-The article provides a comparative analysis of existing models for forecasting prices in the foreign exchange market. The authors propose an improved modification of the model that is most suitable for predicting the behaviour of the EUR / USD currency pair. The calculation is performed by the developed software tool that processes and analyses the entered parameters of the time series.
INTRODUCTION
Currently, there is great interest in the stock market, and in particular in the foreign exchange market.
Many researchers create algorithms and methods for its prediction [1-4]. Therefore, it is of great interest to analyze existing approaches, as well as develop a model that has a number of advantages over existing analogs.
Given the increased volume of data necessary for analysis, new, often nonclassical approaches are required. These include problem solving using artificial intelligence. Currently, it is used in various fields of activity, including the financial sector. The authors set the task of finding the model most suitable for predicting the behavior of the EUR / USD currency pair, as well as its improvement.
Many authors conducted research on this topic and they got different results. For example, an interesting approach is described in the article on fractal time series analysis [1].
A number of articles describe works on expert shortterm forecasting of the foreign exchange market [2-4]. The paper [3] presents research on the fundamental analysis of world currency
markets. The article [5] describes the campaigns used in predicting changes in time series parameters.
II.
Consider the currently most popular models for forecasting prices in the foreign exchange market. The model of autoregression is moving average (ARMA) is one of the mathematical models used to analyze the forecasting of stationary time series in statistics. The ARMA model generalizes two simpler time series models: the autoregressive (AR) model and the moving average (MA) model [4]. As a rule, ARMA models, although they have a more complex structure, in comparison with AR and MA models, are characterized by fewer parameters. ARMA models also have a
number of other properties that determine their practical attractiveness [6].
The predicted value is determined as the following linear = + + −1 + , Copyright © 2020 for this paper by its authors.
is coefficient before , is “White noise” is residues that do not correlate with the remnants of the previous period, is constant.
Another well-known model is the autoregressive model of conditional heteroskedasticity - (G) ARCH. The meaning of this model is encrypted in its name. So, the ARCH model uses past values of the series for forecasting (autoregression), which in turn is heterogeneous, which is manifested in the variability of the variance of random error (heteroskedasticity) [7] [8]. This model was proposed by Robert Engle in 1982 [9], it can be represented as the following equation (2):
2 = 0 + 1 2−1 + ⋯ + 2−1, where is volatility function, m is base volatility, 2−1 is squares of past asset returns, is model coefficients showing the effect of past asset returns on the current value of volatility.
The (G) ARCH model has a number of disadvantages,
for example, the need to choose a large order of the model
so that the results are better [
Let's try to compare the effectiveness of the ARMA and
G (ARCH) models[
Fig.2. The
graph of the behavior of a currency pair over a time interval with the highlighting of the end date of the study.
At the beginning of the study period on November 19, 2019, the currency pair was at around 1.08301. At the end of this period, 1.2138 (Figure 2).
For this study period, a forecast was made using both ARMA and G (ARCH) models. Table 1 shows the results of these models.
According to the results of the models shown in table we can see that the G (ARCH) model for this currency pair turned out to be more accurate. In this model, the standard error has decreased, and there is also a more accurate total amount.
The authors implemented modifications G (ARCH) model, which showed the best result to the III.
MODIFICATION AND THE RESULTS
The model was modified on the basis of an empirical approach related to the fact that the value of a currency depends on many factors (events of a different nature taking place in the world) that the expert can somehow evaluate.
Such an assessment can be made at each step of the model, making certain adjustments to it (3). These changes can be made, for example, daily, so that the next day a more accurate result is obtained. The result of such a model will depend not only on mathematical calculations, but also on the experience of the person who implements it.
2 = 0 + ( 1 + ₁) 2−1 + ⋯ + ( + ) 2−1 (3) where is a parameter whose value the expert sets depending on events that, in his opinion, could affect the value of the currency.
Applying the modification of the G model (ARCH), the accuracy of the calculations increased by almost 4%. For the foreign exchange market, this is a good indicator.
A modified model with adjustments 2 times a day gives an even more accurate result (Table 3).
In this case, the model prediction accuracy increased by 7% relative to the base model G (ARCH).
An obvious factor is a significant increase in the complexity of this model. Optimization of the modified model is possible due to the replacement of the expert’s work with elements of artificial intelligence. As an artificial intelligence, we are developing a neural network that will select this CF, instead of a human. A neural network will be able to more objectively consider factors and the same sample will be an order of magnitude higher.
Using a neural network, we plan to increase the accuracy of predictions to at least 10% relative to the base method.
IV.
The paper presents popular time series forecasting models. The authors developed an improved modification of the (G) ARCH model, which showed the best result for predicting the EUR / USD currency pair. To assess the performance of the models, a comparative analysis of the forecast data with the real one was carried out. The results showed some advantage in the accuracy of forecasting an improved model over existing analogues. This is very important, because this result allows us to evaluate the benefits of using additional parameters in classical models. Additional parameters are expert, but potentially they can be obtained analytically using various tools, for example, neural networks. The authors plan further work to improve models for forecasting the behavior of the foreign exchange market. As a result, we can say that our neural network will predict the currency price, otherwise than usual, since we will not train it on simple mathematical equations, but also take into account the opinions of a person who is versed in this area and will be able to adjust the model based on his experience, and not just numbers, we believe this is the uniqueness of this development.We use not just mathematics, but also human experience. Which can help find those factors that the neural network cannot see.