Wing Lon Ng

Intraday high-frequency FX trading with adaptive neuro-fuzzy inference systems

This paper introduces an adaptive neuro-fuzzy inference system (ANFIS) for financial trading, which learns to predict price movements from training data consisting of intraday tick data sampled at high frequency. The empirical data used in our investigation are five-minute mid-price time series from FX markets. The ANFIS optimisation involves back-testing as well as varying the number of epochs, and is combined with a new method of capturing volatility using an event-driven approach that takes into consideration directional changes within pre-specified thresholds. The results show that the proposed model outperforms standard strategies such as buy-and-hold or linear forecasting.

Enhancing risk-adjusted performance of stock market intraday trading with Neuro-Fuzzy systems

Whilst the interest of many former studies on the application of AI in finance is solely on predicting market movements, trading practitioners are predominantly concerned about risk-adjusted performance. This paper provides new insights into improving the time-varying risk-adjusted performance of trading systems controlled by Artificial Neural Networks (ANNs), Adaptive Neuro-Fuzzy Systems (ANFIS) or Dynamic Evolving Neuro Fuzzy Systems (DENFIS). Contrary to most former studies which focus on daily predictions, we compare these models in an intraday stock trading scenario using high-frequency data. Firstly, we propose a dynamic extension of the popular moving average rule and enhance it with a model validation methodology using heat maps to analyse favourable profitability in specific holding time and signal regions. Secondly, we study the effect of realistic constraints such as transaction costs and intraday trading hours, which many existing approaches in the literature ignore. Thirdly, unlike most former studies that only aim to minimise statistical error measures, we compare this approach with financially more relevant risk-adjusted objective functions. To this end, we also consider an innovative ANFIS ensemble architecture which on an intraday level dynamically selects between different risk-adjusted models. Our study shows that accounting for transaction costs and the use of risk-return objective functions provide better results in out-of-sample tests. Overall, the ANN model is identified as a viable model, however ANFIS shows more stable time-varying performance across multiple market regimes. Moreover, we find that combining multiple risk-adjusted objective functions using an ANFIS ensemble yields promising results.

Improving risk-adjusted performance in high frequency trading using interval type-2 fuzzy logic

We investigate the viability of Type-2 fuzzy systems in high frequency trading.We propose Type-2 models based on a generalisation of the popular ANFIS model.Type-2 models score significant risk adjusted performance improvements over Type-1.Benefits of Type-2 models increase with higher trading frequencies. In this paper, we investigate the ability of higher order fuzzy systems to handle increased uncertainty, mostly induced by the market microstructure noise inherent in a high frequency trading (HFT) scenario. Whilst many former studies comparing type-1 and type-2 Fuzzy Logic Systems (FLSs) focus on error reduction or market direction accuracy, our interest is predominantly risk-adjusted performance and more in line with both trading practitioners and upcoming regulatory regimes. We propose an innovative approach to design an interval type-2 model which is based on a generalisation of the popular type-1 ANFIS model. The significance of this work stems from the contributions as a result of introducing type-2 fuzzy sets in intelligent trading algorithms, with the objective to improve the risk-adjusted performance with minimal increase in the design and computational complexity. Overall, the proposed ANFIS/T2 model scores significant performance improvements when compared to both standard ANFIS and Buy-and-Hold methods. As a further step, we identify a relationship between the increased trading performance benefits of the proposed type-2 model and higher levels of microstructure noise. The results resolve a desirable need for practitioners, researchers and regulators in the design of expert and intelligent systems for better management of risk in the field of HFT.

Nonparametric estimation of general multivariate tail dependence and applications to financial time series

In order to analyse the entire tail dependence structure among random variables in a multidimensional setting, we present and study several nonparametric estimators of general tail dependence functions. These estimators measure tail dependence in different orthants, complementing the commonly studied positive (lower and upper) tail dependence. This approach is in line with the parametric analysis of general tail dependence. Under this unifying approach the different dependencies are analysed using the associated copulas. We generalise estimators of the lower and upper tail dependence coefficient to the general multivariate tail dependence function and study their statistical properties. Tail dependence measures come as a response to the incapability of the correlation coefficient as an extreme dependence measure. We run a Monte Carlo simulation study to assess the performance of the nonparametric estimators. We also employ selected estimators in two empirical applications to detect and measure the general multivariate non-positive tail dependence in financial data, which popular parametric copula models commonly applied in the financial literature fail to capture.

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

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 prioritizes 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 network NN trend prediction model, our results show significant performance improvements compared with both standard NN and buy-and-hold approaches. Copyright

Revealing Intraday Market Efficiency -- Estimating Diurnal Price Densities in Limit Order Books

Liquidity plays an important role in trading and represents a nontrivial economic concept that is difficult to measure as it involves several three dimensions to investigate. In this paper, we explore the liquidity in electronic markets by estimating a time-varying Gamma distribution of volume adjusted prices for both bid and ask side of in the order book. Applying this framework on London Stock Exchange SETS order books during the continuous trading hours, we found a seasonal behaviour for the distribution parameters, implying a certain periodic intraday pattern of the markets liquidity and, hence, its efficiency.

Simulation based calibration using extended balanced augmented empirical likelihood

This paper introduces an extension of the balanced augmented empirical likelihood (eBAEL) method for calibrating simulation models. We illustrate the efficiency of our method in two simulation studies, where we calibrate moments of different distributions and parameters of a geometric Brownian motion process, comparing our approach against other simulation based methods. In these benchmark experiments we observe converging mean squared errors of the empirical likelihood approach. In fact, the results demonstrate that the eBAEL approach is able to provide the best mean squared errors for calibration and in particular is the most robust calibration method, particularly in the presence of noise.

Nonparametric Tail Copula Estimation: An Application to Stock and Volatility Index Returns

In this study, we measure asymmetric negative tail dependence and discuss their statistical properties. In a simulation study, we show the reliability of nonparametric estimators of tail copula to measure not only the common positive lower and upper tail dependence, but also the negative “lower–upper” and “upper–lower” tail dependence. The use of this new framework is illustrated in an application to financial data. We detect the existence of asymmetric negative tail dependence between stock and volatility indices. Many common parametric copula models used in finance fail to capture this characteristic.

Nonlinearities in stochastic clocks: trades and volume as subordinators of electronic markets

This paper discusses the possibility of recovering normality of asset returns through a stochastic time change, where the appropriate economic time is determined through a simple parametric function of the cumulative number of trades and/or the cumulative volume. The existing literature argues that the re-centred cumulative number of trades could be used as the appropriate stochastic clock of the market under which asset returns are virtually Gaussian. Using tick-data for FTSE-100 futures, we show that normality is not always recovered by conditioning on the re-centred number of trades. However, it can be shown that simply extending the approach to a nonlinear function can provide a better stochastic clock of the market.

Precise time-matching in chimpanzee allogrooming does not occur after a short delay

Allogrooming is a key aspect of chimpanzee sociality and many studies have investigated the role of reciprocity in a biological market. One theoretical form of reciprocity is time-matching, where payback consists of an equal duration of effort (e.g. twenty seconds of grooming repaid with twenty seconds of grooming). Here, we report a study of allogrooming in a group of twenty-six captive chimpanzees (Chester Zoo, UK), based on more than 150 hours of data. For analysis, we introduce a methodological innovation called the “Delta scale”, which unidimensionally measures the accuracy of time-matching according to the extent of delay after the cessation of grooming. Delta is positive when reciprocation occurs after any non-zero delay (e.g. A grooms B and then B grooms A after a five second break) and it is negative when reciprocation begins whilst the original grooming has not yet ceased. Using a generalized linear mixed-method, we found evidence for time-matched reciprocation. However, this was true only for immediate reciprocation (Delta less than zero). If there was a temporal break in grooming between two members of a dyad, then there was no evidence that chimpanzees were using new bouts to retroactively correct for time-matching imbalances from previous bouts. Our results have implications for some of the cognitive constraints that differentiate real-life reciprocation from abstract theoretical models. Furthermore, we suggest that some apparent patterns of time-matched reciprocity may arise merely due to the law of large numbers, and we introduce a statistical test which takes this into account when aggregating grooming durations over a window of time.