Full capability inventory.

Every component. Every integration. Every model. All included in the platform — no add-ons, no hidden tiers.

22-model ensemble across 4 tiers

The prediction engine runs all models concurrently via asyncio.gather(). Results are combined by a weighted vote (Tier weights: 25% statistical, 45% ML, 15% sentiment, 15% macro/regime) and then re-weighted by a PPO meta-controller trained via stable-baselines3.

Tier 1

Statistical Models (25% weight)

Model	Method	Signal Output	Primary Use
ARIMA	AutoARIMA parameter selection	Price direction forecast	Short-term trend
GARCH(1,1)	Volatility estimation	Volatility sigma	Position sizing input
Cointegration	Engle-Granger / Johansen	Pair spread signal	NAV arbitrage
Correlation	Rolling 30-day pairwise	Regime-conditional corr	Portfolio construction
Kalman Filter	State-space estimation	Smoothed price signal	Noise reduction
Prophet	Facebook Prophet (seasonality)	Trend + seasonality	Macro cycle signals
VAR	Vector Autoregression	Multi-asset joint signal	Cross-asset dynamics
VECM	Vector Error Correction	Cointegrated system drift	Mean-reversion pairs

Tier 2

Machine Learning Models (45% weight)

Model	Framework	Signal Output	Primary Use
AutoGluon	AutoGluon tabular	Direction probability	General alpha
Gradient Boosting	XGBoost + LightGBM	Directional score	Feature-rich alpha
Random Forest	scikit-learn ensemble	Probability distribution	Stable baseline
LSTM	PyTorch (CPU, no GPU)	Sequence forecast	Temporal pattern recognition

Tier 3

Sentiment Models (15% weight)

Model	Source	Signal Output	Primary Use
FinBERT	Hugging Face (ProsusAI)	Positive/Negative/Neutral	Earnings text, SEC filings
NewsSentiment	News feed NLP	Sentiment score	Macro news impact
MarketSentiment	Market commentary	Composite index	Crowd sentiment timing
SocialSentiment	Social flow analysis	Social momentum	Early breakout detection
UnifiedSentiment	Weighted aggregate	Combined score	Input to ensemble

Tier 4

Macro / Regime Models (15% weight)

Model	Method	Signal Output	Primary Use
HMM	Hidden Markov Model	BULL / BEAR / RANGE / CRISIS	Regime classification
RegimeModel	Multi-state classifier	Regime probability	Orchestrator strategy switch
FactorModel	Fama-French inspired	Factor loadings	Risk decomposition
GaussianProcess	Bayesian inference	Probabilistic forecast	Confidence estimation
PatternModel	Technical pattern detection	Pattern signal	Chart structure recognition

PPO Meta-Controller

A Proximal Policy Optimization agent (stable-baselines3) acts as the meta-controller. Observation space: ensemble predictions, volatility, regime, market microstructure. Action space: continuous tier weight adjustments. Trained continuously against live P&L outcomes. Results in dynamic weight reallocation — in a CRISIS regime, Tier 1 statistical weights increase; in trending markets, Tier 2 ML weights dominate.

LangGraph state machine — a structured debate before every trade

No signal reaches execution without passing through a 7-node LangGraph StateGraph. The deliberation is not cosmetic. A dissent_score > 0.7 triggers unlimited deep research loops (capped at 3). Consensus < 0.2 fast-tracks to allocation. The debate is adversarial by design.

7 Nodes — Summary

research_agent

Perplexity sonar-pro call for grounded market narrative

debate_agent

Bull / Bear / Neutral LLM agents run in parallel; dissent_score computed

deep_research

High-dissent path — focused Perplexity deep dive, injected back into debate

allocator_agent

Kelly Criterion + 4 constraints = AllocationResult

red_team_node

Adversarial GPT-3.5-turbo: find every reason NOT to trade

sanity_guard_node

Price reality check, liquidity check, hallucination detection

execution_agent

IC-level deal memo generation + Decision Gateway routing

Human-in-the-Loop Approval Tiers

Tier	Trigger	Timeout	Channel
Auto-Execute	confidence > 0.85	Immediate	Direct to broker
Trader Review	confidence 0.65 - 0.85	5 minutes	WebSocket cockpit + Telegram
Senior Review	confidence < 0.65	10 minutes	WebSocket cockpit + Telegram

Timeout behavior: all unanswered approvals auto-reject at timeout.

8-stage sequential risk chain. Each stage is an independent veto.

Risk is not a single check. It is a pipeline. Every stage runs in sequence. Any stage can halt the trade. Nothing executes without clearing all 8.

MicrostructureAgent

Veto-capable

Order Book Imbalance (OBI) calculation. Hostile order wall detection (5x average volume = veto). Ensures the market can absorb the trade before models deliberate.

CVaR Risk Manager

Veto-capable

95th-percentile Conditional Value-at-Risk over 252-day rolling window. Position-level and portfolio-level limits enforced simultaneously.

Portfolio Optimizer

Veto-capable

Half-Kelly position sizing. Four constraints: correlation cap (>0.75 vs existing positions = reject), sector concentration limit, inverse-volatility sizing, ATR-based stop-loss anchoring.

ChaosEngine

Veto-capable

Jump-diffusion Monte Carlo: dS = μS dt + σS dW + J·S dN. Three simulation paths (Brownian, fat-tailed Pareto, crisis-cluster correlation). Stress-tests the proposed position across thousands of shock scenarios.

StressTester

Veto-capable

Four historical crisis replays: 2008 Financial Crisis, 2020 COVID Crash, 1987 Black Monday, 2010 Flash Crash. Position fails if worst-case exceeds loss limit.

TailHedgingService

Advisory

Black-Scholes option pricing. Three hedging strategies recommended: OTM Put Spread, Volatility Overlay (long VIX call), Correlation Hedge (gold offset). Does not veto — recommends protective instruments.

RedTeamAgent

Veto-capable

Adversarial LLM (GPT-3.5-turbo) role-playing as a risk-averse analyst. Explicit prompt: find every reason not to trade. Returns veto=bool, kill_reason, risk_flags[].

SanityGuard

Veto-capable

Three checks: price reality (>5% LLM price deviation = hallucination veto), liquidity (order size vs ADV), structural hallucination detection (contradiction with TimescaleDB ground truth).

Direct broker connections. No middleware. No latency overhead.

HedgeVision connects directly to broker APIs using native protocols. Order lifecycle is fully tracked from placement to fill confirmation.

IBKR (Interactive Brokers)

Equities, Options, Futures

Protocol:ib-insync TWS connection

Port:4002 (paper) / 4001 (live)

Order types:Market, Limit, Stop, Bracket

Fill method:Event-driven (orderStatusEvent, execDetailsEvent)

Options:Full chain pricing via qualifyContracts()

IC Markets MT4

Forex

Protocol:REST bridge to MT4 Expert Advisor (localhost:8005)

Symbols:GBPUSD, EURUSD, USDJPY, AUDUSD + full IC Markets roster

Fill method:Polling every 2 seconds until FILLED

Order types:Market, limit, SL, TP via REST body params

Binance

Crypto

Protocol:CCXT singleton

Assets:BTC, ETH, BNB, SOL, ADA + all USDT pairs

Order types:Market, Limit, Stop-Loss, Take-Profit

Fill method:Polling via fetch_order() every 3 seconds

Paper Adapter

Safe Testing

Activated:paper_mode=True in config, or when all live brokers are circuit-broken

Behavior:Simulates fills at live Redis price +/- 0.1% slippage

Recording:Full order lifecycle recorded — identical to live behavior

NAV Arbitrage — Cross-Broker Dual Execution

Z-score spread monitoring. When |Z| > 2.0: simultaneous asyncio.gather(leg1, leg2) across IBKR + Binance or IBKR + IBKR (futures vs equity). Exit when |Z| < 0.3.

Continuous market intelligence. Three tiers of ingestion. Five quality checks.

Data flows through a multi-tier pipeline with strict quality validation before reaching TimescaleDB.

Data Pipeline

Tier 1

Crypto Core

5-minute Binance OHLCV, auto-backfill

Tier 2

Equity / Macro

Hourly — yfinance, Alpha Vantage, FRED API

Tier 3

Gap Assets

On-demand backfill triggered by DataContinuityModule

Quality Gate (5 checks)

CompletenessTemporal orderingPrice logic (High >= Low >= 0)Staleness freshnessContinuity gap detection

Intelligence Modules

Module	Data Source	Output
WhaleScout	Glassnode, CryptoQuant, on-chain	Whale score (40% on-chain / 30% wallet / 30% flow)
TranscriptScanner	Earnings calls, FLTS signals	Semantic earnings surprise signal
OpenBB Adapter	Dark pool prints, FTD data	Institutional flow signal
PerplexitySentimentService	Perplexity sonar-pro (real-time web)	Market intelligence bundle (cached 1hr)
MicrostructureAgent	Live order book	OBI, spread, wall detection, pre-trade veto
Economic Calendar	FMP API	Event-driven signal adjustments

Institutional-grade observability. Fully managed.

Complete visibility into system performance, model health, and trade execution.

Prometheus

Metrics collection (orders, fills, latency, P&L, NAV)

Grafana

Real-time dashboards — order flow, broker fill rates, drawdown

Sentry

Error tracking and alerting

New Relic APM

Distributed tracing across service calls

MLflow

Model experiment tracking, metric logging, run history

Trade alerts, approval requests, circuit breaker notifications

APScheduler — 20 Background Jobs

From 5-minute crypto refresh to weekly ML retraining, all jobs are AsyncIOScheduler UTC-based with no drift. Jobs include crypto_ingestion, trade_lifecycle, signal_generation, position_limits, nav_scanning, pnl_snapshot, model_health, equity_ingestion, EOD sequence (reconciliation, quality gate, features, arbitrage, correlation, cointegration), and weekly ml_retraining (~120min).

Six LLM providers. Automatic fallback. Zero single point of failure.

Circuit breaker rule: 3 failures in 300 seconds marks provider as OPEN. System automatically routes to next in chain. Auto-recovery after 300 seconds + 1 success.

Priority	Provider	Model	Primary Use
1	Groq	llama-3.3-70b-versatile	Debate, critique (speed)
2	Perplexity	sonar-pro	Research, OSINT (web-grounded)
3	OpenAI	gpt-4o-mini / gpt-4o / gpt-3.5-turbo	Deal memos, red team
4	Azure OpenAI	gpt-4o-mini	Enterprise fallback, multi-agent
5	OpenRouter	Free tier models	Low-criticality fallback
6	Google Gemini	gemini-pro / gemini-flash	Last resort, long context

LLM Chain \u2014 Prompt Assembly \u2192 Provider Routing \u2192 Response Parsing

Full capability inventory.

22-model ensemble across 4 tiers

Statistical Models (25% weight)

Machine Learning Models (45% weight)

Sentiment Models (15% weight)

Macro / Regime Models (15% weight)

PPO Meta-Controller

LangGraph state machine — a structured debate before every trade

7 Nodes — Summary

Human-in-the-Loop Approval Tiers

8-stage sequential risk chain. Each stage is an independent veto.

MicrostructureAgent

CVaR Risk Manager

Portfolio Optimizer

ChaosEngine

StressTester

TailHedgingService

RedTeamAgent

SanityGuard

Direct broker connections. No middleware. No latency overhead.

IBKR (Interactive Brokers)

IC Markets MT4

Binance

Paper Adapter

NAV Arbitrage — Cross-Broker Dual Execution

Continuous market intelligence. Three tiers of ingestion. Five quality checks.

Data Pipeline

Crypto Core

Equity / Macro

Gap Assets

Quality Gate (5 checks)

Intelligence Modules

Institutional-grade observability. Fully managed.

Prometheus

Grafana

Sentry

New Relic APM

MLflow

Telegram

APScheduler — 20 Background Jobs

Six LLM providers. Automatic fallback. Zero single point of failure.