Financial mathematics, reinforcement learning, and optimal position sizing. Our research is validated through the emulator before any production deployment.
The mathematical foundation of every position we take. No hardcoded lot sizes, no intuition — just optimal growth.
f* = fraction to bet · b = net odds · p = win probability · q = loss probability
The Kelly Criterion maximizes the expected logarithm of wealth. In trading, it prevents overbetting (risk of ruin) and underbetting (suboptimal growth). Every strategy's Kelly fraction is computed dynamically based on historical win rates, current volatility, and regime detection.
We treat trading as a sequential decision problem. The agent learns optimal policies through continuous interaction with market data.
State: market indicators, order book, portfolio. Action: buy/sell/hold quantities. Reward: PnL with risk penalty. Experience replay breaks temporal correlations.
Direct policy optimization for continuous action spaces. PPO and A2C implementations with generalized advantage estimation (GAE).
Multiple RL agents operating on different timeframes and instruments. Centralized critic, decentralized execution (CTDE) architecture.
Every indicator is computed in zero-copy Go pipelines. No unnecessary allocations — every nanosecond counts.
Relative Strength Index with customizable periods and smoothing. Wilder's and exponential variants. Oversold/overbought regime detection.
Moving Average Convergence Divergence. Signal line crossovers, histogram divergence, and momentum regime classification.
Volatility-based envelopes. Band width indicates market regime. Squeeze detection for breakout prediction.
Average True Range for volatility-adjusted position sizing. Used as direct input to Kelly Criterion calculation.
Real-time order book imbalance, depth pressure, and bid-ask spread analysis. Microstructure signals for short-term prediction.
Volume-weighted price levels. High-volume nodes as support/resistance. Delta analysis for directional conviction.
Models don't stay static. They learn across trading sessions, adapt to new regimes, and continuously improve.
Model-agnostic meta-learning (MAML) enables rapid adaptation to new market conditions. After just K=10 episodes in a new regime, the model achieves near-optimal performance. The outer loop runs continuously on the Worker service.
Knowledge from liquid markets transfers to less liquid ones. Feature extractor layers are frozen, only the policy head is fine-tuned. Reduces cold-start time by 80%. Cross-exchange and cross-instrument transfer supported.