Systematic trading strategy that predicts Brent crude oil from Middle East prediction market signals.
The oil futures market prices 2.7% probability of Iran escalation. Polymarket prices 58%. We found a one-day lag between prediction market repricing and oil price movement, and built a rolling OLS model to trade it.
Sharpe 1.27 | 58% win rate | 2.16 profit factor | -11.3% max drawdown
Ships with a live institutional-grade dashboard.
The ITO-S3 basket is a portfolio of 15 Polymarket contracts tracking Middle East armed conflict escalation. When the basket moves today, Brent crude follows tomorrow (lag-1 cross-correlation r=0.131, p=0.003).
The model is a rolling 63-day OLS regression on Brownian-bridged S3 returns with a 0.5% signal threshold:
alpha(t), beta(t) = OLS[ brent_ret(t) ~ s3_ret(t-1) ] over [t-63, t]
signal(t) = alpha(t) + beta(t) * s3_bridged_ret(t-1)
LONG if signal > +0.5%
SHORT if signal < -0.5%
FLAT otherwise
Institutional light-theme terminal built with TradingView lightweight-charts:
- Signal panel: current LONG/SHORT/FLAT with implied price and confidence interval
- Strategy performance: Sharpe, cumulative return, drawdown, Calmar, win rate, profit factor
- ITO-S3 basket: current level, multi-period returns, all-time-high indicator
- Price chart: Brent crude (left axis) vs ITO-S3 basket (right axis) with Brownian-bridged interpolation
- Probability table: P(Brent > $X) under Gaussian, Student-t, and 21-day horizon models
- Recent trades: last 20 signals with P&L, color-coded by direction
- Rolling beta & R-squared: 63-day rolling regression diagnostics
- Model info: specification summary
Auto-refreshes every 60 seconds.
git clone https://github.com/Ito-Markets/iran-oil-statarb.git
cd iran-oil-statarb
pip install -r requirements.txtRun the dashboard:
uvicorn src.api.server:app --reload
# open http://localhost:8000Or with Docker:
docker-compose up
# open http://localhost:8000Run the research scripts standalone:
python implied_price_model.py # full model estimation + diagnostics
python charts_and_improvements.py # generate all charts
python plot_brent_vs_s3.py # brent vs s3 overlay chartiran-oil-statarb/
src/
api/server.py # FastAPI backend, serves dashboard + API
data/fetcher.py # loads S3 basket CSV, fetches Brent via yfinance
models/
bridge.py # Brownian bridge interpolation for static S3 periods
ols.py # rolling OLS, signal generation, strategy metrics
probability.py # Student-t tail probability extraction
config.py # constants (window size, threshold, basket code)
dashboard/
index.html # institutional terminal UI
css/terminal.css # light institutional theme
js/app.js # frontend logic, TradingView charts
Data/
basket_level_monthly.csv # ITO-S3 daily levels (1,527 obs, 2022-2026)
last_year_monthly_compositions.csv # monthly contract compositions
lagged_implied_series.csv # pre-computed implied prices
implied_price_model.py # full research script (model selection, diagnostics)
charts_and_improvements.py # strategy variants + chart generation
plot_brent_vs_s3.py # main overlay chart
requirements.txt
Dockerfile
docker-compose.yml
We tested 30+ specifications: single-lag (k=1..10), multi-lag (K=1..5), VAR, error-correction, level regressions. Ranked by AIC and BIC, the simplest model wins: single lag-1 OLS. Adding complexity makes it worse.
| model | K | AIC | BIC |
|---|---|---|---|
| single lag 1 | 1 | -2765.2 | -2756.5 |
| pure S3, K=2 | 2 | -2757.7 | -2744.8 |
| VAR (own lag + S3) | 1 | -2763.5 | -2750.6 |
| pure S3, K=5 | 5 | -2740.6 | -2714.7 |
The S3 basket has 247 days of static prices across 11 segments (binary contracts that don't reprice between events). These produce zero returns that attenuate the OLS coefficient toward zero. The Brownian bridge fills static periods with synthetic noise pinned to the real start/end values, calibrated to active-period volatility.
B(t) = B(t0) + (t-t0)/(t1-t0) * [B(t1) - B(t0)] + sigma * W_bridge(t)
Effect: Sharpe goes from 0.42 to 1.13.
Most daily signals are noise. Filtering to |predicted_ret| > 0.5% cuts trades from 487/year to 63/year while improving every metric:
| metric | no filter | 0.5% filter |
|---|---|---|
| Sharpe | 1.13 | 1.27 |
| accuracy | 52% | 59% |
| max drawdown | -26.8% | -11.3% |
| Calmar | 1.36 | 1.76 |
beta +0.01214 (1% S3 move -> 0.012% Brent next day)
t-stat 2.51
p-value 0.012
R-squared 0.0114
Durbin-Watson 1.95 (no autocorrelation)
Breusch-Pagan 1.17 (p=0.28, homoskedastic)
Ljung-Box 14.1 (p=0.17, no serial correlation)
Jarque-Bera 432.6 (p=0.00, fat tails - expected for oil)
Residuals fit Student-t with df=4.2. The model passes every specification test except normality, which is expected for commodity returns.
| metric | value |
|---|---|
| Sharpe Ratio | 1.27 |
| Cumulative P&L | +43.5% |
| Max Drawdown | -11.3% |
| Calmar Ratio | 1.76 |
| Win Rate | 57.8% |
| Avg Win | +1.94% |
| Avg Loss | -1.28% |
| Profit Factor | 2.16 |
| Total Trades | 89 |
| regime | Sharpe | interpretation |
|---|---|---|
| high S3 vol (ME shock) | +1.75 | strategy earns during crises |
| low S3 vol (calm) | -0.24 | strategy loses in quiet markets |
| delta | +2.0 | behaving as designed |
This is a geopolitical shock harvester. It captures the information transfer from prediction markets to oil that activates during Middle East crises.
| endpoint | returns |
|---|---|
GET /api/data |
full aligned time series (brent, s3, bridged s3, returns) |
GET /api/signal |
current signal, implied price, confidence interval, rolling params |
GET /api/strategy |
sharpe, cumret, maxdd, calmar, accuracy, win/loss stats, equity curve |
GET /api/probabilities |
gaussian vs student-t tail probabilities, multi-horizon table |
GET /api/rolling |
rolling beta, R-squared, alpha time series |
- ITO-S3 basket:
Data/basket_level_monthly.csv- 1,527 daily observations, included in repo - Brent crude: yfinance
BZ=F- fetched at runtime, no API key needed - Polymarket: live contract prices via Gamma API - used in
iran_oil_statarb.py
Two markets, same war, different numbers:
| scenario | oil market | Polymarket | gap |
|---|---|---|---|
| resolution | 25.5% | 1.0% | -24.5% |
| prolonged | 71.8% | 40.8% | -31.0% |
| escalation | 2.7% | 58.2% | +55.5% |
Oil-implied expected CL: $92.73. Polymarket-implied: $117.15. Spread: $30/bbl.
R-squared is 1.1%. The model explains almost nothing on any given day. It works in aggregate over many trades. S3 is at 497 (all-time high), well outside the training range (40-300). The 15 contracts change monthly, so basket level is not a stable unit. Beta is 0.012 - a 10% S3 move predicts 0.12% in Brent. The signal is real but small.
This is research code. Not investment advice. Use at your own risk.
- Fama (1970), Efficient Capital Markets
- Granger & Newbold (1974), Spurious Regressions in Econometrics
- Chordia & Swaminathan (2000), Trading Volume and Cross-Autocorrelations
- Campbell & Thompson (2008), Predicting Excess Stock Returns Out of Sample
AGPL-3.0
Built by ITO - research and structured products for prediction markets.