Faza 4 — precompute feature-a

Status: deploy gotov, 4/4 validation tests PASS

Posle Faze 4, svi generatori automatski koriste precomputed features iz cache-a (RSI, EMA-20/50/200, ADX/+DI/-DI, BB width, ATR, supertrend, taker_buy_ratio, vol_z, BTC cross-features). Daily catchup automatski recompute-uje features za simbole sa novim klines-ima.

Šta je deploy-ovano

1. Initial features backfill

Pokrenut tools/history_cache/build_features_cache.py --all — preračunato compute_features() za sve (symbol, timeframe) iz klines_cache.db.

Rezultat:

• 13,818,805 features rows persistirano
• 412 / 413 pairs popunjeno (1 pair preskočen — premalo klines za EMA-200 warm-up)
• DB porasla 2.2GB → 6.0GB (3.8GB indikatora)
• BTC cross-features (btc_close, btc_ema_fast, btc_ret_pct) integrisani gde BTC klines postojali
• Trajanje: ~2.5h (sa starom skriptom koja je list()-irala generator pre print-a)

Patch-ovano build_features_cache.py da printuje progres live — future runs imaju vidljiv progres.

2. Auto-recompute u catchup-u (src/klines_cache_maintenance.py)

daily_catchup() sad ima recompute_features=True flag (default ON):

• _async_daily_catchup tracks touched set svih simbola sa novim klines insert-ima
• Posle catchup-a: _recompute_features() poziva feature_store.compute_features() + persist_features() za poslednjih 500 rows po (symbol, interval) — pokriva EMA-200 warm-up sa rezervom
• BTC klines cache-ovan per-interval u jednom run-u, reused cross-symbol
• Idempotentno (INSERT OR REPLACE)

Summary output:

{
  "status": "finished",
  "touched_symbols": ["..."],
  "features_recompute": {"pairs_recomputed": N, "rows_written": M, "errors": 0}
}

3. Generator integracija (već postojala)

_build_context.get_features() u generator_manager.py već postoji od ranije (IMPROVEMENT_PLAN A.1). Cache-first sa fallback-om na compute:

if config is None:
    conn = _features_conn()
    if conn is not None:
        cached = _feature_store.get_features(conn, symbol, tf, start=start_dt, end=end_dt)
        if cached is not None and not cached.empty:
            return cached  # cache hit
# Fallback: compute on the fly
return _feature_store.compute_features(klines, config=config)

Generator code (growth_threshold_quality_*) već poziva get_features(...) — bez izmena.

Validacija (4/4 PASS)

Skripta: tools/analysis/faza4_verify.py

Test 1: cache hit speedup

• get_features() canonical config (cache path): 180 rows u 6.6ms
• get_features() non-canonical config (compute path): 180 rows u 29.5ms
• 4.5x speedup za isti broj redova — cache je pravilno hit-ovan

Test 2: feature_store.get_features direct read

• 180 rows u 4.7ms
• Sve kolone prisutne (ema_fast, rsi, adx, plus_di)

Test 3: daily_catchup auto-recompute hook

• Status: finished
• touched_symbols: 0 (cache up-to-date posle nedavnog backfill-a)
• features_recompute hook se uredno izvršio
• Catchup elapsed: 33s (svi 4 intervala, 139 simbola)

Test 4: feature values consistency (cached vs fresh compute)

• max diff over last 9 rows:
• ema_fast: 0.000028 (na ~$60k BTC = 4.7e-10 relativno, machine precision)
• rsi: 0.000073 (na 0-100 skali = essentially zero)
• adx: 0.002188 (warm-up window razlika u EWM Wilder smoothing-u)
• Tolerancija 0.01 — PASS

Drift od 0.002 u ADX je posledica EWM beskonačne memorije: cached features su computed nad celom 5y istorijom, fresh compute samo nad 30-dnevnim test prozorom. Apsolutno neuticajno za signal generation.

Operativni značaj

Pre Faze 4

• Svaki generator backfill: klines → compute_features() na svaki poziv
• 20+ generatora × 200 simbola × 4 timeframes = isti indikatori računati stotine puta po ciklusu
• Walkforward over 5y mora rekomputovati features svaku iteraciju → minute traje

Posle Faze 4

• Indikatori se računaju JEDNOM pri features backfill-u, persistovani u features tabeli (idempotent kroz INSERT OR REPLACE)
• Generatori čitaju gotov DF iz SQLite-a u milisekundama (4.5x speedup u verifikaciji; realan speedup za heavy multi-indicator backtest može biti 10-100x)
• Daily catchup automatski recompute-uje samo touched (simbol, tf) parove sa 500-row warm-up
• ML meta-score training (IMPROVEMENT_PLAN A.2) sad može da koristi milione feature rows bez recompute-a

Hotspot smanjen za walkforward

Backtest over 5y više ne mora da:

1. Fetch klines (već u cache-u — Faza 2)
2. Compute RSI/EMA/ADX/BB inline (već u cache-u — Faza 4)

Promene parametara koje BIRAJU druge indikatore (canonical config) ne troše ništa — cache se koristi. Promene parametara koje MODIFIKUJU indikator parametre (npr. ema_fast=21 umesto 20) padaju na fallback compute, kao i pre.

Pre-flight provere (sve ✅)

• [x] features tabela popunjena (412/413 pairs, 13.8M rows)
• [x] cache-hit path 4.5x brži od compute path
• [x] BTC cross-features prisutne u svim pairs gde je BTC klines postojao
• [x] catchup auto-recompute hook radi (struktura validirana)
• [x] feature values match unutar EWM warm-up tolerancije

Faza 5 (kasnije, nije ovde)

Multi-exchange podrška kroz ccxt — Kraken, Bybit, OKX, Coinbase. Treba exchange kolonu u klines PRIMARY KEY, sanity-check eksperiment BTC tri-exchange compare, manifest per-generator "exchange" polje.

Provenance

• Izmene: tools/history_cache/build_features_cache.py (live progress), src/klines_cache_maintenance.py (_recompute_features + touched tracking)
• Existing infrastructure: src/feature_store.py (compute + persist), _build_context.get_features u GM, features tabela u klines_cache.db
• Nova test skripta: tools/analysis/faza4_verify.py
• Linked: crypto-scanner-plan-klines-cache, crypto-scanner-session-2026-05-20-faza3