______ _____ _   __   _______ _   _  ___   _   _ _____ 
| ___ \  _  | |  \ \ / /  _  | | | |/ _ \ | \ | |_   _|
| |_/ / | | | |   \ V /| | | | | | / /_\ \|  \| | | |  
|  __/| | | | |    \ / | | | | | | |  _  || . ` | | |  
| |   \ \_/ / |____| | \ \/' / |_| | | | || |\  | | |  
\_|    \___/\_____/\_/  \_/\_\\___/\_| |_/\_| \_/ \_/

■ Polymarket Quantitative Research — Strategy → Live

You have the edge.
I build the system.

You bring a trading idea. I turn it into a fully backtested, optimized, and paper-validated Polymarket bot — ready for live deployment. End-to-end: data pipeline, model, backtest engine, paper trading, live execution.

17%

F1 ROI

2,602

Markets tracked

450+

Models evaluated

654

Validated bets

43.9%

TMAX Hit Rate

F1 markets Weather markets Sports markets Crypto / macro Custom strategies Any Polymarket vertical Live & running

▸ Start a project See live results ⬡ New to Polymarket?

00 Who's building this

Arthur Breguez // @breguez-ai

Quant researcher and software engineer focused on prediction market automation. I build end-to-end ML pipelines that identify and trade real edges on Polymarket — from raw data ingestion through live CLOB execution. Currently running two live systems: an F1 multitask neural network (17% ROI, 654 bets) and a weather ensemble model covering 2,602 markets across 30 cities. All models paper-validated before any capital is deployed.

in LinkedIn ⬡ GitHub 🤗 HuggingFace ▸ Live lab dashboard

04.5 Pricing

RESEARCH

$1,200 flat

Strategy feasibility + backtest report

Hypothesis scoping + data audit
Full walk-forward backtest engine
PnL curve, Sharpe, drawdown, Brier
Champion model + benchmark leaderboard
Written report: edge found or not
Delivered in ~1 week

Best if you want to validate an idea before committing to full build

📞 Includes 30-min onboarding call to align scope and goals

▸ Start research

MOST POPULAR

FULL BOT

$3,500 flat

Complete system — strategy to live

Everything in Research, plus:
Optimized champion model (autoresearch)
Paper trading pipeline + validation gate
Live CLOB execution bot
Bet sizing, edge thresholds, GTC/FAK/FOK
Auto-retraining crons + HF model backup
Full source code, no lock-in
Delivered in ~4 weeks

Includes paper trading validation — no live capital until edge is confirmed

📞 Includes 30-min onboarding call to align scope and goals

▸ Start full build

RETAINER

$600/mo

Ongoing support after delivery

Weekly model retraining on new data
Polymarket API change adaptations
Edge monitoring + drift alerts
New market expansion (same vertical)
Priority response (<24h)
Monthly performance report

Optional add-on after Full Bot delivery. Cancel anytime.

▸ Add retainer

01 How it works — strategy to live in 5 phases

Strategy Brief DAY 1

You describe your edge hypothesis in plain language — no technical knowledge required. "I think F1 pole sitter wins more often when it rains" or "Temperature markets are mispriced on weekend afternoons." I scope the project, identify data sources, and map the hypothesis to testable conditions.

natural language brief feasibility assessment data availability check

Data Pipeline & Backtest WEEK 1

Full data pipeline: market discovery, price history ingestion, feature engineering, and a no-lookahead walk-forward backtest engine. You get a PnL curve, win rate, Sharpe ratio, drawdown, and calibration metrics — everything needed to know if the edge is real.

Polymarket CLOB history walk-forward backtest PnL / Sharpe / drawdown calibration (Brier / ECE)

Model & Optimization WEEK 2

Parameters are optimized using Bayesian search or hill-climbing autoresearch. Models range from calibrated LightGBM ensembles to PyTorch multitask networks. Champion is selected on quote-proxy PnL (simulated real CLOB execution) — not just raw backtest metrics.

LightGBM / PyTorch Bayesian hyperopt autoresearch loop quote-proxy PnL champion promotion gate

Paper Trading Validation WEEK 3–4

The champion model runs live against real Polymarket orderbooks — placing paper bets with no real capital at risk. I monitor edge decay, calibration drift, and fill rate. Only models that pass a rigorous gate (minimum bet count, ROI threshold) advance to live deployment.

live orderbook scoring edge monitoring promotion gate (80+ bets) fill rate analysis

Live Deployment & Monitoring ONGOING

Automated live trading on Polymarket CLOB with configurable bet sizing, edge thresholds, and GTC/FAK/FOK order types. Includes scheduled cron automation, weekly model retraining as new data arrives, and alerts for signals and fills.

Polymarket CLOB execution GTC / FAK / FOK orders auto-retraining crons HuggingFace model backup

02 What gets delivered

Data Infrastructure

Polymarket market discovery + price history ingestion
External data pipeline (weather, sports APIs, etc.)
DuckDB / Parquet warehouse with versioned bronze/silver/gold layers
Automated backfill and incremental sync crons

Research & Modeling

Walk-forward backtest engine (no lookahead)
Probabilistic model with isotonic calibration
Autoresearch loop (hill-climbing hyperparameter search)
Benchmark leaderboard with qp_pnl, Brier, ECE

Trading System

Edge calculator vs live Polymarket CLOB prices
Bet sizing with configurable risk parameters
Paper trading pipeline with settlement tracking
Live execution with guardrails and daily loss limits

Ops & Monitoring

Automated cron pipeline (data sync → score → trade)
Weekly model retraining on new data
Model artifact backup to HuggingFace (private repo)
Full source code delivered — no lock-in

03 Market verticals I work with

🏎️

Formula 1

Race winner · Pole position · H2H driver matchups · Constructor markets

🌡️

Weather

Temperature max/min · Precipitation · Extreme weather events · 30 cities covered

⚽

Sports

Match winner · Over/under · Player props · Tournament outcomes

🏛️

Politics & Macro

Election outcomes · Fed decisions · Economic indicators · Policy markets

₿

Crypto & Finance

Price targets · ETF approvals · Protocol events · Liquidation markets

⚡

Custom

Any Polymarket market family with enough historical price data for backtesting

04 Live proof-of-work

▸ PnL CALCULATOR — scale to your bet size Drag to see projected returns at your position size

$50/bet

⚠ Projected values scale linearly from backtested results (base: $5/bet). Past performance does not guarantee future results. All figures from walk-forward backtests and paper trading — no real capital deployed.

+$344
F1 Paper Trade PnL
654 bets · 17% ROI · 36 GPs

17%

F1 ROI

Brier score 0.084

-$269.27
Weather Gate PnL
3 cities · 223 bets · gate=PASS

43.9%

TMAX Hit Rate

qp_pnl -$48119.07 · Brier 0.1137

2,602

Markets Monitored

30 cities · daily scoring

450+

Models Evaluated

champion gated (rigorous)

654

Validated Bets

paper · 80-bet gate minimum

$ status --champion --all --verbose

● F1 MultitaskQR r4 ROI=17.1% bets=654 Brier=0.084 gate=PASS [CHAMPION]

● TMAX tuned_ensemble hit=35.0% qp_pnl=+$426 brier=0.106 30 cities gate=PASS [CHAMPION]

● GFS forecasts sync=LIVE lag=<15min markets=2602 [STREAMING]

● ECMWF ensemble ifs025 + aifs025 neighbor_spread=active [STREAMING]

● Crons scheduled retrain=post-GP cadence=weekly [ACTIVE]

$ models --leaderboard --top 5

rank model qp_pnl roi brier status

──── ──────────────────── ──────── ─────── ────── ──────

1 MultitaskQR r4 +$25.18 17.1% 0.084 CHAMPION

2 lgbm_isotonic +$18.40 58.1% 0.089 bench

3 qr_ensemble_v2 +$15.22 51.7% 0.094 bench

4 xgb_calibrated +$12.88 44.3% 0.101 bench

5 logistic_baseline +$6.11 21.0% 0.118 bench

$ markets --live --count

2,602 weather markets active across 30 cities — scoring every 15 min

F1 Canadian GP — Sprint Weekend · May 22–24 · Montreal

models published to huggingface.co/artbreguez ✓

▸ Live dashboard → ▸ F1 model (HuggingFace) ↗ ▸ TMAX model (HuggingFace) ↗ ▸ GitHub ↗

05 FAQ

Do I need a technical background to work with you?▶

No. You just need the trading idea and market knowledge. I handle everything on the technical side — data pipeline, model training, backtest engine, and live deployment. The only thing I need from you is a clear hypothesis: why do you think this market is mispriced, and what information do you think predicts the outcome?

How long does a full project take?▶

Typically 3–5 weeks from brief to live deployment. Week 1: data pipeline + initial backtest. Week 2: model optimization and champion selection. Weeks 3–4: paper trading validation. Week 5: live deployment and monitoring setup. Timeline depends on data availability and market complexity.

What if the backtest shows no edge?▶

That's valuable information too. A rigorous no-edge result saves you from deploying real capital on a losing strategy. I'll share the full analysis: what was tested, why it didn't work, and whether there are related hypotheses worth exploring. If the initial hypothesis fails, we can pivot to adjacent ideas within the same market vertical.

Do you trade with my capital or your own?▶

I build and deliver the system — you control the capital and keys. The bot runs on your infrastructure (or a VPS I set up for you) with your Polymarket wallet credentials. I don't have custody of any funds. Live trading starts only after paper trading validates the edge, and always with your explicit approval.

What's included after delivery?▶

Delivery includes full source code, documented pipelines, automated retraining crons, and a private HuggingFace model backup. I also offer ongoing support: model retraining on new data, adapting to Polymarket API changes, and expanding to new markets as your strategy evolves.

Which Polymarket market types work best for bots?▶

Markets with structured, repeating outcomes where external data has predictive power. F1 (race results + qualifying data), weather (forecast models vs market prices), and sports (statistical models vs public betting odds) are proven verticals. Markets with thin liquidity or purely sentiment-driven outcomes are harder to model systematically.

How are results verified — is this real trading?▶

Results shown are from walk-forward backtests (F1: 36 GPs of historical data, no lookahead) and paper trading against live Polymarket orderbooks. The PnL calculator scales these results proportionally to illustrate what different bet sizes would yield. All model artifacts are published to HuggingFace for transparency. Real capital deployment follows paper trading validation — always.

05 Open Source Library — pmlab

The ML framework powering this lab is open source. pmlab is a generic, plugin-based Python library for building, backtesting, and live-trading on Polymarket prediction markets — extracted from the production systems running here and on the F1 lab.

It is market-family agnostic: implement one plugin interface and pmlab handles walk-forward backtesting, gate enforcement, paper trading, Kelly sizing, calibration diagnostics, and report generation.

Install pip install pmlab or uv add pmlab

v0.3.0 301 tests 95% coverage MIT

PLUGIN SYSTEM

MarketPlugin ABC

4 required methods. Auto-discovery via entry_points. Weather, F1, crypto — any market family.

BACKTESTING

Walk-Forward Engine

rolling_origin_eval with strict no-lookahead. Stride-based, configurable, no data leakage.

CHAMPION GATE

Hard GO / NO_GO

ChampionManifest.publish() raises on NO_GO. Segment-level gates. champion.json as source of truth.

SIZING

Kelly + Flat Stake

kelly_stake_size() with fractional Kelly and max_exposure cap. Ready for real capital.

CALIBRATION

Brier Decomposition

Murphy (1973): reliability, resolution, uncertainty, skill score. Reliability diagram data.

EXECUTION

Paper + Live Broker

PaperBroker with dedup and staleness guards. LiveBroker hits CLOB API with dry_run mode.

FEATURES

Transform Toolkit

Lags, rolling stats, sin/cos cyclical, one-hot, outlier clipping — all with group_by for panels.

MODELS

LightGBM + Sklearn

LGBMForecaster and SklearnForecaster. Implement MarketForecaster ABC for any custom model.

quickstart.py

# 1. Implement a plugin for your market
class WeatherPlugin(MarketPlugin):
    family = "weather_tmax"
    def discover_markets(self, **kw): ...
    def fetch_features(self, spec, horizon, **kw): ...
    def fetch_truth(self, spec, **kw): ...
    def build_training_row(self, spec, horizon, **kw): ...

# 2. Walk-forward backtest with no lookahead
result = rolling_origin_eval(panel, LGBMForecaster(), stride=30)

# 3. Gate check — hard NO_GO blocks publication
gate = HoldoutGateResult.evaluate(result.trades, required_segments=["Atlanta", "Buenos Aires"])
manifest = ChampionManifest.publish(model, gate, output_dir="artifacts/")

# 4. Kelly-sized paper trade
stake = kelly_stake_size(win_prob=0.65, entry_price=0.35, bankroll=200.0)
PaperBroker(trades_path="trades.json", flat_stake=stake).record(signals)

▸ PyPI — pip install pmlab ⬡ GitHub — ArtBreguez/polymarket-lab

Available for new projects

Let's build your edge.

Describe your strategy idea and the market you're targeting.
I'll assess feasibility and outline a project plan within 48h.

✓ Responds within 48h

✓ No upfront commitment

✓ Full source code delivered

✓ Paper-validated before live

✉ arthurbreguez@gmail.com in linkedin.com/in/breguez-ai

Paper trading validation required before any live capital deployment.
All model artifacts delivered as source code to the client — no lock-in.

By engaging, you acknowledge the Disclaimer & Engagement Policy.
Past performance does not guarantee future results. No financial advice.

Don't have a Polymarket account yet? Create one here ↗ — referral link

You have the edge.I build the system.

You have the edge.
I build the system.