Backtest slippage modeling + offline point-in-time LLM signal eval

[biswajeetdev]

What

Two improvements surfaced from a review of the trading bot against current open-source LLM-trading practice:

1. Slippage modeling in backtest.py

Backtests were commission-only (0.05%/side), so net fills were optimistic. Adds a configurable SLIPPAGE (0.05%/side, override via params["slippage"]) applied adversely at every fill point — BUY entries fill higher; SELL/STOP/TARGET/EOD exits fill lower. Commission + slippage are now printed in the run header.

2. eval_llm.py — offline point-in-time eval of the LLM debate brain (new)

The rule layer is validated by backtest.py/god_mode.py, but the LLM debate layer (signals/debate_brain.py) — the bot's claimed source of alpha — was previously scored only by live online loops (rag/strategy_evolver, self_improver). This harness gives it an offline test:

• Replays historical as-of dates, rebuilds point-in-time price/technical indicators (rolling windows are backward-only), and calls the real debate_decide brain.
• Scores each call against realized forward returns: per-action avg forward return, directional hit-rate, and BUY edge vs baseline.
• Look-ahead-bias guard: --anonymize masks the ticker so the model can't recall a known stock's trajectory (cf. arXiv 2510.07920 Profit Mirage, 2601.13770 Look-Ahead-Bench); warns about pre-cutoff memorization.
• Honest scoping: only reconstructable price/technical context is fed; live-only alt-data (social, Congress, options flow, VIX, fundamentals) is passed empty.

Verification

• Slippage: returns drop monotonically with trade count (NVDA −0.06% / 12 trades, MSFT −0.02% / 5 trades).
• eval_llm.py: end-to-end smoke test (NVDA, 2 as-of dates) ran clean — point-in-time indicators → real brain → forward-return scoring → report.

Not included (follow-ups)

• TradingAgents-style explicit risk-debate round.
• Crisis-scenario stress sizing.

🤖 Generated with Claude Code