The Hedge FundEveryone Can Use.AI-Powered. Institutional-Grade.
GeoHorizon protects and grows capital the way hedge funds do — with institutional-grade risk architecture, AI-powered hedging, and a closed learning loop that sharpens with every market cycle. Hard drawdown guards, regime-gated strategies, and stress-tested position sizing work together to protect your capital first. 89.0% directional accuracy across 47+ scored cycles.
The trading engine isn’t the product — it’s the tool that makes effective hedging possible. Every outcome feeds back as a scored trajectory, continuously sharpening regime detection, strategy calibration, and the risk parameters that govern every subsequent position. Institutional risk management, open to everyone.
Defensive performance through historical crises.
12 new strategies are currently running in shadow mode and being evaluated by the World Model and TIMG. The swarm continuously tests, reflects, and promotes the strongest performers.
Swarm Intelligence
Live agents monitoring in real time
Crypto Leverage
$10,000
Options Strategies
$10,000
Stock Leverage
$10,000
Strategy Simulator
Simulated P&L across 3 strategy accounts
Russia-Ukraine ceasefire stalls
US Treasury sanctions Iranian entities
Fed hawkish surprise — rates held higher
South China Sea tensions escalate
Live Backtest Results
Forecasts scored against real outcomes
Simulated results only — not financial advice. No real capital is deployed on your behalf.
A Different Problem. A Different Design.
The AI-driven markets space is attracting significant capital for good reason — hundreds of billions are at stake in portfolios that remain exposed during regime shifts and correlation breakdowns. Most projects in this category focus primarily on signal generation and alpha within traditional equity markets.
GeoHorizon is built around a more critical and less addressed problem: protecting capital when markets break, while still competing on intelligence and alpha generation.
In short: GeoHorizon solves the hardest part of portfolio management — surviving regime shifts and correlation breakdowns — while building one of the most sophisticated cross-asset intelligence systems in the space.
The Protection Layer That
Activates Before the Market Moves.
GeoHorizon was built on a structural observation: when geopolitical events hit, macro hedge funds hedge in real time — adjusting crypto delta, activating put protection, rotating to defensive exposures — while retail investors absorb the full drawdown with no tools to respond. That asymmetry is structural, not cyclical. It has repeated across every major shock since 2013. GeoHorizon exists to close that gap.
Two dedicated protection swarms — one for crypto, one for equities and ETFs — share a single macro foundation: the Living World Model. Every 4 hours it synthesises regime, posture, velocity, and narrative from 50+ geopolitical signals and publishes to both swarms simultaneously. When conditions shift to CRASH_FEAR, both swarms adapt in the same cycle — hedges activate before positions are caught wrong-footed, not after.
Both swarms learn from every outcome. Every forecast is Brier-scored against real price moves within 8 hours — overconfident agents are recalibrated, and every high-confidence failure feeds back as an Anti-TIMG warning before the next similar setup. Each new event is matched against 48K+ historically similar episodes before the first agent runs. The system that protects capital today is sharper than the one that ran yesterday, and sharper still than the one that ran last month.
Nine Layers of Capital Protection
Each layer answers one question: how does this help the portfolio survive when conditions turn dangerous? Individually, each capability is institutional-grade. Together — connected through the World Model and TIMG learning loop — they form a system that detects regime shifts early, hedges before markets move, and learns from every crisis it has seen since 2013. The key is not the components; it is the closed protection loop that connects them.
Living World Model
The system's early warning radar. Every 4 hours, 50+ geopolitical signals synthesise into a single macro state — detecting when conditions are shifting from stable to dangerous, hours before it shows in price action. When regime transitions to RISK_OFF or CRASH_FEAR, every agent re-prices in the same cycle: hedges activate, leverage reduces, and no position runs against the new posture. Coherence is enforced by architecture — this is how institutional funds avoid being caught wrong-footed by macro events while retail investors absorb the full drawdown.
Adversarial Debate System
The guard against the most expensive mistake in portfolio management: high-conviction wrong bets. A Bull Advocate and Bear Devil's Advocate argue every forecast before a Debate Arbitrator nets the result — capping probability adjustments at ±15pp and widening uncertainty bounds when both sides converge. Structural prevention of the overconfident single-direction position that causes catastrophic drawdowns. When the system can't reach consensus, it defaults to caution.
Strategy Lab
Strategies that stop protecting capital in the current regime are retired before they cause losses. The Strategy Lab continuously monitors (regime, strategy family) performance gaps and proposes replacements via LLM analysis — each candidate validated through a 2-of-3 cross-timeframe walk-forward gate before it can touch a live position. Promoted variants blend in weighted by stability, not raw returns. An anomaly detector triggers automatic self-pause when rolling accuracy drops >15pp below the TIMG baseline. The system knows when it is no longer working, and stops.
Portfolio Construction Engine
The system that prevents catastrophic single-event losses. Hard pre-trade constraints block every position that would breach concentration limits — single-asset ≤40%, correlated group ≤60%, strategy family ≤50%, net portfolio beta ≤1.5×. A five-tier drawdown schedule automatically de-risks at −3% / −5% / −8% / −12%, with full halt at −15%. Position-level stops add a second protective layer: close-only mode at −12%, forced close at −20%. Five macro signals inject into every sizing decision independently of regime classification — so no position is sized as if conditions are stable when they are not.
Episodic Memory & TIMG
The system's institutional memory of every crisis since 2013. Before any agent forecasts, TF-IDF retrieval surfaces the 5 most regime-similar historical episodes — matched by regime, asset class, and crisis signature. The Anti-TIMG layer captures every high-confidence wrong prediction (≥60% probability, wrong direction) as an explicit warning trajectory — injected before the next similar setup to prevent repeating costly mistakes. Nightly compaction distils the most durable crisis patterns into 3× weight lesson nodes. A late entrant with identical architecture starts with zero episodes. This system has already lived through every liquidation cascade, Fed pivot surprise, and sanctions shock since 2013.
Polymarket Alpha-Gap
An early warning signal sourced from prediction markets — where informed participants price risk before it shows in equities or crypto. When the swarm's probability diverges from Polymarket and Kalshi crowd consensus by ≥10pp on material events (≥$25k volume), the gap injects into the Forecaster context. Historical pattern: prediction market divergences have preceded price dislocations. The system captures the signal before mainstream consensus catches up — and before the hedge opportunity closes.
Elite Regime Detector
The regime detector that tells you danger is coming before price action confirms it. An HMM-inspired Bayesian state machine tracks eight mutually exclusive macro regimes across six synthetic timeframes — computing velocity, acceleration, and persistence of each state. Detecting a shift to CRASH_FEAR or LIQUIDITY_CRUNCH hours before it shows in price is the difference between entering a hedge at cost and scrambling for protection at peak volatility. On every confirmed regime transition, prior-regime history is automatically distilled into a 3× weight lesson memory — the system starts every new regime informed by what worked and failed in the last one.
Cross-Asset Correlation Monitor
Correlation breakdowns are the signature of systemic risk events — when assets that normally move independently suddenly converge, it signals a regime shift in progress. This monitor tracks Pearson correlations across seven cross-asset pairs (BTC/SPY, ETH/BTC, SPY/VIX, BTC/GOLD) at 15m, 1h, 4h, and 24h simultaneously. Shift thresholds (±0.15) trigger regime alerts that propagate into the World Model — activating defensive posture adjustments before the breakdown reaches portfolio positions.
Execution Realism
Strategies that look good on paper often fail in practice because friction wasn't modelled honestly. Every simulated position accounts for tiered round-trip slippage (0.10% for BTC to 1.00% for options), accrued 8-hour perpetual funding costs, and per-strike implied volatility for each options leg — so OTM puts correctly carry their full put-skew premium. Pre-trade stress testing runs every proposed position against four historical shock scenarios (2008, 2020, 2022, 2025). If a position would have been catastrophic in a prior crisis at the proposed size, it doesn't open.
Built for Institutional Portfolios.
GeoHorizon meets the standards institutional allocators require: transparent methodology, verifiable track records, stress-tested risk controls, and full audit trails.
-11.1%
Low Drawdown
Maximum drawdown on $15M traditional portfolio vs -20.9% for 60/40 benchmark over the same period.
8 States
Regime-Aware Positioning
Bayesian HMM detects regime shifts across 8 macro states and adjusts position sizing within 1 cycle.
Every Trade
Full Explainability
Natural-language audit log for every entry: signals, regime context, sizing rationale, and counterfactual analysis.
4 Shocks
Stress-Test Resilience
Current rules stress-tested against 2008 GFC, 2020 COVID, 2022 crypto winter, and 2025 tariff shock scenarios.
All metrics are from simulated (paper) trading — no live capital at risk. GeoHorizon is non-custodial and does not hold user funds. Backtested performance figures (11.1% CAGR, −11.1% max drawdown) reflect the four core strategy types — crypto leverage, scalping, equity/ETF leverage, and options. Spot + On-Chain Hedging strategies are in active development and will be incorporated into future performance reporting.
How It Protects Your Portfolio.
Four independent protection layers operate simultaneously — each one capable of overriding the position sizer before capital is ever at risk.
Regime Persistence Filter
When the same macro regime persists for 3+ consecutive cycles, position sizes are automatically reduced 20–40% to protect against mean-reversion risk — before the reversal hits.
Drawdown Recovery Mode
If portfolio drawdown exceeds 5%, all sizing contracts by 40%. Beyond 8%, leverage and scalp trading halt entirely — only defensive macro hedges are permitted until equity recovers.
Portfolio VaR Guard
Every new position passes a pre-trade VaR and concentration check. Single-asset exposure is capped at 40%, correlated groups at 60%. Positions that would breach limits are blocked or resized.
Full Explainability
Every trade carries a natural-language audit log — the exact signals, regime context, sizing rationale, and a counterfactual analysis of what would have happened on the opposite side.
Five Strategies. One Risk Gate.
These aren't five independent trading systems. Each serves a defined role in the capital protection and alpha generation architecture — and all five run under the same unified risk gate.
Each strategy serves the hedging mandate. Options provide direct downside protection. Spot hedges activate in crash regimes. Leverage and scalping generate the returns that fund the hedge. Every position runs through the same four-layer pre-trade gate — no strategy is exempt, no capital is committed unguarded.
| Strategy Type | Regime Gate | Vol / Surface Check | Sizing Method | Learning Feedback |
|---|---|---|---|---|
Crypto Leverage | Risk_On · Neutral — off in CRASH_FEAR, HIGH_VOL | 25D risk reversal; funding rate overlay; HIGH_VOL surface disables | Half-Kelly × regime scalar × funding cost | Thompson Sampling per regime/family; RSM tracks win rate |
Scalping (15–30 min) | Risk_On · Neutral only — disabled in CRISIS, HIGH_VOL | No vol gate — RSI + Fibonacci + MACD required | Fixed × regime mult; ATR dynamic stops 1.5–2.5× | RSM per regime/scalping family; per-hold scored |
Equity / ETF Leverage | Risk_On · Neutral — off in GEO_SHOCK, LIQUIDITY_CRUNCH | VIX term structure; credit spread proxy; 4h+24h agreement | Half-Kelly × drawdown tier; multi-TF confirmation required | RSM per regime/trend family; Brier-scored vs 8h outcomes |
Options Strategies | Full library in RISK_OFF — CRASH_FEAR blocks premium-selling | Per-strike IV (live); VRP ratio; 25D skew; PCR + GEX | Half-Kelly × vol regime check × 4-scenario stress overlay | Thompson Sampling per regime/volatility family; vs premium collected |
Spot + On-Chain Hedging | Spot active in Risk_On / Neutral; on-chain hedges across all regimes — intensified in CRASH_FEAR, HIGH_VOL | On-chain funding rates; DEX liquidity depth; DeFi protocol health; slippage-capped at entry | Kelly-adjusted spot; liquidity-weighted on-chain; IL-aware LP exposure; slippage budget enforced | RSM per (regime, spot/liquidity family); scored vs spot price outcomes and on-chain hedge PnL |
Verifiable. Real-Time. Continuously Learning.
Every agent decision, forecast, and learning cycle is transparent and auditable in real time. No wallet required.
| Event | Asset | Brier |
|---|---|---|
| Russia-Ukraine ceasefire stalls-4.2% pred / -3.9% actual | BTC | 0.08 |
| US Treasury sanctions Iranian entities-2.8% pred / -3.1% actual | ETH | 0.11 |
| Fed hawkish surprise — rates held higher-2.1% pred / -1.8% actual | SPY | 0.09 |
| South China Sea tensions escalate-5.1% pred / -4.7% actual | NVDA | 0.10 |
The live dashboard shows real-time agent activity, backtesting results, World Model state, and the continuous learning timeline — updated every 30 seconds.
Open DashboardDual-Swarm, 15-Stage Pipeline
Two parallel swarms of 9 specialist agents each — one for crypto, one for traditional markets — coordinated by a shared World Model and a deterministic 15-stage orchestration pipeline. Both swarms pass fully enriched context downstream at every stage and learn continuously from Brier-scored backtested outcomes.
Regime Agent
Classifies macro environment in real-time using VIX, yield spread, FOMC proximity, funding rates, and Fear & Greed data. Outputs regime object (risk_on / neutral / risk_off / crisis) that multiplies every downstream decision. Crisis regime expands all hedge multipliers 1.5×.
Ingestion Agent
Aggregates CoinGecko + Chainlink (crypto), yfinance + Polygon (equities), NewsAPI, FOMC calendar, on-chain Aave, DeFiLlama TVL, Coinglass exchange flows, Deribit options flow, Reddit sentiment, and real-time X/Twitter breaking headline detection for geopolitical shocks — all fetched asynchronously in parallel before the agent pipeline runs.
Sentiment Agent
Multi-pass LLM scoring pipeline with source credibility weighting (Reuters 0.92×, Bloomberg 0.87×). Enriched with Reddit community sentiment across 5 subreddits, CryptoPanic news ranked by importance votes, trending narrative detection, and risk keyword scanning.
Forecaster Agent
Three independent Forecaster instances run in parallel across Claude, Gemini, and Grok — each routed by the task-aware ModelRouter to the model that fits the regime and how much confidence the call demands — producing probabilistic forecasts across 4h / 24h / 1W. A confidence-weighted vote merges outputs, and a Reflection Ensemble gate forces posture to neutral when the models disagree. Context pre-enriched with the top-5 regime-similar historical episodes from TIMG, live on-chain intelligence, and Deribit options flow.
Correlation Agent
Maps forecasts to portfolio exposures using crisis-regime correlation matrices. BTC–ETH correlation strengthens to 0.92+ during stress events. Computes per-asset VaR at 24h horizon and flags second-order contagion pathways across DeFi, crypto, and equity markets.
Risk Agent
Aggregates upstream signals into unified risk assessment. Monitors Aave health factors in real-time, applies Kelly Criterion for position sizing, and triggers EMERGENCY_EXIT when health factor < 1.1 or VaR exceeds portfolio thresholds.
Strategy Execution Agent
Selects and sizes positions across crypto leverage, scalping, equity leverage, and options through a mandatory four-layer gate: (1) Regime-Strategy Performance Matrix — 20+ strategies tracked per regime, underperformers automatically down-weighted via Thompson Sampling bandit; (2) vol surface regime check — CRASH_FEAR surface blocks premium-selling regardless of IV rank; (3) Portfolio Constructor hard constraints on concentration and beta; (4) pre-trade stress overlay running every proposed size against four historical shock scenarios before the position is opened. Options legs are priced with per-strike implied vol from live options chains — not a single ATM estimate — so OTM puts correctly carry their put-skew premium.
Adversarial Debate
Three-agent adversarial debate: Bull Advocate argues upside with price targets, Bear Devil's Advocate surfaces cascade pathways and black swans, Debate Arbitrator nets probability adjustments (±15pp cap), injects black swan entries, and widens uncertainty bounds when both sides converge.
Meta-Orchestrator
Coordinates the full pipeline as a clean 9-stage coordinator: async pre-fetch before any agent runs, World Model injection, ensemble forecasting, adversarial debate, Portfolio Construction snapshot, Strategy Lab context, episode persistence, TA signal injection, and a ModelRouter-managed Haiku 4.5 → Sonnet 4.6 fallback chain. Each stage runs as an independent function with a shared _CycleCtx state object, ensuring timing, freshness tracking, and error isolation across all 9 core agents without any single-function monolith.
Four Structural Moats
GeoHorizon’s defensibility is architectural, temporal, and data-driven — not dependent on a single LLM provider or any one data source.
Compounding Data Moat
Every scored forecast generates a labelled crisis trajectory — cross-tagged by regime, asset class, and shock signature. 48K+ episodes compounding every 8 hours. Nightly compaction distils the most durable patterns into 3× weighted protection lessons. A late entrant with identical architecture starts with zero crisis memory. This protocol has already processed every major macro shock since 2013 — each one deepening the system's ability to protect capital when analogous conditions return.
Closed Learning Loop
The protection is not in any single component — it is in the closed loop. World Model state propagates into every agent simultaneously; adversarial debate outputs become calibration trajectories that prevent the same overconfident error twice; Strategy Lab walk-forward results retire strategies before they cause losses; Anti-TIMG injects prior high-confidence failures into the next similar forecast. Each output is the next cycle's protection signal. Replicating individual pieces is achievable. Replicating the closed protection loop is not.
Verifiable Calibration Record
Effective hedging requires forecasts that are accurately calibrated — a hedge deployed on an overconfident forecast wastes capital on risks that don't materialise. Every forecast is Brier-scored against real price outcomes within 8 hours: no cherry-picking, no holdout gaps, no simulation-only validation. Agents that degrade are recalibrated; those that outperform gain higher ensemble weight. The result is an externally auditable accuracy record that any investor can verify — and a hedge engine that activates when conditions genuinely warrant it.
Governance-Aligned Protocol Architecture
GEO token holders govern the decisions that shape the protocol's trajectory — treasury allocation, fee structure within hard-coded ceilings, insurance fund parameters, new market and strategy-family prioritisation, major upgrades, and emergency pause mechanisms. Live risk parameters are deliberately excluded: strategy gates, position sizing, drawdown halts, and concentration limits stay under the World Model, hard-coded safety rails, and multi-sig oversight, so capital protection cannot be voted away by a transient majority. Value accrues through demonstrated capital protection and risk-adjusted performance, not speculative token mechanics. Phase 3 on-chain deployment converts this governance structure into enforceable smart contract constraints.
Not three separate moats — one compounding system. Each cycle sharpens all three simultaneously.
Geopolitical Signal
50+ global news feeds, sanction filings, election outcomes, military movements, FOMC shifts — every signal scored, weighted, and mapped to affected asset classes across crypto and equities.
Dual-Market Hedging
Crypto swarm: GMX shorts, Aave collateral, USDC flight. Markets swarm: protective puts, sector rotation, cash positions, inverse ETFs. All cross-calibrated against the same World Model regime signal.
Non-Custodial by Design
All proposals are unsigned. GeoHorizon never touches your wallet or brokerage. Every recommendation is reviewed before it executes. Full transparency, zero custody risk.
Built on Continuous Learning
- 8-hour backtesting — every forecast scored against real price outcomes across crypto and equities via Brier scoring
- LLM reflection — agents critique their own predictions, update calibration weights, and inject lessons back into the knowledge base
- Episodic memory — TF-IDF retrieval of the 3 most historically similar crises before each forecast cycle
- Concept drift detection — strategy win rates tracked per regime; drifting strategies auto-reduced by 20%
- Strategy Lab — LLM-driven gap analysis proposes variants; shadow-tested for 10 cycles before promotion to live
One Protocol.
Two Markets. Already Running.
Both AI swarms are live, learning, and accumulating intelligence every 8 hours — across crypto and traditional market cycles simultaneously. No wallet required to explore.
Not financial advice. GeoHorizon generates unsigned proposals only. You review and approve every action. Past accuracy does not guarantee future results.