Skip to main content

Grid Trading Configuration

The grid trading strategy uses a separate configuration file config.grid.yaml:

cp config.grid.yaml.example config.grid.yaml

Trading Parameters​

trading:
  symbols: [ETH]                 # trading pair(s) for grid trading
  max_total_investment: 500      # total capital allocated (USD)
  max_leverage: 5                # maximum leverage

  # Grid order features
  grid_limit_order_take_profit_enabled: true   # place TP after grid fill
  grid_limit_order_stop_loss_enabled: true     # place SL after grid fill
  grid_reduce_only_exit_orders_enabled: true   # layered reduce-only exit orders

Exit Order Modes​

OptionDescription
grid_limit_order_take_profit_enabledAfter each grid level fills, place a limit take-profit order at the next grid boundary
grid_limit_order_stop_loss_enabledAfter each grid level fills, place a stop-loss order below the entry
grid_reduce_only_exit_orders_enabledUse reduce-only layered exit orders instead of full close — reduces slippage on large positions
tip

Enable all three exit order modes for full protection. grid_reduce_only_exit_orders_enabled is especially important for larger position sizes.

Agent (AI Decision) Parameters​

agent:
  grid_width:
    min_pct: 0.02           # minimum grid spacing as % of price (2%)
    max_pct: 0.15           # maximum grid spacing (15%)
    fallback_pct: 0.05      # fallback if data is unavailable (5%)
    ai_blend_weight: 0.35   # weight given to AI output vs. market data

Grid Width Blending​

The final grid width is a weighted combination:

final_width = (market_data_width × 0.65) + (ai_suggested_width × 0.35)

The market data component uses ATR and volatility to compute a data-driven width. The AI component allows the LLM to adjust based on broader context (news, sentiment, regime). The 65/35 blend ensures the math engine dominates while AI adds directional nuance.

LLM Configuration​

llm:
  client_type: langchain_nvidia
  model: deepseek-ai/deepseek-v3.2
  temperature: 0.3               # slightly higher for grid — more creative width estimation

Scheduler​

scheduler:
  interval_minutes: 5    # grid re-evaluation interval

Grid decisions run every 5 minutes. Between intervals, the GridManager monitors fills and places exit orders automatically — no LLM call required for fill handling.

Data Configuration​

data:
  timeframe: 15m    # shorter timeframe for grid — more responsive to price action
info

Grid trading benefits from shorter timeframes (15m, 5m) compared to the perpetual agent (1h). This gives more granular volatility estimates for grid spacing.

Complete Example​

llm:
  client_type: langchain_nvidia
  model: deepseek-ai/deepseek-v3.2
  temperature: 0.3

trading:
  symbols: [ETH]
  max_total_investment: 500
  max_leverage: 5
  grid_limit_order_take_profit_enabled: true
  grid_limit_order_stop_loss_enabled: true
  grid_reduce_only_exit_orders_enabled: true

agent:
  grid_width:
    min_pct: 0.02
    max_pct: 0.15
    fallback_pct: 0.05
    ai_blend_weight: 0.35

scheduler:
  interval_minutes: 5

data:
  timeframe: 15m

Grid State Persistence​

The grid manager saves its state to grid_state.json after every order operation. The file is written atomically (via tempfile + rename) to prevent corruption on unexpected shutdowns.

If you restart the bot, it reads grid_state.json and resumes from the last known state — active orders are reconciled with the exchange.

Do Not Edit grid_state.json Manually

Editing grid_state.json directly can cause the grid manager to lose track of active orders, potentially resulting in orphan orders on the exchange.

Safety Mechanisms​

The GridManager includes several built-in safety features:

  • Orphan trigger cleanup — detects and cancels trigger orders with no corresponding position
  • Reduce-only layered exit — exits large positions gradually to minimize market impact
  • Cancel hard timeout (20s) — if a cancel order doesn't confirm within 20 seconds, retries and alerts
  • Atomic state writes — prevents state file corruption on process interruption

Next Steps​