A Dynamic Fuzzy Money Management Approach for Controlling the Intraday Risk‐Adjusted Performance of AI Trading Algorithms

• Author: Vince Vella,Wing Lon Ng
• DOI: http://doi.org/10.1002/isaf.1359
• Date: 01 April 2015
• Published date: 01 April 2015

A DYNAMIC FUZZY MONEY MANAGEMENT APPROACH FOR

CONTROLLING THE INTRADAY RISK-ADJUSTED

PERFORMANCE OF AI TRADING ALGORITHMS

VINCE VELLA AND WING LON NG*

Centre for Computational Finance and Economic Agents (CCFEA),University of Essex, Colchester, UK

SUMMARY

The majority of existing artificial intelligence (AI) studies in computational finance literature are devoted solely to

predicting market movements. In this paper we shift the attention to how AI can be applied to control risk-based

money management decisions. We propose an innovative fuzzy logic approach which identifies and categorizes

technical rules performance across different regions in the trend and volatility space. The model dynamically pri-

oritizes higher performing regions at an intraday level and adapts money management policies with the objective to

maximize global risk-adjusted performance. By adopting a hybrid method in conjunction with a popular neural net-

work (NN) trend prediction model, our results show significant performance improvements compared with both

standard NN and buy-and-hold approaches. Copyright © 2014 John Wiley & Sons, Ltd.

Keywords: artificial neural network; dynamic moving average; fuzzy clustering; high-frequency trading; money

management

1. INTRODUCTION

The efficient market hypothesis (EMH) (Fama, 1965, 1970), which resulted in Eugene Fama being

awarded the Nobel Prize in Economics in 2013, remains without any doubt one of the most debated

theories. According to EMH and the implied random walk of asset prices, the use of trading rules on

historical prices should not result in excess returns after accounting for transaction costs. However, al-

though the EMH is a strong landmark in the finance and economics literature, the application of tech-

nical rules is a widespread practice. Research claims from a number of authors (LeBaron, 1999;

Schulmeister, 2006; Gradojevic & Gençay, 2013; Han, Yang, & Zhou, 2013; Holmberg, Lönnbark,

& Lundström, 2013) indicate a relationship between the possible periodic breakdown of market effi-

ciency and volatility. We extend this research at a more granular intraday level by identifying a link be-

tween the profitability of technical rules with different levels of price volatility at short intraday time

horizons and explore how this information can improve the risk-adjusted performance of high-

frequency trading algorithms. To our best knowledge this has not been investigated in the literature

so far.

Our approach stems from two gaps in the computational finance literature. First, existing literature

related to algorithmic trading is used to focus on return (only), putting risk (if at all) second; see surveys

by Krollner, Vanstone, and Finnie (2010) and Tsai and Wang (2009). Although the proliferation of

high-frequency data led to new measures of variation (Andersen, Bollerslev, Diebold, & Labys,

* Correspondence to: Wing Lon Ng, Centre for Computational Finance and Economic Agents (CCFEA), University of Essex,

Wivenhoe Park, Colchester CO4 3SQ, UK. E-mail: wlng@essex.ac.uk

Copyright © 2014 John Wiley & Sons, Ltd.

INTELLIGENT SYSTEMS IN ACCOUNTING, FINANCE AND MANAGEMENT

Intell. Sys. Acc. Fin. Mgmt. 22, 153–178 (2015)

Published online 7 November 2014 in Wiley Online Library (wileyonlinelibrary.com)DOI: 10.1002/isaf.1359

2001; Barndorff-Nielsen & Shephard, 2004, 2006), making it possible to be predicted with a good de-

gree of accuracy up to an intraday level (Andersen, Bollerslev, & Cai, 2000), the use of this information

for intraday trading purposes is rarely considered. In our opinion, reverting to artificial intelligence (AI)

models solely for market movement predictions with little consideration to the time-varying market un-

certainty (Son, Noh, & Lee, 2012) portrays an incongruent view from financial markets practitioners,

since investors are mostly interested in risk-adjusted performance.

Second, although most AI literature focuses on identifying market direction, traders in financial mar-

kets, on a daily basis, are repetitively presented with a sequence of decisions, with market direction be-

ing only one piece of the puzzle. This presents a very limited view on the applicability of AI in trading

scenarios. Moreover, many studies in the existing literature do not reflect the rigid constraints that typ-

ically govern the trading desk—see also Vanstone and Tan (2003). In particular, Pardo (2008) states that

position sizing is often not appreciated and poorly understood in trading strategy design.

Our goal in this paper is to address these two, albeit interrelated, literature gaps with the ultimate ob-

jective to enhance the risk-adjusted performance of trading algorithms in an intraday stock trading sce-

nario. In particular, we present an innovative method that, in addition to trend direction signals,

identifies the optimum capital allocation across an intraday trading period by dynamically adapting

the levels of trade frequency and position sizes (two common decisions taken by traders) based on dif-

ferent degrees of expected return and volatility (uncertainty) over short intraday horizons. This also

sheds more light on the theoretical claims of Holmberg et al. (2013) that market inefficiency, and hence

the profitability of technical rules, can be linked to different volatility periods.

Finding the optimum level of trade frequency and position size along continuously changing intra-

day market conditions can be a nontrivial task. Albeit volatility is typically linked with risk, suffi-

cient market volatility is required to ensure that changes in prices exceed transaction costs.

Several studies suggest a fixed return threshold filter to avoid small unprofitable movements (Kaastra

& Boyd, 1996; Vanstone & Finnie, 2009, 2010). Holmberg et al. (2013) show that increasing the

return filter size results in a better success rate and average return. However, the same authors show

that this leads to a reduction in the number of trades, hence reducing the investors’potential profits.

Other authors (e.g. Brabazon & O’Neill, 2006) prefer to avoid high volatility (uncertainty)

completely by staying out of the market during periods whenvolatility goes beyond a specified fixed

threshold, hence avoiding the risk of possibly strong adverse market movements. This, however,

reduces the opportunity of possibly extraordinary gains. Also, Pardo (2008) claims that the effective

rate at which trading equity is compounded will remain suboptimal if a sound position sizing

strategy or algorithm is not employed.

These arguments cast the trade frequency and position size decisions in the context of better man-

agement of uncertainty. The modelling of uncertainty is at the core of fuzzyset theor y (Zadeh,1997).

Hence, as an extension to a neural network (NN) trend prediction model that is very popular in trad-

ing applications (Tsai & Wang, 2009; Krollner et al., 2010; Choudhry, McGroarty, Peng, & Wang,

2012; Vella & Ng, 2014), this paper presents a hybrid model consisting of an NN that is then ex-

tended with a dynamic fuzzy logic money management controller (NN-FMM). Core to our controller

is a fuzzy c-means (FCM) clustering algorithm (Bezdek, 1981; Dutta Baruah & Angelov, 2010) that

identifies unique trading performance regions across two dimensions: (a) the intraday realized vola-

tility (Andersen & Bollerslev, 1997), which is used as a proxy for uncertainty, and (b) the predicted

return size, which indicates a trend direction. The identified fuzzy clusters allow the extraction of

fuzzy rules, and their combined result produces a decision surface across the trend and volatility

space that is used to a dapt trade frequency and posit ion sizing levels based on local (r ather than

global) regional performance.

154 V. VELLA AND W. L. NG

Copyright © 2014 John Wiley & Sons, Ltd. Intell. Sys. Acc. Fin. Mgmt., 22, 153–178 (2015)

DOI: 10.1002/isaf