In recent years, artificial intelligence (AI) has attracted growing attention as a promising approach to time-series forecasting, with an increasing number of studies exploring and discussing its applications across various domains (e.g., [2]; [13]; [12]). Stock price prediction is one of the important research topics in financial markets, and a wide range of forecasting methods have been examined, from statistical approaches to machine learning and, more recently, deep learning. Because stock prices and stock returns exhibit complex nonlinearities and temporal dependencies, recent studies have employed machine learning models such as Random Forest, Decision Tree, Extreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine (LightGBM), as well as deep learning models such as Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN), with the aim of capturing these characteristics (e.g., [5]; [10]).

Previous studies have examined the effectiveness of deep learning models, including LSTM, as well as the possibility of utilizing information other than prices, such as investor sentiment on social media (e.g., [6]). In addition, research has progressed on addressing the non-stationarity of financial time series and on the generalization performance of models across different markets (e.g., [4]).

On the other hand, when using time-series information from stock prices themselves, there remains room for further investigation into the extent to which different machine learning models exhibit different predictive performance and how predictive performance changes depending on the sample period and the specification of explanatory variables.

This study therefore applies five machine learning models—Random Forest Regression, Decision Tree Regression, XGBoost Regression, LightGBM Regression, and K-Nearest Neighbors Regression—to daily Nikkei 225 data to forecast the closing price on the forecast day and compare their predictive performance. As shown in Figure 1, the Nikkei 225 has exhibited a pronounced upward trend in recent years, making it a useful benchmark for evaluating and comparing different forecasting models. In addition, this study uses a model that includes only lags of the closing price as explanatory variables, which serves as the baseline model, and compares its predictive performance with that of models that additionally include lags of the opening, high, and low prices.

Furthermore, six different forecast target periods, ranging from five to ten years, are considered to examine the effects of differences in the forecast target period on the prediction results. The objective of this study is to empirically clarify, through these comparisons, the relative predictive performance of the respective machine learning models in forecasting the closing price of the Nikkei 225 and the differences in predictive performance resulting from the specification of explanatory variables and forecast target periods.

The results of the analysis show broadly similar tendencies across different forecast target periods for each of the three evaluation measures, the R-squared value (R2), mean absolute error (MAE), and root mean squared error (RMSE). In particular, Random Forest Regression and Decision Tree Regression exhibit relatively high predictive performance, followed by K-Nearest Neighbors Regression.

In addition, adding lags of the opening, high, and low prices to the lags of the closing price as explanatory variables is found to slightly improve predictive performance in some cases. Furthermore, comparison across different forecast target periods indicates that predictive performance tends to improve as the period is extended from five to ten years. These findings are significant contributions of this study.

The remainder of this paper is organized as follows. Section 2 reviews the previous literature, and Section 3 describes the data and models. Section 4 presents the empirical results, followed by the conclusion in Section 5.

Figure 1. Evolution of the Nikkei 225 closing prices: January 5, 1970, to December 30, 2025.

In recent years, research on stock price prediction has progressed through the use of machine learning and deep learning in addition to conventional statistical methods. An increasing number of studies have employed neural networks such as LSTM and CNN to capture the complex nonlinearities and time-series characteristics of stock prices and stock returns. This section reviews previous studies on stock price prediction using AI in chronological order.

[9] applied LSTM to stock index prediction and demonstrated the usefulness of Recurrent Neural Networks (RNNs) and LSTM for time-series data such as stock prices. This study is important in that it indicated the potential effectiveness of LSTM, which can learn long-term dependencies, for predicting future price movements using historical stock price information.

Subsequently, [5] presented an approach for classifying stock price increases and decreases by combining large-scale data processing using Spark with machine learning models such as Random Forest and Decision Tree. This demonstrated the potential for applying not only deep learning but also conventional machine learning models to large-scale financial data.

Moreover, [7] examined stock time-series forecasting using LSTM and found that the number of training iterations and the period of data used affected predictive performance. These studies indicate that, in stock price prediction, not only the choice of models that appropriately capture the characteristics of time-series data but also training conditions and data handling have important effects on predictive accuracy.

Later, [8] examined the potential of utilizing AI and machine learning in stock market analysis by incorporating sentiment analysis of posts on Twitter in addition to technical analysis based on historical stock price data. In this way, the information used for stock price prediction has expanded from market data such as prices and trading volume to unstructured data containing investors’ opinions and emotions.

Furthermore, [6] applied Natural Language Processing (NLP) to a large volume of social media data and quantified investor sentiment to analyze the relationship between sentiment and stock prices. Their results showed a significant positive synergy between investor sentiment and stock prices, while also indicating that this relationship could reverse or disappear depending on the period. In addition, the study showed that investor sentiment on social media reflects expectations regarding future market trends, suggesting the potential usefulness of sentiment information for stock price prediction.

Moreover, [4] focused on the non-stationarity of financial time series and proposed a forecasting model for non-stationary financial time series using meta-learning. In this model, CNN is used as the predictor and LSTM as the meta-learner, and long-term non-stationary time series are divided into multiple shorter subsequences for forecasting. The experiments showed that the proposed model achieved better predictive performance than conventional CNN and autoregressive (AR) models, demonstrating the effectiveness of an adaptive forecasting approach for non-stationary financial data.

Subsequently, [3] compared multiple neural network architectures, including Multilayer Perceptron (MLP), RNN, LSTM, and CNN, examined their respective stock price forecasting performance, and analyzed the applicability of forecasting models to different markets. This indicates that, in addition to accuracy for a particular market or dataset, the generalization performance of models is also an important research issue. Furthermore, [1] applied CNN, LSTM, and hybrid models combining these approaches to financial markets in the BRICS countries, thereby extending the application of deep learning models to multiple countries and markets.

These studies indicate that research on stock price prediction has evolved from forecasting based on a single time series toward forecasting in non-stationary market environments and different markets, as well as forecasting using combinations of multiple deep learning models. However, relatively few studies have carefully examined well-established models by changing settings such as forecast target periods and explanatory variables. Therefore, we believe that research such as the present study fills the research gap left by existing studies and thus is highly meaningful.

This section describes the data used in the analysis and the forecasting models. This study employs the five machine learning models, Random Forest Regression, Decision Tree Regression, XGBoost Regression, LightGBM Regression, and K-Nearest Neighbors Regression, for model comparisons.

The analysis uses daily data for the Nikkei 225. Specifically, the four daily price variables—opening price (Open), high price (High), low price (Low), and closing price (Close)—are used. The overall sample period used in the analysis is from January 5, 1970, to December 30, 2025. To clarify the differences by forecasting periods, six forecast target periods are considered: five years from January 4, 2021, to December 30, 2025; six years from January 6, 2020, to December 30, 2025; seven years from January 4, 2019, to December 30, 2025; eight years from January 4, 2018, to December 30, 2025; nine years from January 4, 2017, to December 30, 2025; and ten years from January 4, 2016, to December 30, 2025. Each of these six periods is analyzed separately for the five machine learning models.

Accordingly, in model training and prediction, for each of the six period settings, the observations outside the respective forecast target period within the overall sample period constitute the training period, while the respective forecast target period constitutes the test period.

The prediction target is the closing price (Close) of the Nikkei 225. First, taking into account the temporal relationship with the closing price, a baseline model is specified using lags 1 through 3 of the closing price as explanatory variables. Moreover, three additional six-variable models are constructed by adding lags 0 through 2 of the opening price (Open), lags 1 through 3 of the high price (High), and lags 1 through 3 of the low price (Low), respectively, to the explanatory variables of the baseline model.

These four explanatory-variable specifications are then applied to the five forecasting models described above for comparative analysis. For lag 0 of Open, it is assumed that the opening price of the day is known at the time of prediction. In other words, the analysis assumes that the closing price of the day is predicted once the opening price of that day becomes known.

This section compares and examines the results obtained from each forecasting model. The five models—Random Forest Regression, Decision Tree Regression, XGBoost Regression, LightGBM Regression, and K-Nearest Neighbors Regression—are applied to the price series of the Nikkei 225 and compared across six forecast target periods of five, six, seven, eight, nine, and ten years. The results for the predictive accuracy of each period are presented in Tables 1-6, respectively. For all forecasting models, model specifications were determined based solely on data from the training period, and observations from the forecast target period were not used for model fitting.

Each of the six tables presents the results of forecasting the closing price on the forecast day using past 3-day closing prices as explanatory variables in Panel A; the results of forecasting using past 3-day closing prices and same-day and past 2-day opening prices as explanatory variables in Panel B; the results of forecasting using past 3-day closing prices and past 3-day high prices as explanatory variables in Panel C; and the results of forecasting using past 3-day closing prices and past 3-day low prices as explanatory variables in Panel D.

First, across the three evaluation measures—R2, MAE, and RMSE—shown in Tables 1-6, broadly similar tendencies are observed with respect to the predictive accuracy of the daily closing price of the Nikkei 225 across the six forecast target periods. In particular, Random Forest Regression and Decision Tree Regression exhibit relatively high predictive accuracy, followed by K-Nearest Neighbors Regression. In contrast, XGBoost Regression and LightGBM Regression underperform the other three models.

Furthermore, as shown in Tables 1-6, adding lags of Open, High, or Low as explanatory variables to the baseline models, which use only lags of Close as explanatory variables, slightly improves predictive accuracy in some cases. At the same time, predictive accuracy tends to improve as the forecast target period is extended from five to ten years. As the values of R2, MAE, and RMSE inTables 1-6 indicate, these tendencies remain highly consistent across the six forecast target periods.

Table 1. Results of 5-year forecasts of the Nikkei 225 daily closing prices.

Notes. Our full sample period is from January 5, 1970, to December 30, 2025, and the test period is from January 4, 2021, to December 30, 2025. R2 denotes the R-squared value, MAE denotes mean absolute error, and RMSE denotes root mean squared error.

Table 2. Results of 6-year forecasts of the Nikkei 225 daily closing prices.

Notes. Our full sample period is from January 5, 1970, to December 30, 2025, and the test period is from January 6, 2020, to December 30, 2025. R2 denotes the R-squared value, MAE denotes mean absolute error, and RMSE denotes root mean squared error.

Table 3. Results of 7-year forecasts of the Nikkei 225 daily closing prices.

Notes. Our full sample period is from January 5, 1970, to December 30, 2025, and the test period is from January 4, 2019, to December 30, 2025. R2 denotes the R-squared value, MAE denotes mean absolute error, and RMSE denotes root mean squared error.

Table 4. Results of 8-year forecasts of the Nikkei 225 daily closing prices.

Notes. Our full sample period is from January 5, 1970, to December 30, 2025, and the test period is from January 4, 2018, to December 30, 2025. R2 denotes the R-squared value, MAE denotes mean absolute error, and RMSE denotes root mean squared error.

Table 5. Results of 9-year forecasts of the Nikkei 225 daily closing prices.

Notes. Our full sample period is from January 5, 1970, to December 30, 2025, and the test period is from January 4, 2017, to December 30, 2025. R2 denotes the R-squared value, MAE denotes mean absolute error, and RMSE denotes root mean squared error.

Table 6. Results of 10-year forecasts of the Nikkei 225 daily closing prices.

Notes. Our full sample period is from January 5, 1970, to December 30, 2025, and the test period is from January 4, 2016, to December 30, 2025. R2 denotes the R-squared value, MAE denotes mean absolute error, and RMSE denotes root mean squared error.

This study used daily data for the Nikkei 225 to forecast its closing price using five machine learning models—Random Forest Regression, Decision Tree Regression, XGBoost Regression, LightGBM Regression, and K-Nearest Neighbors Regression—and compared their predictive performance. The analysis employed baseline models using lags of the closing price as explanatory variables, as well as models that added lags of the opening, high, and low prices, respectively, to the explanatory variables of the baseline models. In addition, comparisons were conducted across six forecast target periods ranging from five to ten years.

The results showed broadly similar tendencies across different forecast target periods for each of the three evaluation measures, R2, MAE, and RMSE. In particular, Random Forest Regression and Decision Tree Regression exhibited relatively high predictive performance, followed by K-Nearest Neighbors Regression.

In addition, adding lags of the opening, high, and low prices to the lags of the closing price as explanatory variables slightly improved predictive performance in some cases. This suggests that incorporating information related to intraday price formation, in addition to historical closing-price information, may contribute to improving predictive performance.

Furthermore, comparison across different forecast target periods indicated that predictive performance tends to improve as the period is extended from five to ten years. This result suggests that, when forecasting financial time series such as the Nikkei 225 using machine learning models, the period of data used in the analysis may affect predictive performance.

Overall, this study demonstrated that, in forecasting the daily closing price of the Nikkei 225, not only the type of model but also the composition of explanatory variables and the forecast target period are important factors to consider when evaluating predictive performance. This clarification is the most significant contribution of this study. Moreover, we found that, among the models compared in this study, Random Forest Regression and Decision Tree Regression performed relatively well. This finding is also an important contribution of this study.

This study is a comparative analysis centered on price data for the Nikkei 225, and whether similar results would be obtained for other markets, or when additional information such as trading volume, news, and investor sentiment is incorporated, remains a subject for future research. Nevertheless, the present study highlights the critical importance of meticulous foundational research. From this perspective, we anticipate that the findings of the present study will make a significant contribution to further research in this area and serve as a valuable basis for further investigation.

The author is grateful for the repeated cordial invitations from Joy Deng, Managing Editor of the Journal. The author also thanks the anonymous reviewers for their supportive and constructive comments on this paper. Additionally, the author appreciates the financial support from the Chuo University Grant for Special Research. Lastly, the author also thanks all the editors of this journal for their kind attention to this paper.

The author designed and conducted all of the research.