There has been a lot of research on stock price prediction in the stock market, and experiments using various methodologies are still ongoing. In particular, with the advent of artificial intelligence, various techniques are being explored for stock price prediction using machine learning methodologies to predict stock market movements (7). Related studies that have recently been developed in this field are as follows, and a brief overview is shown in Table 1. Recognizing the difficulties of forecasting stock prices, tree bagging, random forests, and logit models have been developed under the premise that they are more successful at predicting stock price directions. As a result of performing a performance comparison, it was suggested that tree bagging and random forest are more useful methods for predicting stock prices (8). It created a model to predict the prices of 10 high-yielding stocks in the CSI300 index, and random forests outperformed ANN and logistic regression (9). Sharma and Juneja (10) used LSBoost and showed that a model using technical indicators outperformed support vector regression (SVR). The proposed model achieved better prediction performance than did SVR (10). Bucci (11) developed a model to predict the volatility of the S&P 500 index. Capturing the long-term dependence of volatility through a nonlinear autoregressive model process using LSTM (long short-term memory) and NARX (exogenous input) neural network models was shown to improve forecasting accuracy even during periods of high volatility (11). Guo et al. (12) presented that a hybrid model combining the LSTM model and the LightGBM model showed high performance.

In addition, studies using data processed to suit the characteristics of time series data were also conducted. First, Yong et al. (13) created a deep neural network (DNN) model to predict the FTSE Strait Time Index t days in the future and proposed a trading system with a 70% profitability (13). Montenegro and Molina (14) presented a sliding window technique based on S&P 500 index data. This has contributed to providing scholars with experience with new approaches and supporting investment decisions in the stock market (14). Wu et al. (15) combined the k-means and AprioriAll algorithms to create a model and showed that it can be used for a long period of time to make profits in real markets (15). Wen et al. (16) presented results showing that models using augmentation methods for time series classification models had low accuracy. However, it was confirmed that the performance slightly improved when the scaling method was applied to the DNN algorithm, and in this study, we developed a model by applying the scaling method to machine learning and deep learning algorithms.

Like the previous studies discussed above, stock prediction research uses a variety of methodologies to test predictability. These studies used ML algorithms and technical indicators. However, the stock price can be affected by macroeconomic factors in addition to information about the company that issued the stock, transaction price, and volume. Kim and Kim (17) analyzed the correlation between macroeconomic variables and stock prices. Granger causality analysis results showed that many variables were correlated. In this study, various technical and economic indicators are used, and stock price prediction is performed using various and latest artificial intelligence algorithms.

The selection of major indices used the indices of major countries used in previous studies, and included exchange rate information and digital currency between major countries. Futures included major raw materials and food. For analysis, we extract 12 years of stock price information data from finance.yahoo.com. Two of those years are used to generate technical variables and are not used for training or testing. The types of variables used after the null processing process are stock indices such as KOSPI, KOSDAQ, FTSE, S&P500, DowJones, NASDAQ, Nikkei, and HSI. Exchange rate information is Korea, UK, EUR, Japan, China, and Bitcoin. Bond is 10 years, 2 years. Future is Gold, Copper, Crude Oil, Natural Gas Corn, Rough Rice, Soybean Meal, and Coffee. The technical variables are used by extracting data from Yahoo Financial. That is simple moving average (SMA), weighted moving average (WMA), Relative Strength Index (RSI), Stochastic K% D%, Moving Average Convergence Divergence (MACD).

The model training and optimization process was performed as follows. First, by dividing the training set, various parameters were tested using the 5-fold cross-validation technique. This is a method of calculating the performance of a test set by selecting the best parameter according to the criteria and learning the best parameter by combining the five divided training sets. This method was used for model training in the random forest and LightGBM models. Second, early stopping, a method mainly used to prevent overfitting of the DNN model, is a method for stopping learning if there is no improvement in the performance of the validation set through loss calculation while learning the training set. In this study, the training set and the validation set were classified as 8:2, and when the loss reduction in the validation set was less than 0.00001 while learning the training set, early stopping was used to stop the learning.

The following steps are taken for prediction: (1) Extract 12 years of data. (2) The first 2 years are used only for technical indicator generation. (3) 8 years of data out of 10 years are used for training/validation. (4) Provide 2 years of test results.

For the test set results, see Tables 2–6. The data collection period was 12 years before the reference date (FTAS 2022/11/03, GSPC 2022/10/31, HSI 2022/11/04, KS11 2022/10/28, and N225 2022/11/02), and 10 years were used for training/validation/testing. That is, the remaining two years are used for feature calculation of the moving average. The training/validation set and test set division should measure time intervals that are not included in the training/validation set; 8 years should be used for training/validation, and 2 years should be used for testing. That is, data from 2020 to 2022 were used (KOSPI test set: 2020/10/29 to 2022/10/28; training set: 2012/10/29 to 2020/10/28). Therefore, for training/validation, various parameters were used to improve the performance because the pattern of the stock market changes greatly due to a sharp rise in 2020 and a rise in late 2021 and 2022 after a decline in 2019.

See summary (Table 7). We applied several models to predict the composite stock indices of major countries and selected the optimal model. The optimal model learns according to the optimal parameter selection method presented above. For each type of feature, one of nine features is selected as the DNN, RF, or LGBM for each type of basic, technological, overall index, or algorithm. The optimal model has 15 technical indicators and 5 basic/total indicators for each type of feature; in the algorithm, 16 LGBMs, 5 DNNs, and 4 RFs are used. Therefore, the features that should be applied first for stock price prediction include basic indicators and technical indicators (SMA and WMA of 5, 10, 20, 60, and 120 days, RSI, Stochastic K%, Stochastic D%, MACD, signal line, and histogram, respectively) combined. The reason that the performance of the technical indicator data, including basic indicators, is greater than that of the basic indicators when using the fewest indicators and that of the overall indicators (including basic, technical, and macroeconomic indicators) when using the most data is that indices with a high correlation (over 0.995) of T+1 to 5, such as the KOSPI and S&P, are included in the foundation; thus, it was concluded that the inclusion of technical indicators would improve the optimal performance. In addition, an algorithm that should be applied prior to RF or DNN for stock price prediction is proposed as a LightGBM model. The structure is considered to minimize overfitting and maximize performance due to the characteristics of the algorithm.

In previous studies, exchange rates, raw materials, and bonds are mentioned as indicators that affect stock price prediction, and some studies have shown them to be important factors, but the corresponding indicator was effective in Nikkei and Hang Seng. In other words, the results of this study provide important data for reference in the use of various algorithms and indicators depending on the indicator.