Dilemmas in knowledge-based evolutionary computation for financial investing

Abstract

Can knowledge about financial statements be incorporated in an investing system that improves itself via evolutionary computing? Experiments using neural logic networks and genetic algorithms were implemented. A neural logic network for processing financial ratios and biasing financial forecasts proves resistant to neural network learning. A genetic algorithm for weighting financial attributes and considering industry category also did not demonstrate gradual improvement. The experiments reveal the dilemmas of missing data and inherently unpredictable attribute values. More importantly, the results show the challenges of getting the representation, the fitness measure, and the change operators to mesh in such a way that the search space manifests gradualness. Recommendations include exploiting the declarative nature of Excel programming and involving the user in guiding changes.