Meta-score training — regime_cycle / qcv2 / gtc (2026-05-09)

↑ Nazad

Cilj: posle wired profit-lock-a, da li klasifikator za P(realized_pnl > 0) postavlja AUC > 0.55 i top-bottom quartile spread > 0.10 — pragovi koje tools/analysis/train_meta_score.py koristi za "score has signal".

Trening pokrenut nad prod backup-om (crypto_scanner_prod_20260509_safe.db) preko DATABASE_URL env override-a. Features cache: 73K rows / 96 simbola / 4h iz database/klines_cache.db.features.

Rezultati (walk-forward 5-fold)

Generator	usable	win_rate	AUC	top Q	bottom Q	spread	Verdikt
growth_threshold_regime_cycle	44	29.5%	0.377	20.0%	40.0%	−0.20	NO SIGNAL — ispod random
growth_threshold_quality_cycle_v2	94	34.0%	0.415	30.0%	50.0%	−0.20	NO SIGNAL — ispod random
growth_threshold_cycle (workhorse)	6942	26.9%	0.515	31.4%	29.2%	+0.022	marginalan, ispod 0.10 praga

Interpretacija

regime_cycle i qcv2 su sample-too-small. 44 i 94 usable trade-a nakon walk-forward split-a ostavljaju ~18 i 38 test sample-a po fold-u. Negativan spread (−0.20) je čisti noise — sa toliko malo podataka model nauči šum.
gtc ima dovoljno podataka ali features ne nose signal. Sa 6942 trade-ova, AUC 0.515 je realna procena edge-a, ne sample noise. Top-quartile win 31.4% vs bottom-quartile 29.2% — razlika od 2.2pp u win-rate-u između najbolje i najgore četvrtine kandidata. Premali za ranking webhook prioriteta (treba > 10pp da se realno smatra signalom).
Features set (12 indikatora) nije dovoljan. Trenutni feature columns: EMA fast/slow, RSI, ADX/DI, ATR, BB width, taker_buy_ratio, vol_z. Sve per-symbol per-candle indikatori. Nedostaju:
regime/market features (BTC trend, BTC vol, ALT vs BTC strength)
sector/cohort momentum (corr-cluster signali)
cross-timeframe signali (1h/1d agregati)
macro context (DXY, RVOL, weekend efekat)

Sta ovo znaci za odluke

Ne ramp-uje webhook prioritetiziaciju na ovim modelima. AUC 0.5x ne opravdava overhead-a; rang radije po ML score plus diverzifikacija.
regime_cycle ostaje GREEN-LIGHT za paper-trading bez meta-score filtera. Kline counterfactual rezultat (+9.09%) stoji nezavisno od ML.
qcv2 isti zaključak. Profit-lock je validan; meta-score nije trenutno koristan filter.
gtc: webhook-only rol, meta-score ranking dodaje minimalnu vrednost (+2pp u top-quartile win-rate-u) — može se koristiti samo ako se priority sistema već implementira; nije sam za sebe vredan truda.

Sledeci koraci (research tier)

Dodaj regime features — BTC EMA trend, BTC ATR, ALT/BTC ratio. Ovo direktno meri ono na čemu regime_cycle generator radi pa bi meta-score trebao da hvata sličan signal.
Pre-symbol vs cohort feature — u koje "regime cluster" simbol pripada (high-vol vs trending vs range).
Ponovi sa 2-3x istorije. Posle 3-6 meseci dodatnog paper-trading-a regime_cycle/qcv2 imaće 200+ trade-a što je minimum za stabilnu meta-score procenu.

Reproduce

cd crypto_scanner
$env:DATABASE_URL = "sqlite:///$(pwd -W)/database/crypto_scanner_prod_20260509_safe.db"
.venv39/Scripts/python.exe -m tools.analysis.train_meta_score \
  --generator growth_threshold_regime_cycle \
  --report-json reports/analysis/meta_score_regime_cycle.json \
  --out models/regime_cycle.json

Modeli sačuvani u models/<generator>.json. Slim metrics+weights u reports/analysis/meta_score_<gen>.json.

Izvor: 43_2026-05-09_meta_score_training.md