Stochastic Optimization for Market Return Prediction Using Financial Knowledge Graph

Abstract

Interactive prediction of financial instrument returns is important. It is needed for asset managers to generate trading strategies as well as for stock exchange regulators to discover pricing anomalies. In this paper, we introduce an integrated stochastic optimization technique, namely genetic programming (GP) with generalized crowding (GC), GP+GC, as an integrated approach for a market return prediction system, using a financial knowledge graph (KG). On the one hand, using time-series data for twenty-nine component stocks of the Dow Jones industrial average, we show that our stochastic local search method can give a better prediction performance by providing a comparison of its return performances with two traditional benchmarks, namely a Buy & Hold strategy and the Moving Average Convergence Divergence (MACD) technical indicator. On the other hand, we use features extracted from a time-evolving knowledge graph constructed from fifty component stocks of the SSE50 Index. These features are used to a GP variant and then incorporate the knowledge extracted from the expression learnt from GP into a KG. Overall, this work demonstrates how to integrate GP+GC with KGs in a powerful manner.