Rank113 alpha-beta abstain clean replication

结论：park / evidence pool

这轮最小 clean replication 没把 abstain gate 变成更诚实的 fib-retest overlay：成本后收益、假跟随率与留存没有同时改善，因此当前更适合 park。

这轮故意只挂 1 条 archetype：fib_retest_long。先在训练段冻结 hold_bars=12、lower_q=0.15、upper_q=0.80，再在测试段统一比较 baseline / lower_band_only / dual_band 三臂，口径固定为 signal 当根及之前数据 + next-bar open + no-overlap。

主读法（6bps/side）

baseline：mean_total_return = 1.79%；retention = 78.79%；false_follow_through_4bars = 25.00%
lower_band_only：mean_total_return = 0.80%；retention = 66.67%；false_follow_through_4bars = 0.00%
dual_band：mean_total_return = 0.50%；retention = 33.33%；false_follow_through_4bars = 0.00%

这轮到底测了什么

noise-band：若 fib 回踩确认后，(close - fib_618) / ATR 仍太小，就当它还没真正走出来，不做。
shock-band：若同一个 proxy 已经太大，就当它更像追尾，不追。
profit-window：不按 accuracy 选窗口，而是先在训练段冻结一个更像能赚钱的持有窗口，再拿到测试段看 gate 是否还成立。

Overall summary

variant	cost_bps_per_side	mean_total_return	positive_asset_ratio	mean_trades	mean_trade_retention	mean_avg_net_return	mean_win_rate	mean_false_follow_through_4bars	mean_proxy_distance_atr
baseline	6	1.79%	100.00%	6.0	78.79%	0.35%	80.00%	25.00%	2.17
dual_band	6	0.50%	100.00%	1.0	33.33%	0.50%	100.00%	0.00%	1.68
lower_band_only	6	0.80%	100.00%	2.0	66.67%	0.40%	100.00%	0.00%	3.14
baseline	10	1.30%	100.00%	6.0	78.79%	0.27%	75.00%	25.00%	2.17
dual_band	10	0.42%	100.00%	1.0	33.33%	0.42%	100.00%	0.00%	1.68
lower_band_only	10	0.64%	100.00%	2.0	66.67%	0.32%	100.00%	0.00%	3.14

Asset summary（6bps）

variant	asset	cost_bps_per_side	baseline_signals	trades	trade_retention	total_return	avg_net_return	win_rate	false_follow_through_4bars	avg_proxy_distance_atr
baseline	ETH-USD	6	3	2	66.67%	0.80%	0.40%	100.00%	0.00%	3.14
baseline	SOL-USD	6	11	10	90.91%	2.78%	0.30%	60.00%	50.00%	1.21
dual_band	ETH-USD	6	3	1	33.33%	0.50%	0.50%	100.00%	0.00%	1.68
lower_band_only	ETH-USD	6	3	2	66.67%	0.80%	0.40%	100.00%	0.00%	3.14

Chosen thresholds

asset	train_rows	lower_q	upper_q	lower_thr	upper_thr	proxy_train_median
BTC-USD	11	0.15	0.80	0.85	1.74	1.41
ETH-USD	9	0.15	0.80	0.53	1.79	0.99

Train band / horizon grid

hold_bars	lower_q	upper_q	train_mean_total_return	train_positive_asset_ratio	train_mean_retention
12	0.15	0.80	5.17%	100.00%	59.60%
12	0.15	0.85	5.17%	100.00%	59.60%
12	0.20	0.80	5.17%	100.00%	59.60%
12	0.20	0.85	5.17%	100.00%	59.60%
12	0.25	0.80	4.83%	100.00%	55.05%
12	0.25	0.85	4.83%	100.00%	55.05%
12	0.15	0.75	4.55%	100.00%	55.05%
12	0.20	0.75	4.55%	100.00%	55.05%
12	0.25	0.75	4.22%	50.00%	50.51%

Veto reason summary（6bps）

variant	asset	veto_reason	count
dual_band	ETH-USD	above_shock_band	1
dual_band	SOL-USD	missing_proxy	11
lower_band_only	SOL-USD	missing_proxy	11

Time bucket summary（6bps）

variant	bucket	trades	mean_net_return	total_return	false_follow_through_4bars
baseline	bucket_1	4	-0.65%	-2.61%	75.00%
baseline	bucket_2	4	0.52%	1.95%	50.00%
baseline	bucket_3	4	1.07%	4.34%	0.00%
dual_band	bucket_3	1	0.50%	0.50%	0.00%
lower_band_only	bucket_1	1	0.30%	0.30%	0.00%
lower_band_only	bucket_3	1	0.50%	0.50%	0.00%

Sample meta

asset	symbol	bars	signals	sample_start_utc	sample_end_utc
BTC-USD	BTCUSDT	11531	11	2025-11-16T09:30:00Z	2026-03-16T12:00:00Z
ETH-USD	ETHUSDT	11531	12	2025-11-16T09:30:00Z	2026-03-16T12:00:00Z
SOL-USD	SOLUSDT	11531	11	2025-11-16T09:30:00Z	2026-03-16T12:00:00Z

Artifacts: overall_summary.csv / asset_summary.csv / chosen_thresholds.csv / train_band_horizon_grid.csv / veto_reason_summary.csv / trade_log.csv / summary.json

Rank 113 / alpha-beta abstain / profit-window — minimal clean replication