Method for asset price prediction through high frequency time series data processing

Abstract

The present invention relates to a method for asset price prediction through high frequency time series data processing, whereby high prices and low prices at a future point in time can be predicted by: collecting reference time series data and high frequency time series data for price data of assets observed in a time series; and machine-learning each piece of time series data through a machine learning model using a recurrent neural network or some other artificial neural network.

Description

Asset Price Prediction Method Using High-Frequency Time Series Data Processing

[0001] The present invention relates to a method for predicting asset prices through high-frequency time series data processing, and more specifically, to a method for predicting asset prices through high-frequency time series data processing, which collects reference time series data and high-frequency time series data for price data of assets observed in a time series format, and performs machine learning on each time series data through a machine learning model using a recurrent neural network or other artificial neural network, thereby predicting high and low prices at a future point in time compared to the present point in time.

[0002] The field of artificial intelligence (AI) has seen rapid progress since the introduction of machine learning, moving away from traditional algorithms or heuristics-based techniques. Machine learning is also widely applied in finance, and research and development on predicting asset prices using artificial neural networks (ANNs) have been actively pursued. Stocks, a representative financial asset, often have data in sequence form, particularly time series formats. Therefore, machine learning using recurrent neural networks (RNNs) is frequently attempted, but ensemble machine learning, combining convolutional neural networks (CNNs) and other artificial neural networks, is also being used. However, stock price prediction was a challenging field even before the introduction of machine learning, regardless of the prediction period (e.g., daily, weekly, etc.) or method. Despite the numerous attempts at machine learning, very few AI models have been developed and released that are accurate enough to generate high returns in real-world trading. At this point, human judgment should not be involved in the AI ​​model's predictions, the accuracy of these predictions should be significantly high across a variety of market conditions, and the returns of algorithmic trading based primarily on these predictions should significantly exceed the benchmark returns. To truly validate the effectiveness of the AI ​​model, the algorithmic trading itself must be free from human judgment.

[0003] Stock-related data typically exists in a well-refined time series format, with daily prices as the basis. This facilitates data collection and preprocessing for machine learning. However, even if daily time series data spans decades, the quantity is often insufficient for machine learning purposes, making it prone to overfitting. Separately from this data shortage, high-frequency intraday trading accounts for a significant portion of global stock market activity. Consequently, stock-related machine learning is increasingly employing shorter, high-frequency time series data, rather than daily data.

[0004] High-frequency time series data contain more transaction information and more detail than daily time series data. Therefore, if properly utilized in machine learning, high-frequency time series data can yield more detailed information and, to some extent, alleviate the overfitting problem caused by insufficient data. However, stock transaction data is generally considered noisy, so there is a potential problem that noise can increase as the frequency of time series data increases.

[0005] Furthermore, in the past, when analyzing stock price data or predicting prices using machine learning, the price value of the stock price was often directly applied. However, when applying the actual price data as the input and output data of machine learning, even after going through preprocessing steps such as normalization, considerable noise may occur in the machine learning and its prediction results. In this way, if stock price data is used as the input and output of machine learning with the actual price value as the main value, considerable noise may occur. However, existing methods such as ARIMA (auto-regressive integrated moving average) and GARCH (generalized autoregressive conditional heteroskedasticity) have limitations in that such noise is removed only with methods.

[0006] The present invention aims to solve the above problems by collecting reference time series data and high-frequency time series data for price data of assets observed in time series form, and machine learning each time series data through a machine learning model using a recurrent neural network or other artificial neural network, thereby providing a method for predicting asset prices through high-frequency time series data processing capable of predicting high and low prices at a future point in time compared to the present point in time.

[0007] In addition, the present invention provides a method for predicting asset prices through processing high-frequency time series data, which constructs a machine learning model using one or more neural networks among recurrent neural networks and other artificial neural networks to predict asset prices, and performs machine learning on reference time series and high-frequency time series data through this, thereby enabling learning of more detailed information and alleviating the overfitting problem of machine learning compared to performing machine learning using only reference time series data.

[0008] In addition, the present invention provides a method for predicting asset prices through high-frequency time series data processing, which uses price data in the form of at least one of a value representing a price fluctuation range as a percentage and a value corresponding to a percentage as input and output of machine learning for asset price prediction, thereby significantly reducing noise in price prediction results compared to using actual values ​​of stock prices as input and output of machine learning.

[0009] In addition, the present invention aims to provide a method for predicting asset prices through high-frequency time series data processing, which can significantly increase the accuracy of prediction by predicting the high and low prices of the next time interval corresponding to a future time interval compared to the current time interval of a reference time series, and can utilize the prediction results for a wider variety of stock trading strategies.

[0010] According to one embodiment of the present invention, a method for predicting asset prices through high-frequency time series data processing may include, when price data for an asset is observed in a time series format, the steps of: setting reference time series data that serve as a basis for predicting asset prices and high-frequency time series data observed at a higher frequency than the reference time series data; collecting and preprocessing the reference time series data and high-frequency time series data for the asset to extract valid data for machine learning and price prediction; configuring a machine learning model using at least one neural network among a recurrent neural network and another artificial neural network to predict the price of the asset, and performing machine learning on the valid data; using the price data as input and output of the machine learning in the form of at least one of a value expressing a price fluctuation range as a percentage and a value corresponding to the percentage, and predicting a high price and a low price of a next time interval corresponding to a future time interval compared to a current time interval of the reference time series, respectively.

[0011] Preferably, the high-frequency time series data may be data having a higher data extraction frequency than the reference time series data for price data of the same nature for the asset, so that the time series is composed of a shorter period.

[0012] In one embodiment, the market price of the asset is determined in real time in the market to which the asset belongs, and the price data is data of the asset that has a high price, a low price, and a closing price within each section of the time series, and may include one or more of stock data, bond data, derivative data, foreign exchange rate data, cryptocurrency data, and raw material data.

[0013] In one embodiment, the step of machine learning the valid data may include a step of configuring a machine learning model by connecting one or more input layers, a hidden layer to which one or more artificial neural networks are applied, and an output layer, a step of configuring one or more input layers by synchronizing the time difference between the reference time series data and high-frequency time series data in the input layer, and a step of configuring the machine learning model by connecting the input layer to one or more neural networks among the recurrent neural network and another artificial neural network in the hidden layer.

[0014] In one embodiment, the step of using the input and output of the machine learning may include a step of using the price data from a preset past point in time to a present point in time of the reference time series in an input layer in the form of at least one of a value representing a price fluctuation range as a percentage and a value corresponding to the percentage, and a step of using the price prediction value of the next point in time of the reference time series in an output layer in the form of at least one of a value representing a price fluctuation range as a percentage and a value corresponding to the percentage.

[0015] In one embodiment, the step of predicting the high and low prices of the next time interval corresponding to a future time interval compared to the current time interval of the reference time series may include the step of predicting the fluctuation range of the high price through a first machine learning model and the step of predicting the fluctuation range of the low price through a second machine learning model.

[0016] According to one aspect of the present invention, there is provided an advantage in that reference time series data and high-frequency time series data for price data of assets observed in a time series format are collected, and each time series data is machine-learned through a machine learning model using a recurrent neural network or other artificial neural network, thereby predicting high and low prices at a future point in time compared to the present point in time.

[0017] In addition, according to the present invention, by constructing a machine learning model using one or more neural networks among a recurrent neural network and other artificial neural networks for predicting asset prices and performing machine learning on reference time series and high-frequency time series data through this, more detailed information can be learned compared to performing machine learning using only reference time series data, and the overfitting problem of machine learning can be alleviated.

[0018] In addition, according to the present invention, by using price data in the form of one or more of a value representing the price fluctuation range as a percentage and a value corresponding to the percentage for asset price prediction, as input and output of machine learning, there is an advantage in that noise in the price prediction result can be significantly reduced compared to using the actual value of the stock price as input and output of machine learning.

[0019] In addition, according to the present invention, by predicting the high and low prices of the next time interval corresponding to a future time interval compared to the current time interval of the reference time series, the accuracy of the prediction can be greatly increased compared to predicting the closing price of the next time interval, and the prediction results can be utilized for a wider variety of stock trading strategies.

[0020] FIG. 1 is a diagram showing a method for predicting asset prices through high-frequency time series data processing according to one embodiment of the present invention in a series of sequential steps.

[0021] Figure 2 is a diagram illustrating a concept of implementing an asset price prediction method through high-frequency time series data processing as illustrated in Figure 1.

[0022] Figure 3 is a diagram illustrating the concept of machine learning of reference time series data and high-frequency time series data in a machine learning model using a neural network.

[0023] Figures 4a, 4b, 4c and 4d are diagrams that more specifically illustrate price data used as input and output of machine learning for asset price prediction.

[0024] FIG. 5 and FIG. 6 are diagrams showing the results of a simulated trading test on KOSPI 200 futures using an asset price prediction method through high-frequency time series data processing according to the present invention.

[0025]

[0026] Hereinafter, preferred examples are presented to aid in understanding the present invention. However, the following examples are provided solely to facilitate a better understanding of the present invention, and the scope of the present invention is not limited by the examples.

[0027] In implementing a method for predicting asset prices through high-frequency time series data processing according to one embodiment of the present invention, the target of the prediction is the KOSPI 200 futures index of the Korean stock market, and the total period of time series data related to KOSPI 200 futures may be approximately 20 years or more. Furthermore, the time series data related to KOSPI 200 futures may include the price and trading volume of the KOSPI 200 index, the price and trading volume of KOSPI 200 futures, the prices of major overseas stock indices, and trends in the Korean Won-Dollar (USD / KRW) exchange rate.

[0028] Here, "stocks" can refer to individual stocks traded on the general stock market, various derivatives, funds, or stock indices. Derivatives include index and individual stock futures and options, while funds can include exchange-traded funds (ETFs).

[0029] Hereinafter, with reference to FIGS. 1 to 6, a method for predicting asset prices through high-frequency time series data processing according to the present invention will be specifically examined.

[0030] FIG. 1 is a diagram showing a method for predicting asset prices through high-frequency time series data processing according to an embodiment of the present invention in a series of sequential steps, FIG. 2 is a diagram showing a concept of implementing the method for predicting asset prices through high-frequency time series data processing shown in FIG. 1, and FIG. 3 is a diagram showing a concept of performing machine learning on reference time series data and high-frequency time series data in a machine learning model using a neural network.

[0031] First, in the present invention, when price data for an asset is observed in a time series format, reference time series data that serves as a basis for asset price prediction and high-frequency time series data that is observed at a higher frequency than the reference time series data are set (S101).

[0032] Here, asset may refer to data whose price is determined in real time in the market to which the asset belongs, and price data may be interpreted as data of an asset that has a high price, low price, and closing price within each section of the time series, and includes at least one of stock data, bond data, derivatives data, foreign exchange rate data, virtual asset data, and raw material data.

[0033] At this time, the base time series (BT) is the time series that is the target of asset price prediction, for example, the current point in time (BT) of the time series (t) ) compared to the next point (BT) (t+1) ) can be in units of daily (trading days), intraday units of several hours or tens of minutes, or units of several trading days or more. In addition, the next time interval compared to the current time interval of the reference time series theoretically means the interval from immediately after the end of the current time interval to the end of the next time interval, and in practice, it can mean the time period within the interval when the asset market opens and the asset to be predicted is traded.

[0034] In addition, in the present invention, high-frequency time series (HT) may mean a time series in which the data sampling (extraction) period is shorter than that of the reference time series, and thus the frequency of the extracted data is higher than that of the reference time series. For example, the interval between each point in the high-frequency time series (HT) (t') ~ HT (t'+1) ) If the reference time series is daily, the high-frequency time series can be shorter, from tens of seconds to several hours, and if the reference time series is several trading days or longer, the high-frequency time series can be shorter, from tens of minutes to several days.

[0035]

[0036] Next, in the present invention, reference time series and high-frequency time series data are collected for the assets discussed above, and preprocessed to extract valid data for machine learning and price prediction (S102).

[0037] In step S102, a preprocessing step may be performed, including normalizing the reference time series data, which serves as the basis for asset price prediction, and high-frequency time series data, which is observed more frequently than the reference time series data. Furthermore, step S102 may include a preprocessing step to ensure that the reference time series data and high-frequency time series data can be appropriately utilized in a machine learning model using a neural network, as described below.

[0038] The data preprocessing process may include performing data normalization to enhance the effectiveness of machine learning on reference time series data and high-frequency time series data based on a machine learning model utilizing a neural network. For example, data normalization may be performed using one or more of the following methods: a min-max scaler, mean-variance normalization, and Z-score normalization. However, the present invention is not limited to these methods, and any method for normalizing data may be applied.

[0039] Additionally, in one embodiment, to synchronize the time difference between the reference time series data and the high-frequency time series data, an operation may be performed to ensure that the format of the date and time field values ​​in each time series data is identical. Furthermore, if necessary, an operation may be performed to distinguish between high-frequency time series data that falls within the actual time series interval of the reference time series data and those that fall outside the actual time series interval.

[0040] Additionally, in one embodiment, a task of distinguishing labels of output data among time series data can be performed for machine learning in a supervised learning manner.

[0041]

[0042] Next, in the present invention, in order to predict the price of an asset, a machine learning model using at least one neural network among a recurrent neural network (RNN) and another artificial neural network (ANN) is constructed, and a process of machine learning the reference time series data and high-frequency time series data is performed (S103).

[0043] More specifically, step S103 may be performed by connecting one or more input layers, a hidden layer to which one or more artificial neural networks are applied, and an output layer to form a machine learning model using a neural network, synchronizing the time difference between the reference time series data and high-frequency time series data in the input layer to form one or more input layers, and connecting the input layer in the hidden layer to one or more neural networks among a recurrent neural network and another artificial neural network to form the machine learning model.

[0044] In addition, the artificial neural network applied to the above machine learning model may be a convolutional neural network (CNN), a multi-layer perceptron (MLP), or other various forms of artificial neural networks in addition to a recurrent neural network.

[0045] In addition, although the present invention did not attempt machine learning that does not use artificial neural networks at all, a case in which a machine learning model using one or more artificial neural networks and a machine learning model not using artificial neural networks are connected to form a machine learning model using neural networks as a whole can be interpreted as being included in the scope of application for modifying and changing the present invention.

[0046]

[0047] In addition, the input layer of the present invention configures one or more input layers by synchronizing different frequencies, or time differences, between multiple time series data, such as reference time series data and high-frequency time series data. High-frequency time series data contains more and more detailed transaction information than reference time series data, but may have different time differences from the reference time series data. Therefore, the present invention synchronizes the time differences between these multiple time series data. In one embodiment of the present invention, the frequency of the high-frequency time series data may be synchronized with the cycle for updating the price prediction value of the reference time series data to be predicted until the end of the corresponding future time period. However, the present invention is not limited thereto.

[0048] In addition, the hidden layer of the present invention configures a machine learning model by connecting the previously configured input layer with one or more neural networks, including a recurrent neural network and other artificial neural networks. At this time, one artificial neural network used in the hidden layer may be configured as a multi-layer structure with a depth (or layer, depth) of two or more levels, and one or more neural networks, including a recurrent neural network, a convolutional neural network, a multilayer perceptron, or other artificial neural networks, may be configured by various combinations within the hidden layer.

[0049] Additionally, when applying a recurrent neural network to a machine learning model, as in one embodiment, the recurrent neural network may be formed into a multi-layer structure of two or more layers to form a hidden layer to enhance the effectiveness of machine learning. Furthermore, in one embodiment of the present invention, the sequence length of the recurrent neural network may be approximately 250 (or trading days, in the case of daily units) or less. However, the present invention is not limited thereto.

[0050] In addition, in the present invention, when constructing a machine learning model using a neural network, one or more input layers, one or more hidden layers to which an artificial neural network is applied, and an output layer can be connected to construct a single machine learning model. In addition, two or more machine learning models of the above method can be constructed, and then each machine learning model can be appropriately connected to construct an ensemble machine learning model overall. In this case, it is sufficient that the predicted result value of the same form is output from the final output layer in any machine learning model.

[0051]

[0052] Next, in the present invention, price data corresponding to one or more of the values ​​expressed as a percentage of price fluctuation range and values ​​corresponding to the percentage are used for machine learning to predict the price of an asset (S104).

[0053] More specifically, price data from a preset past point in time to the present point in time of the reference time series can be used in the input layer in the form of one or more of a value representing the price fluctuation range as a percentage and a value corresponding to the percentage, and a price prediction value for the next point in time of the reference time series can be used in the output layer in the form of one or more of a value representing the price fluctuation range as a percentage and a value corresponding to the percentage.

[0054] This is because asset-related data, especially market price data (opening / high / low / closing prices, etc.), often reflects the actual price of the asset, and asset buys and sells are often based on actual prices. Therefore, analysis, machine learning, and prediction of asset price data often rely on actual price values. However, this approach has its limitations. Specifically, using actual asset price values ​​as input and output data for machine learning can lead to increased data noise, even after preprocessing steps such as normalization.

[0055] For example, the price P1 of a stock at different points in time (t) , P2 (t) are all equal to $100.0, but the price P1 at the previous point in time (t-1) , P2 (t-1) If these are $90.0 and $110.0 respectively, then P1 (t) , P2 (t) The fluctuation rates compared to the previous point in time are +11.11% and -9.09%, respectively, which are significant differences. As another example, the price P1 of a certain stock at different points in time (t) , P2 (t) The fluctuation range compared to the previous point is all the same at +$10.0, but the price at the previous point is P1 (t-1) , P2 (t-1) If these are $100.0 and $200.0 respectively, then P1 (t) , P2 (t) The change rates compared to the previous point in time are +10.0% and +5.0%, respectively, which is a large difference.

[0056] Therefore, in the present invention, price data corresponding to one or more of the values ​​expressed as a percentage of price fluctuation range and values ​​corresponding to the percentage are used as input and output of machine learning to predict the price of an asset.

[0057]

[0058] Next, in the present invention, the high and low prices of the next time interval corresponding to a future time interval compared to the current time interval of the reference time series are predicted (S105).

[0059] More specifically, in the present invention, a high-price fluctuation range is predicted through a first machine learning model, and a low-price fluctuation range is predicted through a second machine learning model.

[0060] At this time, the first and second machine learning models may be configured in a form that is completely independent of each other, or may refer to machine learning models that are configured in a form that partially or completely shares one or more steps.

[0061] Additionally, in step S105, the present invention can apply dropout or regularization to enhance the effectiveness of machine learning. For example, dropout can be applied to an artificial neural network, and the keep probability of dropout during training can be set to 0.7 or higher. Furthermore, L2 or L1 generalization can be applied to generalize training variables, and in this case, the regularization strength of L2 generalization can be set to between 10.0 and 0.1. However, the present invention is not limited thereto.

[0062]

[0063] Furthermore, in one embodiment, the present invention can be utilized for virtual or actual trading of an asset based on the predicted high price fluctuation range through the first machine learning model and the predicted low price fluctuation range through the second machine learning model. For example, based on the predicted high and low price fluctuation ranges, if the real-time current price of the asset matches or approaches a certain value, or if the real-time current price of the asset exceeds a certain value upward or downward, a buy or sell signal for the asset can be generated in real time. In this case, the real-time buy or sell signal may be for the intraday market (current) price or for the intraday fixed price. However, the present invention is not limited thereto.

[0064] In addition, when generating a real-time buy or sell signal for an asset through a method for predicting asset prices through high-frequency time series data processing according to one embodiment of the present invention and utilizing the signal for virtual or actual trading, it should be noted that a trading environment must be established in which virtual or actual trading for the asset can be executed within a minimum real-time delay time immediately after the trading signal is generated.

[0065]

[0066] Figures 4a, 4b, 4c, and 4d are diagrams that more specifically illustrate the price data used as input and output for machine learning for asset price prediction. First, Figure 4(a) is an example in which the reference time series data is collected as a daily time series. This shows that when the asset price prediction target is daily prices, the reference time series data becomes a daily time series. Figure 4(b) is an example in which the high-frequency time series data is collected as an intraday 10-minute time series. This shows that when the reference time series data is daily, high-frequency time series data is extracted at a high frequency from a shorter interval.

[0067] Figure 4(c) illustrates an example of price prediction results for KOSPI 200 futures using a machine learning model utilizing a neural network. The present invention predicts the daily high and low prices of KOSPI 200 futures, updates these prices in real time every 10 minutes during trading, and predicts the percentage (%) of the stock price fluctuations rather than the actual price. Furthermore, the present invention provides the actual price of the stock price, which can be easily utilized for trading, along with the predicted results. Through this, the present invention reduces noise that may appear in the input and output data of machine learning, thereby increasing the accuracy of the price prediction results.

[0068] Figure 4(d) shows the results of a daily price prediction test for KOSPI 200 futures using a machine learning model using a neural network, and the results are from prediction tests conducted in 2018, 2019, and 2020, respectively. Prediction accuracy was measured by calculating the error between the predicted high and low price fluctuations (%) and the actual high and low price fluctuations (%) for the same day of KOSPI 200 futures. At this time, the method of summing the errors of the predicted values ​​updated every 10 minutes during the day is not a simple average, but a time-weighted average that uses the time remaining from each prediction time to the market close as a weight, so that predictions made early in the market have a greater weight than predictions made later in the market. Accordingly, the average error in predicting the daily high and low prices of KOSPI 200 futures over the past three years is ±0.20%p based on the daily volatility (%). Even considering that 2020 was a period of high market volatility, the prediction error is considered to be quite low.

[0069]

[0070] FIG. 5 and FIG. 6 are diagrams showing the results of a virtual trading test conducted on KOSPI 200 futures using an asset price prediction method through high-frequency time series data processing according to the present invention.

[0071] First, looking at Figure 5, using the asset price prediction method through high-frequency time series data processing according to the present invention, the results of predicting the daily high and low prices of KOSPI 200 futures were utilized to automatically generate real-time buy / sell signals during the day based on an algorithm, and a test was performed in a virtual trading environment in which trading was automatically executed according to the signals, and the results as shown in the table in Figure 5 were confirmed.

[0072] More specifically, as a result of conducting virtual trading tests for 2018, 2019, and 2020, the benchmark (KOSPI 200 Index) return for the same period averaged 8.44% per year, while the virtual trading return for KOSPI 200 futures using machine learning prediction results recorded an average annual return of 55.59%. At this time, the futures trading conditions for the virtual trading test were applied, such as futures leverage of 2x, a trading commission of 0.003%, and a trading execution error (slippage) of 1 tick (0.05p). In addition, the accumulation method of the futures return on the day was simple, not compound.

[0073] These results indicate that even without futures leverage, the average annual return was more than three times that of the benchmark. Considering that futures leverage is more than twice that of actual trading, this suggests that returns that significantly exceed the benchmark can be achieved even in actual trading, not just virtual trading.

[0074] Additionally, Figure 6 presents the results of a virtual trading test on KOSPI 200 futures, graphing the cumulative daily virtual trading returns and daily benchmark returns for a total of one year, from January 2 to December 30, 2020. The compound interest method was used to accumulate the benchmark returns.

[0075]

[0076] Although the present invention has been described above with reference to preferred embodiments thereof, it will be understood by those skilled in the art that various modifications and changes may be made to the present invention without departing from the spirit and scope of the present invention as set forth in the claims below.

Claims

1. When price data for an asset is observed in the form of a time series, a step of setting base time series data that serves as a basis for asset price prediction and high-frequency time series data that is observed at a higher frequency than the base time series data; Here, the high-frequency time series data is data whose data extraction frequency is higher than that of the reference time series data for price data of the same nature for the asset, and thus the time series is composed of shorter periods. For the above assets, a step of collecting and preprocessing the reference time series data and high-frequency time series data to extract valid data for machine learning and price prediction; In order to predict the price of the above asset, a step of performing machine learning on the valid data by constructing a machine learning model using at least one neural network among a recurrent neural network and other artificial neural networks; In order to predict the price of the above asset, the step of using the price data as input and output of the machine learning in the form of at least one of a value representing the price fluctuation range as a percentage and a value corresponding to the percentage; and A method for predicting asset prices through high-frequency time series data processing, characterized by including a step of predicting the high and low prices of the next time interval corresponding to a future time interval compared to the current time interval of the above-mentioned reference time series.

2. In paragraph 1, The above assets are, A method for predicting asset prices through high-frequency time series data processing, characterized in that the price is determined in real time in the market to which the asset belongs, and the price data is data of an asset that has a high price, a low price, and a closing price within each section of the time series, and includes at least one of stock data, bond data, derivative data, foreign exchange rate data, virtual asset data, and raw material data.

3. In paragraph 1, The step of machine learning the above valid data is: A step of configuring a machine learning model by connecting one or more input layers, one or more hidden layers to which an artificial neural network is applied, and an output layer; A step of configuring one or more input layers by synchronizing the time difference between the reference time series data and the high-frequency time series data in the input layer; and A method for predicting asset prices through high-frequency time series data processing, characterized in that it comprises a step of configuring the machine learning model by connecting the input layer in the hidden layer with at least one neural network among the recurrent neural network and other artificial neural networks.

4. In paragraph 1, The steps used for input and output of the above machine learning are: A step of using the price data from a preset past point in time to the present point in time of the above-mentioned reference time series in the input layer in the form of at least one of a value representing the price fluctuation range as a percentage and a value corresponding to the percentage; and A method for predicting asset prices through high-frequency time series data processing, characterized in that it includes a step of using, in an output layer, a price prediction value of the next time interval of the above-mentioned reference time series in the form of at least one of a value expressing the price fluctuation range as a percentage and a value corresponding to the percentage.

5. In paragraph 1, The step of predicting the high and low prices of the next time interval corresponding to the future time interval compared to the current time interval of the above reference time series is as follows: A step of predicting the high price fluctuation range through a first machine learning model; and A method for predicting asset prices through high-frequency time series data processing, characterized in that it includes a step of predicting the fluctuation range of the low price through a second machine learning model.

Images