System Overview

ML-Driven Edge Detection
for Prediction Markets

74 Data Sources  |  Category Ensemble AUC 0.686  |  Backtested Sharpe 3.97

74
Data Sources
335
Features
256,296
Markets
19
Services
55.9%
Win Rate
3.97
Sharpe
Development Pipeline
01Data InfrastructureComplete
74 data sources across 14 categories. 86 harvest scripts. data_processor.py (1,800+ lines) produces 335 features per market. Output: market_features.parquet (256,296 x 335, 33 MB).
02Model Training60%
Category ensemble AUC 0.686 (+6.2%). 6 custom models (Sports 0.955, Health 0.794, Entertainment 0.731). LightGBM 80% + MLP 10% + XGBoost 10%. Autoresearch: 129+ experiments on 3090.
03Live Market FeedComplete
live_scanner.py (1,501 lines). Gamma API polling, 246 features with batch enrichment, model scoring in 4.4 seconds.
04Real-Time FeaturesComplete
246 features generated per market. Classification (instant), financial markets (near-instant), batch enrichment (volume rank, uniqueness, network).
05Edge DetectionComplete
Bayesian blend: P_final = 0.3 * P_model + 0.7 * P_market. Trump-code + MiroFish enrichment.
06Position SizingComplete
Kelly criterion (25% fractional). Max 10% per position, 50% portfolio, 30% per category. 20% drawdown breaker.
07Liquidity FilterComplete
Order book depth, bid-ask spread analysis, slippage estimation. Rejects trades with slippage > 1%.
08ExecutionComplete
Paper trading (default) + live via py-clob-client. Limit orders, manual approval, kill switch.
09P&L TrackingComplete
Sharpe ratio, max drawdown, Calmar, calibration analysis. Category-level breakdown. CSV export.
10Live News & SentimentComplete
7 real-time sources: USGS, HN, Reddit, GDELT, NWS, Google Trends, financial markets.
Phase 1

74 Data Sources

Organized across 14 categories with full 4-year coverage (2023-2026)

Phase 2

Model Architecture

Category-specific ensemble with Bayesian market-price anchoring and multi-agent AI consensus

Base Model
0.745validation AUC
LightGBM · 232 trees · 247 features · 15s training
Category Ensemble
0.686+6.2%
6 custom models routing by market category
MLP Neural Net
0.661
3-layer MLP (128-64-32) · 10% ensemble weight
Bayesian Blend (Live Scoring)
Pfinal = 0.3 × Pmodel + 0.7 × Pmarket
Market price receives 70% weight. Model identifies mispricings via 30% contribution. Adversarial AUC = 1.0 confirms significant distribution shift.
MiroFish AI Consensus
AnalystContrarianInsiderBayesianSuperforecaster
5 agents debate each opportunity via Kimi/Moonshot LLM. Confidence-weighted consensus.
Validation

Backtesting Results

Out-of-sample performance on 7,346 resolved 2026 markets

55.9%
Win Rate
4.49
Profit Factor
3.97
Sharpe Ratio
8.5%
Max Drawdown
Category Performance
CategoryTradesWin RateStatus
Geopolitics1070.0%Best
Other9861.2%Strong
Crypto19456.2%Good
Sports5143.1%Weak
Monthly P&L
Key Metrics
Markets Scanned: 7,346
Trades Taken: 358 (4.9% selectivity)
Avg Edge at Entry: 9.6%
Edge Accuracy: 55.9%
Live System

Paper Trading

Active paper trading with $500 bankroll, scanning every 5 minutes

Bankroll
$500
Deployed
$250
50% exposure
Positions
5
P&L
$0.00
Open Positions
#MarketSideEntrySizeEdge
1BTC above $70K March 16NO0.983$50-27.0%
2BTC above $72K March 16NO0.885$50-24.2%
3No change Fed rates AprilNO0.935$50-24.2%
4BTC reach $75K MarchNO0.855$50-22.3%
5Gen.G vs JD Gaming (LoL)NO0.885$50-21.8%
Documentation

Technical Whitepaper

Comprehensive system documentation covering architecture, data, models, risk management, and competitive edge

Interactive Visualizations

Visualization Suite

11 interactive dashboards — click any card to explore

Satellite Intelligence
🛰️
Global Satellite Map
ESRI satellite imagery with geopolitical markers, live USGS earthquakes, fire zones, chokepoint routes. 17 monitored targets.
Three.js Simulation
🌙
Moon & Sun 3D
Real-time 3D orbiting moon with procedural terrain, glowing sun with corona, starfield. Live phase calculation + trading signal HUD.
3D Mathematical Surfaces
📐
Quant Viz Engine
6 interactive 3D surfaces: Black-Scholes PDE, Heston Vol Smile, Lorenz Attractor, Efficient Frontier, Greeks, Polymarket Surface.
Market Intelligence
📊
Treemap & Sentiment
Bloomberg-style market treemap (size=volume, color=edge). 5 animated gauges: VIX Fear, Crypto F&G, Social Buzz, Stress, Confidence.
Network Analysis
🕸️
Market Graph & Agent Radar
Force-directed market similarity graph. MiroFish 5-agent debate radar showing Analyst, Contrarian, Insider, Bayesian, Superforecaster.
Advanced Analytics
🔬
Solar/Lunar · Sankey · Drawdown
Lunar cycle win rates, trade flow Sankey diagram, drawdown underwater chart with circuit breaker overlay.
Model Diagnostics
🧠
Sunburst · Calibration · Heatmap
Feature importance sunburst, prediction calibration reliability plot, daily win rate calendar heatmap.
Trading Tools
🔧
Kelly · Vol Bands · Order Flow
Interactive Kelly calculator with live sliders, VIX volatility regime bands with FOMC markers, order flow waterfall chart.
Satellite Data
🌍
24 Satellite Features
NDVI vegetation cycle, daylight hours, fire risk heatmap, economic activity gauges, Maxar coverage, urban activity, 24 feature cards.
Real-Time
⚡
Live WebSocket Dashboard
Auto-updating opportunities table, P&L charts, news feed, portfolio panel. Scans every 60 seconds via WebSocket.
System Design

Architecture

Data
74 Sources
→
Features
335 Signals
→
ML Model
AUC 0.686
→
Bayesian
30/70 Split
→
Edge
9.6% Avg
→
Kelly
25% Frac
→
Executor
Paper/Live
News Monitor
7 real-time sources. USGS, HN, Reddit, GDELT, NWS, Trends, Financial.
Trump-Code
100 validated rules. 414 days Truth Social. 6.2hr early signal window.
MiroFish
5 AI agents via Kimi/Moonshot. Analyst, Contrarian, Insider, Bayesian, Superforecaster.
Reference

Commands

Trading Commands
# Full orchestrator python -m services.orchestrator --bankroll 500 --interval 300 # Scan only python -m services.live_scanner --top 20 --min-volume 5000 # Paper trading python run_paper_trading.py --bankroll 500 --loop 300 # Dashboard server python -m services.dashboard_server --port 8080
3090 VM Commands
# Check autoresearch ssh douglaswhittingham@10.0.0.3 "tail ~/autoresearch_log.txt" # Train model ~/ml_venv/bin/python ~/train_model.py # Category models python -m backtesting.scripts.train_category_models # Historical backtest python -m backtesting.scripts.backtest_engine --bankroll 500
Main Dashboard
Hero Stats
Key system metrics. 74 data sources feed 335 features across 256,296 markets. Win rate and Sharpe from out-of-sample 2026 backtesting.
Pipeline
10-phase development roadmap. Click phases to expand. All complete except model optimization (ongoing autoresearch on 3090).
Data Sources
All 74 sources in 14 categories. Green dot = full 2023-2026 coverage. Ranges from Reddit to lunar phases to satellite imagery.
Models
Base LightGBM improved by category routing (+6.2%). Bayesian blend: 30% model / 70% market price. MiroFish = 5 AI agents debating each market.
Backtesting
Simulated trading on 7,346 resolved 2026 markets. Win rate 55.9%, profit factor 4.49, Sharpe 3.97.
Paper Trading
Live simulated trades, $500 bankroll. BUY NO = model thinks overpriced. Edge = model vs market disagreement.
Whitepaper
Full 14-section technical documentation. Click to expand each section.
DASHBOARD LIVE 3D VIZ SATELLITE TREEMAP NETWORK ANALYTICS MODEL TOOLS