Implementing a Two-Dimensional State Space Model for Trading Data

Summary

This article continues a practical implementation of the Chimera framework, a two-dimensional state space model that learns dependencies across both time and input features. It describes an MQL5/OpenCL neural network layer: input projections are formed separately for temporal and variable contexts, context-dependent parameters are generated, and a kernel computes hidden states and outputs. The article also outlines gradient propagation through the layer and discusses how the model was trained and evaluated on historical market data.

The reported test produced profit while taking long positions with extended holding periods, suggesting the model captured broad trends and filtered shorter fluctuations even on an M1 chart. This is a single implementation and test, not evidence of general profitability or high-frequency suitability. Results depend on the architecture, training data, and test setup, and the supplied excerpt focuses substantially on implementation details rather than a broad comparison or validation study.

Key ideas

• The Chimera framework models dependencies along both the time axis and the feature axis.
• The implementation projects inputs separately for temporal and variable contexts before generating model parameters.
• OpenCL kernels handle hidden-state updates, outputs, and gradient distribution.
• The reported test showed profitable trend-following behavior with long holding periods on an M1 chart.
• The result is specific to the tested model and historical data and does not establish general performance.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.