Gene expression programming for Efficient Time-series Financial Forecasting




Journal Title

Journal ISSN



Volume Title


De Montfort University


Thesis or dissertation

Peer reviewed


Stock market prediction is of immense interest to trading companies and buyers due to high profit margins. The majority of successful buying or selling activities occur close to stock price turning trends. This makes the prediction of stock indices and analysis a crucial factor in the determination that whether the stocks will increase or decrease the next day. Additionally, precise prediction of the measure of increase or decrease of stock prices also plays an important role in buying/selling activities. This research presents two core aspects of stock-market prediction. Firstly, it presents a Networkbased Fuzzy Inference System (ANFIS) methodology to integrate the capabilities of neural networks with that of fuzzy logic. A specialised extension to this technique is known as the genetic programming (GP) and gene expression programming (GEP) to explore and investigate the outcome of the GEP criteria on the stock market price prediction. The research presented in this thesis aims at the modelling and prediction of short-tomedium term stock value fluctuations in the market via genetically tuned stock market parameters. The technique uses hierarchically defined GP and gene-expressionprogramming (GEP) techniques to tune algebraic functions representing the fittest equation for stock market activities. The technology achieves novelty by proposing a fractional adaptive mutation rate Elitism (GEP-FAMR) technique to initiate a balance between varied mutation rates between varied-fitness chromosomes thereby improving prediction accuracy and fitness improvement rate. The methodology is evaluated against five stock market companies with each having its own trading circumstances during the past 20+ years. The proposed GEP/GP methodologies were evaluated based on variable window/population sizes, selection methods, and Elitism, Rank and Roulette selection methods. The Elitism-based approach showed promising results with a low error-rate in the resultant pattern matching with an overall accuracy of 95.96% for short-term 5-day and 95.35% for medium-term 56-day trading periods. The contribution of this research to theory is that it presented a novel evolutionary methodology with modified selection operators for the prediction of stock exchange data via Gene expression programming. The methodology dynamically adapts the mutation rate of different fitness groups in each generation to ensure a diversification II balance between high and low fitness solutions. The GEP-FAMR approach was preferred to Neural and Fuzzy approaches because it can address well-reported problems of over-fitting, algorithmic black-boxing, and data-snooping issues via GP and GEP algorithms





Research Institute