Elsevier

Big Data Research

Volume 14, December 2018, Pages 112-120
Big Data Research

Forecasting Price Movements in Betting Exchanges Using Cartesian Genetic Programming and ANN

https://doi.org/10.1016/j.bdr.2018.10.001Get rights and content

Abstract

Since the introduction of betting exchanges in 2000, there has been increased interest of ways to monetize on the new technology. Betting exchange markets are fairly similar to the financial markets in terms of their operation. Due to the lower market share and newer technology, there are very few tools available for automated trading for betting exchanges. The in-depth analysis of features available in commercial software demonstrates that there is no commercial software that natively supports machine learned strategy development. Furthermore, previously published academic software products are not publicly obtainable. Hence, this work concentrates on developing a full-stack solution from data capture, back-testing to automated Strategy Agent development for betting exchanges. Moreover, work also explores ways to forecast price movements within betting exchange using new machine learned trading strategies based on Artificial Neuron Networks (ANN) and Cartesian Genetic Programming (CGP). Automatically generated strategies can then be deployed on a server and require no human interaction. Data explored in this work were captured from 1st of January 2016 to 17th of May 2016 for all GB WIN Horse Racing markets (total of 204 GB of data processing). Best found Strategy agent shows promising 83% Return on Investment (ROI) during simulated historical validation period of one month (15th of April 2016 to 16th of May 2016).

Introduction

People have loved to bet on various events since the beginning of written history [1]. In particular, many people are gambling on sporting events. In past few years gambling has grown into multi billion industry. In the UK alone, remote (online) gambling sector generated £4.47bn (April 2015 – March 2016). Out of that, online betting and betting exchanges generated total of £1.72bn Gross Gambling Yield (GGY) [2]. This has led to increased interest in forecasting sporting event outcomes using mathematics and statistics for years.

With the rise of betting exchanges in 2000, punters can not only take bets, but also offer their own, in peer to peer fashion. Following the concept of more familiar stock exchange markets. This has attracted a new sort of customer, the full time, high-volume trader who buy and sell odds just like financial traders buy and sell stock or trade on foreign currency exchange. Although, the value of betting exchanges cannot be compared to financial markets, Betfair – the leading betting exchange – processes more than seven million transactions each day – more than all European stock exchanges combined [3].

This research investigates possible advantages of using machine learning with Feed forward multilayer perceptron (MLP) Artificial Neural Networks (ANN) and Cartesian Genetic Programming (CGP) to predict price movements on pre-race GB horse racing markets. The paper has been structured as follows: Section 2 explores current state of the art approaches, Section 3.1 outlines the machine learning and back testing platform and Section 3.2. describes the machine learning models explored. Furthermore, Section 4 reports on the results achieved.

Section snippets

Related work

There have been various attempts to predict the outcome of various sporting events using machine learning – dog racing (greyhound) [4] [5], tennis [6] [7], soccer [8] [9] [10], cricket [11] and horse racing [12] [13] [14].

However, little work has been done on forecasting the price movements inside the betting exchanges. Because betting exchange is very similar to traditional stock or currency exchange markets, there have been couple of attempts using stock market strategies and analysis on

Framework

Fig. 1 demonstrates the overall framework diagram. At first, technical market data, such as price and volume, is extracted from Betfair API. Data is then transferred to a local server on a weekly basis where it is formatted in open-source Protobuf-net for fast deserialization speeds and processing. Additionally, fundamental data such as horse form and running history is imported into the MySQL database. On the local server, each Strategy gathers the necessary data (if any) and develops a model

Results

This section describes results for multiple strategy models. All models were trained on training set for 24 hours using Intel Core i7 3930 K @ 4.2 GHz CPU (100% of CPU utilisation with parallel processing), then evaluated based on overall profit during training set (Table 4) and for unseen data – testing set (Table 5). Furthermore, all models used a limited liability staking plan of 100 units. Which means that with any single bet maximum potential loss is 100 units. Default Back/Lay bet stake

Conclusion

This research has addressed multiple challenges designing and implementing a back-testing and machine learning framework for betting exchange markets. Furthermore, developed entirely new trading strategies using machine learning algorithms.

The proposed framework is generic enough that it can be adopted to various financial markets, however, the only difference is that betting markets are time constrained, i.e., market is finalized as the winner is determined, while most other financial markets

References (30)

  • A.S. Timmaraju et al.

    Game ON! Predicting English Premier League Match Outcomes

    (2013)
  • D. Buursma

    Predicting Sports Events from Past Results. Towards Effective Betting on Football Matches

    (2010)
  • S. Kampakis et al.

    Using Machine Learning to Predict the Outcome of English County twenty over Cricket Matches

    (2015)
  • C. Kempston

    How to Win at the Track

    (2007)
  • S. Pudaruth et al.

    Horse racing prediction at the Champ de Mars using a weighted probabilistic approach

    Int. J. Comput. Appl.

    (2013)
  • Cited by (6)

    This article belongs to Special Issue: HEST4BDAA.

    View full text