baselines, but 3y/1y generally fail the stricter "Calmar improves and DD does not increase" test.
• Observation shortlist: selected three read-only candidates covering low drawdown, high return, and lower calendar exposure. None should trade before live signal observation.
• Read-only intent generated for A/B/C. Current calendar leg has no entry signal and no theoretical active position as of the latest local candle in the report.
• Neighborhood is mixed: hold 8 is the only useful hold bucket; many adjacent hours/weekday/vol gate variants are weak.

Next step: deploy only the read-only intent builder if observation automation is needed. Do not promote standalone.

Bearish Failure Confirmation

Conclusion: reject for combination search after narrow validation. Trend-exhaustion variants looked promising in recent windows, but the left tail cannot be reduced without destroying the edge.

Status: reject.

Key metrics:

• Selected trend_exhaustion-1H-f50-s240-lb8-th0.012-sl0.03-tp0.045-h72-none: full return 75.77%, annualized 9.30%, max drawdown 36.15%, win rate 46.18%, PF 1.18, trades 275.
• Recent windows are strong: 3y +25.72%, 1y +29.58%, 6m +30.91%, 3m +5.19%.
• Stability issue: 2 losing years out of 7; 2022 -22.79%, 2023 -11.67%.
• Top-row alternative has lower drawdown, th0.032: full return 84.15%, max drawdown 24.77%, PF 1.50, but only 92 trades and 3m return 0.00%.
• Narrow validation: parameter neighborhood pass count was 0/17.
• Best structural filter reduced 2022 loss but full return fell to 3.93% and 3y return became -1.15%.
• Conservative portfolio overlay at 2.5%-10% lowered full Calmar and full return; 3y drawdown worsened.

Next step: stop this branch unless a new premise is introduced.

False Breakout Reversal

Conclusion: reject as an independent direction. The idea is coherent, but the selected candidate trades too rarely for statistical confidence.

Status: reject.

Key metrics:

• Selected false_breakout_reversal-btc-4H-bidir-rl36-cw240-cq0.25-sw0.002-hs80-hl6-sl0.025-tp0.035-h36: full return 24.86%, annualized 3.62%, max drawdown 6.81%, win rate 66.67%, PF 1.95, trades 30.
• Recent windows are superficially clean: 3y +19.70%, 1y +9.64%, 6m +9.64%, 3m +3.55%.
• Sample problem: only 30 full-sample trades, 3 trades in the last year, 1 trade in the last 3 months, and only 22 active months out of 76.
• Short-only variants are also sparse, generally around 20-24 trades for the best rows.

Next step: stop this branch. Do not combine or live-test until a non-sparse formulation exists.

Wick Rejection

Conclusion: reject. The single-candle upper-wick rejection signal has low monthly correlation to the current reference portfolio, but the edge is negative.

Status: reject.

Key metrics:

• Reference candidate: full return -6.40%, trades 157, PF 0.95, win rate 43.31%.
• Required windows are all negative: 3y -2.15%, 1y -3.43%, 6m -2.88%, 3m -3.83%.
• Monthly correlation to equal-2-c0003 was low at 0.0424, but low correlation without positive expectancy is not useful.

Next step: stop this branch.

Source Reports

• reports/eth-exploration/eth-nextgen-staged-entry-evaluation.md
• reports/eth-exploration/eth-nextgen-staged-entry-regime-validation.md
• reports/eth-exploration/eth-relative-momentum-report.md
• reports/eth-exploration/eth-relmom-lb84-regime-gate-report.md
• reports/eth-exploration/eth-btc-calendar-carry-report.md
• reports/eth-exploration/eth-btc-calendar-carry-candidate-stress.md
• reports/long-short-fusion/calendar-carry-fusion-leg.md
• reports/long-short-fusion/fusion-calendar-report.md
• reports/long-short-fusion/calendar-fusion-observation-candidates.md
• reports/long-short-fusion/calendar-fusion-observation-intent.md
• reports/eth-exploration/eth-bearish-failure-confirmation-report.md
• reports/eth-exploration/trend-exhaustion-narrow-validation.md
• reports/eth-exploration/eth-btc-false-breakout-reversal-report.md
• reports/eth-exploration/eth-btc-wick-rejection-report.md
• reports/long-short-fusion/eth-crash-follow-overlay.md
• reports/long-short-fusion/eth-crash-follow-fusion-overlay-search.md