The COMBI Algorithm for GMDH Model Selection and Time-Series Prediction

Summary

The document explains COMBI, a single-layer form of the Group Method of Data Handling. It generates candidate linear models from combinations of input variables, estimates their coefficients on a training sample, and uses performance on a separate test sample to choose a preferred structure. This differs from the multilayered MIA approach described as a related GMDH method. The article presents the MQL5 class structure, including linear prediction, combination generation, fitting, and prediction for time series or multivariable inputs.

The examples include scripts for simple and multivariable datasets and a practical application to Bitcoin daily prices. The discussion indicates that model settings and criteria require tuning, and frames the financial application as evidence of potential rather than proof of reliable forecasting. No detailed out-of-sample results, trading simulation, or profitability analysis is provided in the supplied text. The method is therefore useful as a model-building and selection framework, but its predictive value depends on data preparation, validation design, and parameter choices.

Key ideas

• COMBI builds candidate linear models from combinations of the available inputs.
• Model coefficients are estimated on training data, while a separate test sample selects among candidate structures.
• COMBI uses a single layer, unlike the multilayered MIA approach discussed for comparison.
• The MQL5 implementation supports fitting time series and multivariable datasets.
• The Bitcoin price example illustrates an application, but the supplied material does not establish trading profitability.

Tags

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