Grammar-guided genetic programming

The plan is to improve the two existing methods in the prototype and additionally implement three more recent ones that claim to have better performance: piGE , DSGE and WHGE . Once all five methods are stable, suitable benchmark problems will be collected according to guidelines developed by the genetic programming community. This will allow to compare the methods objectively, analyze their sensitivity to hyperparameter settings and identify the best choice of method and parameter combination for tackling different use cases. All that functionality will be integrated and provided in form of an open-source software package on GitHub and PyPI. The most performant method will then be onboarded on the SingularityNET platform as general-purpose optimization service that can be used to solve hard optimization problems in a wide range of domains.

The plan is to apply the best grammar-guided genetic programming method in a real-world setting: evolutionary search for robust trading strategies, presumably many of which can not be designed by human traders or other machine-learning approaches. This essentially requires the formulation of 1) context-free grammars to define search spaces of candidate strategies and 2) an objective functions to correctly evaluate the quality of candidate strategies. The second task is suspected to be the main challenge, because properly evaluating candidate trading strategies requires a non-trivial setup of backtesting simulations together with proper statistical evaluation techniques to prevent overfitting and suitable risk-adjusted performance metrics to optimize the correct goal. Only if that setup is done right, the evolved trading strategies will have good performance both on the training data but more importantly also on future data in live markets.