In order to forecast stock prices based on economic indicators, many studies have been conducted using well-known statistical methods. Meanwhile, since ~2010 as the power of computers improved, new methods of machine learning began to be used. It would be interesting to know how those algorithms using a variety of mathematical and statistical methods, are able to predict the stock market. The purpose of this article is to model the monthly price of the S&P 500 index based on U.S. economic indicators using statistical, machine learning, deep learning approaches and finally compare metrics of those models. After the selection of indicators according to the data visualization, multicollinearity tests, statistical significance tests, 3 out of 27 indicators remained. The main finding of the research is that the authors improved the baseline statistical linear regression model by 19 percent using a ML Random Forest algorithm. In this way, model achieved accuracy 97.68% of prediction S&P 500 index.

There are a big number of researches which analyzing stock price returns. Some of them is based on fundamental analysis theory. Meanwhile other studies are based on efficient market hypotheses and financial behavior theories. However, there is not enough researches combining the characteristics of these theories into one. Such kind of researches in the scientific literature is usually referred to macroeconomic news, announcements, surprises, expectations studies. These studies examine not only the actual but also the predictive values of macroeconomic indicators announcements, normalizing them and thus creating absolutely a new surprise indicator. Purpose of this paper is modeling EURO STOXX 50 index price returns based on Industrial production surprise indicator. Empirical part shows that the best models for explaining EURO STOXX 50 index price returns was obtained at the 40 and 42 in different surprise indicator scenario. The coefficient of determination was obtained respectively 24.70% and 21,80%. Meanwhile applying machine learning method of artificial intelligence, a much more accurate models were obtained. The coefficient of determination respectively was 33,22% and 26,60%.

Abstract. This article evaluates fluctuations in the Standard