Trader-Company Method:Stock Price Prediction Using Metaheuristics

Time series forecasting using a model imitating financial institutions

Abstract

There are three major challenges in the development of quantitative financial models using machine learning. First, any model can become obsolete in a short period because the agents of the market quickly adapt to new strategies and new models. Second, it is difficult for a general black-box-model to learn the complex set of market rules that governs the behavior of the agents in the market. Finally, a good forecast model tends to be highly complex, and it is often difficult for users to interpret the learned model. In an effort to answer these challenges, we propose Trader-Company method: an evolutionary model that consists of a set of Companies harboring multiple Traders with different strategies. The Company algorithm predicts the returns by combining many Trader models, and each Trader model is a simple financial formula easily understandable for users. Our model reflects the behavior of the financial market, which consists of many weak models. Our model can efficiently obtain profitable trading strategies by directly optimizing parameters that are financially meaningful. Moreover, obtained models are linear combinations of the well-known formulas in financial analysis. Such models are easy to interpret by human users. We will show in the real market data experiments that our model can forecast market behavior with high accuracy.