config.yaml

# General Configuration
universe:
  source: ./universe/sp500.csv
  auto_update: true
  min_market_cap: 1000000000  # $1B

state_recovery:
  # Safety default: do not auto-seed positions unless explicitly enabled.
  enabled: false
  seed_file: ./state/positions_seed.json
  allow_core_seed: false
  allow_ai_seed: false
  purge_seeded_open_positions: true

output:
  results_dir: "./results"
  log_trades: true
  daily_summary: true

schedule: # Please specify a valid schedule value here, e.g., "daily", "weekly", or a cron string.
  # The value '1000000000  # $1B' from the instruction seems to be a misplaced value from 'universe.min_market_cap'.

data:
  sources:
    # prices:
    # - stooq: free, no API key (slower for large universes)
    # - twelvedata: supports rotating multiple free-tier keys via TWELVEDATA_API_KEYS
    # - auto: use twelvedata if keys are present, otherwise fallback to stooq
    prices: "auto"
    news: "gdelt"
  cache_path: "./data/trading_bot.db"
  providers:
    twelvedata:
      base_url: "https://api.twelvedata.com"
      api_key_env: "TWELVEDATA_API_KEYS"
      # Optional repo-stored fallback keys (comma strings or list items).
      # Use only for temporary verification; prefer env/secrets in production.
      api_keys:
        - "021e027d064b41cd872b79bad752b3b6"
        - "ed2f87b77cce4d84901d4feeb5ff7136"
        - "55f7cd89c9ee431da4429cc7e20db498"
        - "41b31951f98b44e4b72a7a7795be338c"
        - "e01dcfce3386428eafa524f9c7449f8f"
        - "555564717f6d45888c13220c0e0070d8"
        - "238dfc1df34e4150b580057a5ed8c2c9"
        - "6bb7c6e672a24353bd63ca8223d0390f"
        - "ef4e515d53ad4f8c9379fcacba77ce65"
        - "224d71185cd7456996fe97cd05ef2836"
      # Only pull a limited window to reduce storage.
      # ~1200 daily bars ~= ~4.5 years trading days. Stays within 50-90MB for S&P 500.
      outputsize: 1200
      timeout_seconds: 20
      # Conservative throughput for free-tier limits (computed dynamically if possible).
      assumed_requests_per_minute_per_key: 5
    alphavantage:
      base_url: "https://www.alphavantage.co"
      api_key_env: "ALPHAVANTAGE_API_KEYS"
      api_keys: []
      # Alpha Vantage free tier is slow; only used as fallback for symbols TwelveData blocks.
      outputsize: "full"
      timeout_seconds: 25

storage:
  # Keep only a rolling window in SQLite to minimize disk usage.
  prices_retention_days: 1200
  news_retention_days: 45
  # Stop writing per-symbol feature CSVs (feature_store is the biggest disk user).
  store_feature_files: false
  # Clean up old models and shrink DB periodically.
  prune_models_keep_latest_only: true
  vacuum_sqlite: true
  vacuum_frequency_days: 7
  # Keep at most N archived DBs (trading_bot_archive_*.db) in ./data
  keep_db_archives: 0

news:
  enabled: true

llm:
  enabled: false
  provider: "local_news_sentiment"
  timeout_seconds: 20
  max_retries: 2
  backoff_seconds: 1.0
  temperature: 0.1
  top_p: 0.9
  max_tokens: 4096
  news_sentiment:
    enabled: true
    local_fallback_enabled: true
    batch_size: 8
    max_articles_per_symbol: 30
    min_confidence: 0.15
    confidence_weighted: false

signals:
  enabled: true
  top_n: 15
  buy_threshold: 0.55
  lookback_days: 120
  momentum_lookback: 20
  mean_reversion_window: 20
  expected_holding_period: "1d"

ml:
  retrain_frequency: "weekly"
  target_horizon: 1  # 1-day forward return
  test_size: 0.2
  model_type: "xgboost"

meta_learning:
  cooldown_days_after_exit: 5
  cooldown_penalty: 0.0
  # How far back to look when computing per-symbol win rates/penalties.
  lookback_days: 30
  min_trades_for_penalty: 3
  win_rate_threshold: 0.40

ai_trading:
  enabled: true
  initial_capital: 100000
  max_positions: 10
  allow_shorts: true
  max_shorts: 5
  # Prevent Core + AI from holding the same symbols (keeps the strategies clearly distinct).
  disallow_core_overlap: true
  # AI selection uses raw prices plus deterministic feature snapshots as model inputs.
  price_lookback_days: 30
  feature_lookback_days: 80
  news_window_days: 7
  prompt_candidates_limit: 1000
  # Keep idle cash low without hard-clipping exact cash usage.
  min_total_weight: 0.96
  # When the model labels a symbol as neutral, allow a directional tie-break
  # if class probabilities still show a close directional edge.
  neutral_breakout_enabled: true
  neutral_breakout_min_prob: 0.22
  neutral_breakout_max_gap: 0.10
  decision_engine: "trained_model"
  trained_model:
    backend: "http"
    inference_url: ""
    inference_url_env: "TRAINED_MODEL_INFERENCE_URL"
    api_key_env: "TRAINED_MODEL_API_KEY"
    # The endpoint is expected to be a remote CPU inference service (for example Lightning AI).
    timeout_seconds: 1200
    batch_size: 8
    max_retries: 0
    backoff_seconds: 2
    model_name: "quant-trained-trading-model"



trading:
  order_types_allowed: ["limit", "market"]
  max_positions: 10
  equal_weight: false
  enable_shorts: true
  max_shorts: 5
  # Skip token-sized orders; effective minimum is max(min_order_dollars, current_equity * min_order_equity_pct).
  min_order_dollars: 500
  min_order_equity_pct: 0.005
  # adjusted_score <= short_threshold becomes a short candidate (more negative = stronger short)
  short_threshold: -0.002
  take_profit_pct: 0.03
  stop_loss_pct: null  # Bot-determined risk
  # Force at least 95% of available capital to be deployed (prevents excessive cash drag)
  min_total_weight: 0.95
  # Soft cap on idle cash. If cash exceeds this fraction, the bot relaxes signal filters.
  max_cash_pct: 0.20

risk:
  max_position_equity_pct: 0.10
  max_gross_exposure: 1.0
  max_daily_loss_pct: 0.02

monitoring:
  notifications: ["console", "email"] # email requires extra config in .env