By Bogdan Zgersky, Portfolio Manager at GL Asset Management AG
The fundamentals of an investment management are based on forecasting future financial indicators, but predicting such data is not an easy task.
The difficulty of accurately forecasting is rooted in the same ground as to why money does not grow on trees – there are too many people ready to rip them off. To win the competition in the long term, it is necessary to constantly advance and develop the methods and models of market senarii. As a result, the one with the best experience and methods for detecting regularities obtains the highest rate of return out of their less-competitive colleagues.
Over the last decades, we have observed a steady growth in the popularity of Technical Indicator Analysis1 – a set of empirical rules based on various indicators and statistic linear dependencies. Linear modelling was the main principle of operation in most fields due to its well-developed procedures for precise optimisation. However, for tasks with poor approximations (such as forecasting the future), linear models are of little use, which is why the growth of computational capacities is coupled with increased interest in neural networks2. Due to their nonlinear nature, they have become an exceptionally powerful modelling method, able to reproduce particularly complex dependencies, even though the data for such analysis require a different kind of structuring.
In order for Neural Networks to produce top results, it is essential to accurately understand the structure and morphology of the input data for analysis in the first place, to get the architecture of the Neural Network right for the data structure.
Implementation of such a simple Neural Network as a multilayer perception will not give noticeably accurate results in forecasting price time series. For instance, upon keying in a number of prices the network will look for dependencies in the data as if prices did not depend on each other, or as if current prices were independent of prices over the previous periods or possible price extremes.
It all looks as if each price was viewed as a separate parameter in such a set of mutually independent parameters as eye colour, height, age, and sex in search for a shared dependency, influencing the future price.
However, if we assume that a dependency exists between the prices of a given time series, we must use the right architecture that allows operating such dependencies, for instance, Convolutional (CNN) and Recurrent Neural Networks (RCC).
If our study aims at identifying patterns to forecast prices, we can use Convolutional Neural Networks, used for image recognition. In so doing we assume that a chart consists of a set of patterns we still have to identify and forecast price behaviour according to a particular pattern. However, we will not be able to consider the impact patterns have on each other. To that end, it is indispensable to analyse the identified patterns using another network and another architecture. For instance, we can use the Recurrent Neural Network.
Recurrent Neural Networks (RNN) are commonly used to analyse such data as text or audio files, wherein data structure implies dependence of data entries from each other’s values–for example, the meaning of a sentence can radically change depending on the context, word order, and the positions of a comma. Consequently, this Neural Network analyses and memorizes such dependencies and learns to identify regularities among them, that is why it is safe to assume such networks are best to use for price charts analysis.
One has to bear in mind that there is no universal approach for Neural Network data analysis. In order to build the right architecture capable of completing tasks at hand, it is crucial to be fully aware of the data morphology and structure and to understand how the data is organised and how it can be interrelated. If the data consists of physical parameters or image data, where the rules of data interrelation, its structure and morphology are most obvious; and if we use the data of such complex systems as text or speech, where the interrelations are variable, we have to build Neural Network clutches with specialised architecture in line with the most accurate description of the nature of the data analysed.
GL Asset Management AG, Portfolio Manager
1 – To know more about Technical Analysis (Investopedia)
2 – “A computer system modelled on the human brain and nervous system.” Wikipedia: Neural Network