Skip to content

S10_THE_REAL_DEAL_PLAN.md

Latest commit

 

History

History
136 lines (92 loc) · 4.1 KB

S10_THE_REAL_DEAL_PLAN.md

File metadata and controls

136 lines (92 loc) · 4.1 KB

🚀 Sprint-10 — "The Real Deal"

Tarih: 15–31 Ekim 2025
Owner: @siyahkare
Amaç: Mock'tan production modeller + gerçek veri düzenine geçiş. ccxt/MEXC akışı, gerçek LGBM/TFT eğitimleri, backtesting çerçevesi ve live trading hazırlığı.

🎯 Hedefler

  • Gerçek veri: ccxt ile REST + WebSocket; stabil akış
  • Production modeller: LightGBM (joblib) + TFT (PyTorch Lightning)
  • Backtesting: 90 günlük geriye dönük simülasyon & rapor
  • Live-Prep: testnet/paper order flow + kill switch

📦 Kapsam

İçerir: veri ingest, eğitim, inference, backtest, testnet entegrasyon, metrikler
Dışında: On-chain oracle & vault (Sprint-11+), mobil arayüz

🧩 Epic'ler ve Görevler

A) Real Data Ingestion (ccxt + Stream)

  • A1: adapters/mexc_ccxt.py (REST OHLCV + orderbook snapshot)
  • A2: WebSocket ticker/trades/mini-agg stream + reconnect/backoff
  • A3: Feeder → engine md sözlüğüne: price, spread, vol, texts (placeholder), funding, oi
  • A4: Time-sync & gap-fill (minute bars; retry)

DoD: 24 saat kesintisiz akış, dropped < %0.1, p95 parse < 5ms

B) Production LGBM (Optuna + joblib)

  • B1: Feature store (Parquet/DuckDB) — leak-safe time split
  • B2: Optuna study (≥ 200 trials), early stop
  • B3: best_lgbm.pkl → EnsemblePredictor load (joblib)

DoD: acc ≥ %65 (val), leakage 0, model card oluşturuldu

C) Production TFT (Lightning)

  • C1: Sequence dataset builder (windowed, scaling)
  • C2: Trainer + early stopping + ckpt + AMP opsiyon
  • C3: best_tft.pt → inference wrapper (thread-safe)

DoD: p95 inference ≤ 40ms (CPU), val-score ≥ baseline LGBM-5%

D) Backtesting Framework

  • D1: backtest/runner.py (vectorized minute sim)
  • D2: Slippage/fee/spread parametrizasyonu
  • D3: Rapor: HTML/MD (Sharpe, Sortino, MDD, turnover)

DoD: 90g BTC/ETH raporu; seed-reproducible sonuçlar

E) Live Trading Prep (testnet)

  • E1: Order adapter (idempotent clientOrderId, rate-limit)
  • E2: Circuit breaker & kill switch (global stop bağla)
  • E3: Paper→Testnet staging pipeline (feature flag)

DoD: 48 saat testnet/paper sorunsuz yürüyüş; 0 fail order

🧪 Test Stratejisi

  • Unit: adapters, feature store, trainers, backtester (≥ %80)
  • Integration: ccxt ingest → engine → signal → risk flow
  • Regression gate (CI): backtest KPI drop > %10 → fail

📈 KPI Hedefleri

Metrik Hedef
Accuracy (val) ≥ %65
Sharpe (90g backtest) ≥ 2.0
Max Drawdown ≤ %12
Inference p95 (CPU) ≤ 40ms
Uptime (paper/testnet 48h) ≥ %99

📅 Takvim

Gün Epic Çıktı
15–17 Eki A REST+WS akış, feeder stabilize
18–21 Eki B LGBM eğitim + load
22–24 Eki C TFT eğitim + inference
25–27 Eki D Backtest + rapor
28–31 Eki E Testnet run + kill switch

🔐 Riskler & Önlemler

  • WS kesintisi: exponential backoff + snapshot replay
  • Leakage riski: time-based split + feature lag testleri
  • Latency spike: batch predict + cache + async pool
  • Slippage hatası: spread/fee param'ları; duyarlılık analizi

🧾 Çıktılar (Dosyalar)

backend/src/adapters/mexc_ccxt.py
backend/src/data/feature_store.py
backend/src/ml/train_lgbm_prod.py
backend/src/ml/train_tft_prod.py
backend/src/backtest/runner.py
reports/backtests/90d_{SYMBOL}.html

✅ Definition of Done (Sprint-10)

  • ccxt ingest 24h stabil, dropped < %0.1
  • best_lgbm.pkl & best_tft.pt gerçek eğitimden
  • 90g backtest raporları üretildi (BTC/ETH)
  • Testnet/paper 48h sorunsuz, kill switch doğrulandı
  • KPI hedeflerinin %80'i sağlandı

Prepared by: @siyahkare
Sprint: S10 — The Real Deal
Status: 🚀 KICKOFF!