Public Alpha

Quantize News,Predict Alpha

Stop guessing. Use RL-aligned agents to predict volatility and automate portfolio rotation. A professional-grade platform for strategy development.

Start Backtest GitHub

System operational

Python SDK

train_policy.py

from olpsbandit.llm import GRPOTrainer, NewsAligner

# 1. Align news tokens with price action
dataset = NewsAligner.sync(tickers=["NVDA", "AMD"], lookback="2y")

# Train Policy on Returns
trainer = GRPOTrainer(model="qwen-3-8b-fin", reward_func="sharpe")

# Forecast Quantiles
forecast = trainer.predict(live_stream)
print(forecast.p90_confidence) 
# >> 0.87 (HIGH CONVICTION)

Backtest Results

/backtest-preview.png

Live Monitor

/live-preview.png

Strategy Editor

/editor-preview.png

NVDA $920.15TSLA $175.30AAPL $145.60MSFT $299.35GOOGL $2,850.00$OLPS +12.4% (LIVE)SPY $512.20NDX $13,450BTC/USD $64,230ETH/USD $3,450

Built for Modern Quants

A complete ecosystem for researching, backtesting, and deploying agent-based strategies.

News-Aligned LLMs

New

Train models on price-aligned news data using RLVR rewards to forecast return quantiles with institutional precision.

VaR (95%): -2.1%

Exp. Return: +4.8%

Backtesting Engine

Event-driven simulations with realistic slippage, liquidity models, and transaction costs.

Portfolio Optimization

Mean-Variance and Black-Litterman models to find the efficient frontier.

Bandit Algorithms

Reinforcement learning to auto-balance exploration vs exploitation.

Market Screener

Filter universe by fundamental ratios and technical signals.

Risk Management

Hard constraints on drawdown, exposure, and beta.

Live Performance

Real-time leaderboard of deployed strategies.

View Dashboard

Strategy	Return	Sharpe	Avg Hold	Trend (1D)

Deep Learning

Price-Aligned News & Quantile Forecasting

Standard sentiment analysis is noisy. We use Reinforcement Learning from Verifiable Rewards (RLVR) to align large language models directly with market returns.

1Ingest news from 50+ financial sources
2Output probability distributions (quantiles)
3Execute on high-confidence P90/P10 divergences

Live News StreamToken Attention

"Fed signals rate cut..." Bullish (0.89)

Projected Return Quantiles

strategies/bandit_ensemble.py

# Initialize Strategy Ensemble
bandit = ContextualBandit(
    arms=[
        MeanReversion(lookback=20),
        TrendFollowing(ema_span=50),
        SentimentModel(source="reuters")
    ],
    policy="LinUCB"
)

# Adaptive Allocation
allocation = bandit.select_arm(live_stream)

Real-time Weight Adjustment

Updating every tick

Exp4.P AlgorithmRegime: Trending

LIVE

PORTFOLIO WEIGHT

100%

50%

Sentiment

Mean Reversion

Trend Follow

T-24HT-12HT-6HNOW

Adaptive Ensemble

Solve the Non-Stationarity Problem

What worked yesterday often fails today. Markets shift between regimes: trends, volatility, liquidity. Instead of relying on a static strategy, we treat every signal as an "arm" in a Multi-Armed Bandit problem.

Smart Exploitation

The system automatically shifts heavy capital allocation to the "hot hand". Reward functions (e.g. Sharpe ratio) allow for regime specialization.

Continuous Exploration

Minimize regret by testing underperforming strategies with small capital, ensuring you never miss a regime shift.

Contextual Bandits

Unlike simple bandits, our agents observe market context (volatility, sentiment, volume) to predict which arm will perform best before the move happens.

Ready to outsmart the market?

Clone the repo, install the SDK, and start training your first agent today.

Read Documentation