Geopolitical Trading Bot
Build History

Full Python bot with signal cycle, corroboration filter, risk manager, trade arbiter, position tracker, and exit manager. Connected to Interactive Brokers via ib_insync on paper account port 7497.

Live options chain data from IBKR for LMT, RTX, USO, GLD, NVDA, TSM. Scores based on: unusual call/put volume (>2× 20-day average), put/call ratio direction, and IVR (implied volatility rank vs 52-week history). IVR ≥ 80 caps score at −0.3 as a volatility warning. Returns 0.0 on IBKR disconnection.

Three separate vol strategies running in parallel: (1) Oil dispersion — crude IV vs energy equity IV mismatch; (2) EM panic harvesting — sell overpriced puts on EWW/EPOL during temporary shocks; (3) Safe haven skew — GLD/FXY ratio spread when crisis pricing diverges.

strategy_optimizer.py logs all signal scores, decisions, and trade outcomes to .jsonl files. Periodically sends full history to Claude via Anthropic API. Claude analyzes what's working and recommends updated parameter values (signal weights, profit targets, thresholds). Bot applies changes to config.py automatically.

Third AI layer added to bot_health.py, running every 6h. Completes the three-layer health stack: Layer 1 (deterministic diagnostics, instant) → Layer 2 (Claude health analysis, every 30 min) → Layer 3 (Claude parameter optimization, every 6h). Layer 3 ingests signal non-zero % over 48h, per-signal trade outcome attribution, Layer 2 health findings, and current market regime. Asks Claude specifically: "what parameter changes would improve signal quality right now?" Returns structured JSON proposals covering signal weights, corroboration thresholds, and profit targets. Changes staged to pending_optimizations table — weight changes auto-approved within bounds, threshold changes require human approval via dashboard. Applied changes hot-reload via importlib.reload(config) so the next signal cycle (up to 15 min later) picks up new values without restart. Guard rails: no single weight below 0.03 or above 0.50, weights must sum to 1.0 after rebalance, max 2 parameter changes per run, disabled during open positions, changes expire after 12h without approval. Full change history written to optimization_history.jsonl. /rollback command available via Telegram.

Immutable per-trade record with: entry/exit timestamps, signal scores at time of decision, which sources agreed, VIX at entry, NAV at entry, exit reason, strategy attribution. Written on entry, updated on exit. Never edited — establishes provable track record.

Two parallel bot instances running simultaneously at all times. Live instance trades real capital on IBKR. Dev instance runs paper trading. Dev performance is directly comparable against live at any time — no guessing whether a change improves outcomes.

Historical geopolitical and macroeconomic events classified into three tiers × three sub-tiers = nine severity levels. Each level has corresponding real historical market data. Green 1–3: stable conditions, capital mostly parked in yield instruments. Yellow 1–3: elevated tension, capital shifting toward active deployment. Red 1–3: crisis conditions, full deployment mode. Used to stress-test the bot before any update goes live.

The nine tiers map to a normal distribution curve. Most days cluster in the middle (Green and low Yellow). True crisis events (Red 3) are rare outliers on the far tail. The system displays where TODAY sits on this curve in real time — a live marker on the distribution showing current geopolitical risk score. This drives capital allocation decisions automatically.

When conditions fall outside the normal distribution beyond Red 3 into true outlier territory, a dedicated outlier protocol fires. System cross-references closest historical analogues and executes with conviction while most participants are paralyzed. A dedicated capital reserve (10–15% of fund) is held specifically for outlier deployment and never touched in normal conditions. Outliers are the highest-earnings-potential events in the system.

Multiple AI models — Claude Opus, Claude Sonnet, ChatGPT, Gemini, Perplexity, Grok — run simultaneously in paper trading accounts, each generating signals independently. Performance compared in real time against the live baseline. Answers a question no platform currently asks: which AI model actually makes better trading decisions? Proven with real paper trading data, not benchmarks.

Beyond competition, models operate collaboratively: one model generates a signal hypothesis, another critiques or refines it, the result feeds back to the first for iteration. Creates a cross-model feedback loop that catches blind spots individual models miss. All iterations logged for analysis and contribute to the signal optimization database.

Capital is always in one of three states: earning yield in liquid safe instruments, partially deployed in active positions, or fully deployed on strong signals. Zero idle cash. The allocation engine reads the current position on the geopolitical distribution curve and automatically adjusts the ratio between safe parking and active deployment as conditions shift. Green: majority in T-bills/HYSA/dividend ETFs. Yellow: capital migrating toward active. Red: maximum deployment. Outlier: crisis reserve fires.

Tab 1 — Strategy management: live trading column + dev/paper comparison columns (Strategy X, Y, Experimental). Tab 2 — AI benchmarking: each configured model's paper trading performance vs live baseline, collaborative mode toggle. Tab 3 — Historical curve explorer: interactive normal distribution with all historical events plotted, live today marker, click-through event details, dropdown filter by category. Tab 4 — Settings: IBKR params, AI API keys, risk thresholds, deployment ratios, notification prefs.