Building Polynomial Prediction Models with GMDH's Multilayer Iterative Algorithm

Summary

The article explains the Group Method of Data Handling (GMDH), focusing on its multilayer iterative algorithm (MIA), and describes an MQL5 implementation for building polynomial prediction models. MIA pairs input variables to create candidate partial models, fits polynomial relationships on training data, and evaluates their prediction errors on a test set. The best-performing candidates supply transformed inputs for the next layer, and the process continues while the layer-level error measure improves. The final model retains selected polynomial components from the successive layers.

The implementation supports linear, interaction, and quadratic forms, and the article outlines model evaluation and stopping criteria. It points to example scripts for time-series and multivariable data, but offers no evidence of trading profitability. The MQL5 version omits multithreaded training and has fewer QR decomposition options than the referenced C++ implementation. The method can require substantial computation, and results depend on data partitioning, model choices, and the selected criterion; the article advises caution when applying it to market data.

Key ideas

• GMDH builds polynomial models from data by generating, evaluating, and combining candidate functions.
• MIA uses pairs of inputs to create partial models and feeds selected outputs into successive layers.
• Candidate models are ranked by prediction error measured on a test set, with training stopping when improvement ends.
• The described polynomial options include linear, interaction, and quadratic forms.
• The MQL5 implementation lacks multithreaded training and has limited solver choices, while iterative model construction can be computationally demanding.

Tags

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