THE MATHEMATICAL FRAMEWORK

The system is grounded in a well-studied class of stochastic differential equations from statistical physics. These equations describe how prices deviate from, and revert toward, dynamic equilibrium levels. The mathematical framework provides three critical capabilities that together form the analytical backbone of the trading system.

The core innovation of the system. Rather than predicting the final outcome of a trade directly, the framework decomposes each trade into its path-dependent components. This decomposition separates the risk dimension from the reward dimension, and the two turn out to be nearly independent. Predicting them separately, then recombining, produces dramatically better filtering than any direct approach.

The decomposed predictions feed a multi-gate architecture that filters trades across independent dimensions. Because the dimensions are approximately orthogonal, filtering power is multiplicative. Three individually modest filters combine into a single powerful gate. Walk-forward cross-validation confirms positive performance across every fold tested.

Exit placement is not heuristic. The system uses closed-form analytical solutions from a class of partial differential equations that describe how stochastic processes interact with barriers. These equations yield exact probabilities, expected durations, and optimal barrier positions without simulation or approximation.

The system's strongest validated result. Two independent approaches to take-profit placement, one derived from analytical solutions and one from path-dependent prediction, are synthesized into a single hybrid that outperforms either approach in isolation. The combination works because each approach corrects the other's weakness across different volatility regimes.

FX returns exhibit fat tails far beyond what Gaussian models predict. Standard normal assumptions catastrophically underestimate the probability of extreme moves. The system addresses this with extreme value theory, a branch of statistics specifically designed to model the behavior of rare, large events.

All models update incrementally as new data arrives. There is no batch retraining cycle, no overnight model refresh, and no lookahead bias. The system learns continuously, adapting to changing market dynamics while maintaining strict temporal integrity.

The system operates at the informational frontier of its prediction problem. Through extensive experimentation, a clear R-squared ceiling has been established and repeatedly confirmed. This ceiling is not a failure. It is an asset, because it provides a definitive test for overfitting: any model claiming to exceed it is, by definition, fitting noise.