| Abstract [eng] |
The efficient market hypothesis states that price of securities fully reflects, almost in real-time, all available information, including historical price data. Hence, the stock price immediately responds to new information. Based on this hypothesis, the optimal future stock price prediction is its current price. However, in practice, market dynamics are observed, which allow modeling part of the future price change using historical data. In order to correctly identify and consequently model market dynamics, the study of S&P 500 stock market index is carried. This process is modeled during the study in order to forecast its changes and to evaluate the possibilities of investing in the stock market. Estimations of daily averaged value of the index with horizons 1 to 4 are evaluated. Data range from 2017 to 2019 has been selected for research. Data sets used in the work are downloaded from Yahoo finance and FRED database. The selected time series are modeled using 4 different models: VAR, ARMA+GARCH, SARIMA, and LSTM. In process of this study, observation that distinct models at different time intervals were able to make smaller biases in predicting 1 to 4 day ahead index values. This structure was modeled in order to improve the accuracy of the forecasts. To achieve this, 3 different architectures of model ensembles were created. Market simulations have also been carried, and the use of different model predictions has been assessed by proposed strategy. |
