Uzun kısa süreli bellek ile altın fiyatı tahmini

Abstract

Thanks to its many physical and chemical properties, gold has been a mineral that has attracted the attention of people since ancient times. Although it has various usage areas such as the defense industry, electrical electronics industry, and jewelry, its economic value can be given as the most important feature from the first ages to the present day. With the use of gold in coins by the Lydians, gold became a tool of exchange and investment. It still maintains this feature today. Thanks to these features, it has become a symbol of power in societies. With the Bretton-Woods system, which came into effect towards the end of World War II, the gold price was indexed to the US dollar. With the increasing tension in the world markets, this system collapsed, and gold started to be priced dynamically as of 1971. The gold price in Turkey is calculated using the worldwide accepted gold price in US dollars and the Turkish lira / US dollar parity. Today, investors want to make a profit in the long or short term with minimum risk factors. The most risk-free investment instruments preferred by investors can be given as precious commodities such as gold, foreign currencies, stocks, real estate, cryptocurrencies, bonds, mutual funds, and government bonds. It is preferable to make estimates of the corresponding returns or losses on such investments. For this reason, forecasting is of great importance when investing and the reaction of investment instruments affected by events in the world should be analyzed. Our main purpose in this research is to predict the future behavior of gold based on past data. The recurrent neural network (RNN), which is a type of deep learning, was chosen as the forecasting method. The problem to be studied in this research is considered a regression problem that requires a nonlinear solution to the time series. In this study, feature selection and gold price prediction in multivariate financial time series were examined. There are many dynamic factors for pricing gold used for investment. Since not every factor has the same effect, the most effective factors on gold should be determined. After the data engineering part, the main factors affecting gold price have been identified. Features were determined using the factors examined, and the most important of these features were selected and formed the basis of the study. The datasets were created using various features such as the US dollar index, cryptocurrencies, commodities, stock market indices, volatility index, inflation, and interest rates. Although gold prices in the real-world act according to basic theory and criteria, gold is a commodity type that is affected by many technical and fundamental parameters. Before the forecasting section, feature selection was made using Random Forest Regression and Linear Regression. In this section, it has been determined that the parameters that affect the gold price the most are the USD/JPY parity, 10-year expected inflation (USA), 10-year real interest rate (USA), US gross national product (GDP), and the amount of USD in circulation. No improvement was observed in the forecasting performance criterion even if more variables were added. In the principal component analysis, the most important variables representing the main dataset were determined as oil, US real interest rate, Bitcoin, silver price, LME index, 10-year inflation (USA), TXBM index, USD money supply M1, and volatility index. A basic recurrent artificial neural network (RNN) and long short-term memory deep learning network (LSTM) were used in the forecasting study. The dataset combinations were created by using 5 variant variables on 3 main datasets so that there are 15 combinations in total. These variant variables are economic indicators LMACD and MACD. Test combinations were created using dataset combinations also batch size and window size values determined for RNN and LSTM networks. It has been tried to give an idea about the reaction of the gold price against these inputs. The window size is the hyperparameter that determines how many days the historical data will be retrieved when creating the observation unit. RNN and LSTM hyperparameters were also derived on each dataset combination and forecasting was made. After the training process, the predictive model performances of the applications were calculated. Parameters with appropriate estimation results were determined in the tests performed. While RNN performs at par with LSTM in one main dataset, LSTM has a higher predictive success than RNN in the other two main datasets. In tests where the datasets created by adding the MACD indicator were trained with lower window sizes, these models gave superior results than other combinations. In addition, it was observed that there were deviation errors in the training of the models due to the Covid-19 crisis, which started in March 2020. In the forecasting study on the training set, it was determined that the network could not perform as well as before this date in the part of the training data after the onset of Covid-19. To improve the model, the existing parameters were selected more precisely, and optimization was made. Although it is not possible to use it professionally yet, it has given promising results for the first study. In this context, the aim and scope of the study have been met. It will also be a starting point for future work.