Cross-generator overlap top 5

2026-05-20 nalaz

Prosečan signal bucket vidi 2.45 od 5 generatora. Efektivno samo 41% nezavisnih trejdova — hibrid od svih pet bio bi uglavnom redundansa.

↑ Svi izveštaji ← Korak B+D — slot-count osetljivost Korak A — istraga deljenog $79.82 max-win-a →

Bucket = entry candle rounded down to 4h. Dva signala su 'isti trade' ako su isti simbol + isti 4h bucket.

Raw counts

Generator	Raw buys	Unique buckets
best_v3	855	855
best_v2	1225	1225
blacklist	1993	1993
best	1492	1492
quality_cycle	2850	2850

Jaccard similarity matrix

Jaccard = |A∩B| / |A∪B|. 0 = disjunktni signali, 1 = identični.

	best_v3	best_v2	blacklist	best	quality_cycle
best_v3	1.000	0.582	0.311	0.450	0.183
best_v2	0.582	1.000	0.496	0.748	0.279
blacklist	0.311	0.496	1.000	0.656	0.520
best	0.450	0.748	0.656	1.000	0.357
quality_cycle	0.183	0.279	0.520	0.357	1.000

Directional overlap (row in col)

Vrednost = |A∩B| / |A|. Procenat signala iz reda A koji se javljaju i u koloni B.

A \ B	best_v3	best_v2	blacklist	best	quality_cycle
best_v3	100.0%	89.5%	79.1%	85.1%	66.9%
best_v2	62.4%	100.0%	87.1%	94.9%	72.5%
blacklist	33.9%	53.5%	100.0%	69.2%	83.1%
best	48.8%	77.9%	92.5%	100.0%	76.6%
quality_cycle	20.1%	31.2%	58.1%	40.1%	100.0%

Bucket membership histogram

Koliko bucket-a se javlja u tačno k generatora? Visok udeo u k=1 = generatori nezavisni. Visok udeo u k=5 = svi vide isto.

Bucket je u k generatora	Br. bucket-a	% od svih
1	1390	40.5%
2	656	19.1%
3	380	11.1%
4	452	13.2%
5	553	16.1%
Total unique buckets	3431	100.0%

Tumačenje

Avg membership = 2.45 (prosečan bucket je u toliko generatora od 5)
Efektivni broj nezavisnih signala = total_unique / sum(raw) = 3431 / 8415 = 40.8%

Ako se napravi hybrid 5×N=2 (10 slot-ova rasporedjenih po generatorima), stvarni broj nezavisnih trade-ova će biti ~ 4.1 (10 / avg_membership). Ostalo je redundansa.

Provenance

Skripta: tools/analysis/cross_generator_overlap.py
Source: signal_events__growth_threshold_quality_* na .88 prod PG, is_backfill=TRUE
Bucket size: 4h

Izvor: crypto_scanner/reports/analysis/2026-05-20_top5_overlap.md