JP7759053B2 - Stock price evaluation device, stock price evaluation program, stock price monitoring program, asset management program, stock price evaluation method, stock price monitoring method, and asset management method - Google Patents

Description

Article 30, paragraph 2 of the Patent Act applies. Publication date: February 15, 2021. Publication address: https://www.ibaraki.ac.jp/news/2021/02/15011101.html https://www.daiwa-am.co.jp/company/press-release/20210215_90.pdf Publication date: February 26, 2021. Publication address: https://www.digital.asahi.com/articles/ASP2T55QGP2TULZU009. html Published date: February 23, 2021 Published address: http://www.ieice-taikai.jp/2021general/jpn/registration.html Meeting name: Complex Communication Science Study Group Date held: November 18-19, 2021 Published date: December 4, 2021 Published address: https://nlsw2021.org/index.html

The present invention particularly relates to a stock price valuation device, a stock price valuation program, a stock price monitoring program, an asset management program, a stock price valuation method, a stock price monitoring method, and an asset management method, all of which are related to the valuation, monitoring, and management of stock prices.

With the recent advancement of ICT (Information and Communication Technology), there has been a growing trend to apply artificial intelligence to financial operations. In particular, technological innovation is progressing in the field of financial engineering, known as FinTech, which combines finance and IT (Information Technology).

Patent Document 1 describes a conventional stock price prediction system that includes a prediction set acquisition means for acquiring a set of stock price prediction values based on input data indicating factors that change stock prices, and a collective intelligence acquisition means for acquiring collective intelligence of stock prices with high prediction accuracy from the set of stock price prediction values.

Japanese Patent Application Laid-Open No. 2018-200513

J. A. Ohlson, “Earning, Book Values, and Dividends in Equity Valuation: An Empirical Perspective”, 2001, Contemporary Accounting Research, vol. 18, no. 1, p. 107-120

However, there was a technical demand for calculating an intrinsic asset value, which is a more realistic theoretical stock price than the technology in Patent Document 1, and evaluating the difference between this and the asset price, which is the actual stock price.

The present invention was made in light of these circumstances, and its objective is to provide a stock price evaluation device that solves the above-mentioned problems.

The stock price evaluation device of the present invention comprises a machine learning unit that performs machine learning on the relationship between financial information of multiple companies over multiple periods and asset prices, and estimates a dynamics function common to the multiple companies; an intrinsic value calculation unit that calculates the intrinsic value of an asset using the dynamics function estimated by the machine learning unit; and a value difference calculation unit that calculates the difference between the intrinsic value calculated by the intrinsic value calculation unit and the asset price as crowd psychology , and is characterized in that the value difference calculation unit draws a list of the crowd psychology based on the period and type of stock .
The stock price evaluation device of the present invention is characterized in that the machine learning unit also uses, in the machine learning, performance forecasts by management or analysts indicated in the financial results summary.
The stock price evaluation device of the present invention is characterized in that the dynamics function is an extension of the OVM model (Ohlson Valuation Model) to a nonlinear model.
The stock price evaluation device of the present invention is characterized in that the value difference calculation unit detects abnormal stock prices based on the crowd psychology .
The stock price evaluation device of the present invention is characterized by further comprising an asset management unit that identifies undervalued stocks and overvalued stocks based on the herd psychology calculated by the value difference calculation unit.
The stock price evaluation device of the present invention is characterized in that the asset management department manages a portfolio by going long on the undervalued stocks and short on the overvalued stocks.
The stock price valuation program of the present invention is a stock price valuation program executed by a stock price valuation device, which performs machine learning on the relationship between financial information of multiple companies over multiple periods and asset prices, estimates a dynamics function common to the multiple companies, calculates the intrinsic value of assets using the estimated dynamics function, calculates the difference between the calculated intrinsic value and the asset price as crowd psychology , and draws a list of the crowd psychology based on the period and type of stock .
The asset management program of the present invention is characterized in that it identifies undervalued stocks and overvalued stocks based on the herd psychology calculated by the stock price evaluation program , and manages a portfolio by going long on the identified undervalued stocks and short on the identified overvalued stocks.
The stock price valuation method of the present invention is a stock price valuation method executed by a stock price valuation device, which performs machine learning on the relationship between financial information of multiple companies over multiple time periods and asset prices, estimates a dynamics function common to the multiple companies, calculates the intrinsic value of assets using the estimated dynamics function, calculates the difference between the calculated intrinsic value and the asset price as crowd psychology , and displays a list of the crowd psychology based on the period and type of stock .
The stock price monitoring method of the present invention is characterized in that abnormal stock prices are monitored based on the crowd psychology calculated by the stock price evaluation method.
The asset management method of the present invention is characterized in that it identifies undervalued stocks and overvalued stocks based on the herd psychology calculated by the stock price valuation method, and manages a portfolio by going long on the identified undervalued stocks and short on the identified overvalued stocks.

This invention provides a stock price evaluation device that can calculate intrinsic asset value and evaluate the difference from asset prices by using machine learning to estimate a dynamics function based on the relationship between asset prices and financial information for multiple companies over multiple time periods, calculating the intrinsic value of assets using the estimated dynamics function, and calculating the difference between this intrinsic value and asset prices as herd psychology.

1 is a block diagram showing a control configuration of a stock price evaluation device according to an embodiment of the present invention. 1 is a flowchart of a stock price evaluation process according to an embodiment of the present invention. FIG. 3 is a conceptual diagram of the stock price evaluation process shown in FIG. 2. FIG. 3 is a conceptual diagram of the list drawing process shown in FIG. 2 . 1 is a graph showing the correlation between the intrinsic value estimate ^V _i,t and the asset price realization P _i,t according to an embodiment of the present invention. 10 is a graph showing the average value and standard deviation of the deviation rate ξ _i,t according to an example of the present invention. 10 is a graph showing the average value and standard deviation of the deviation rate ξ _i,t after correction according to an example of the present invention. 1 is a graph showing the importance of explanatory variables according to an example of the present invention. 1 is a graph showing cumulative returns of a long portfolio and a short portfolio according to an embodiment of the present invention. 1 is a graph showing cumulative active returns of a long portfolio for each quantile according to an embodiment of the present invention. 1 is a graph showing cumulative returns of a long-short portfolio according to an embodiment of the present invention.

<Embodiment>
[Configuration of stock price evaluation device 1]
First, with reference to FIG. 1, the control configuration of a stock price evaluation device 1 according to an embodiment of the present invention will be described.
The stock price evaluation device 1 is an information processing device for evaluating stock prices, such as a PC (Personal Computer), a server, a general-purpose computer, or a smartphone.

The stock price evaluation device 1 includes a control unit 10, an image processing unit 11, a memory unit 12, an input unit 13, a display unit 14, and a transmission/reception unit 15. Each unit is connected to the control unit 10 directly or via a bus, and its operation is controlled by the control unit 10.

The control unit 10 is an information processing unit such as a CPU (Central Processing Unit), MPU (Micro Processing Unit), GPU (Graphics Processing Unit), or ASIC (Application Specific Integrated Circuit).
The control unit 10 reads out a control program stored in the ROM or HDD of the storage unit 12, and loads and executes the control program in the RAM, thereby causing each unit of the functional blocks described below to operate. The control unit 10 also controls the entire device in accordance with predetermined instruction information input from the input unit 13 or an external terminal (not shown).

The image processing unit 11 is a control and calculation means including a GPU (Graphics Processing Unit), a DSP (Digital Signal Processor), an FPGA (Field-Programmable Gate Array), etc. The image processing unit 11 may also function as a machine learning accelerator.

The storage unit 12 is a non-transitory recording medium such as a semiconductor memory such as a ROM (Read Only Memory) or a RAM (Random Access Memory) or a HDD (Hard Disk Drive).
The ROM and HDD of the storage unit 12 store various programs and data including an OS (Operating System) and application software (hereinafter simply referred to as "apps") for controlling the operation of the stock price evaluation device 1. In addition, the storage unit 12 may also store user account settings.

The transmitter/receiver 15 is a network connection unit including a LAN (Local Area Network) board, a wireless LAN (WiFi) board, a Bluetooth (registered trademark) or short-range wireless transceiver, etc. for connecting to an external network. The external network in this embodiment is, for example, an intranet such as a LAN, a WAN (Wide Area Network), or a mobile phone network.
Furthermore, the transmitting/receiving unit 15 may be capable of connecting to various devices via a USB (Universal Serial Bus) or the like.

The input unit 13 is an input device that includes a keyboard and a pointing device. The pointing device of the input unit 13 may be, for example, a mouse, a touch panel, a touchpad, an electromagnetic digitizer, etc. The input unit 13 acquires various instructions from the user.

The display unit 14 is an LCD (Liquid Crystal Display), an organic EL (Organic Electro-Luminescence) display, etc. The display unit 14 displays various types of information in a GUI (Graphical User Interface) or a CUI (Character-based User Interface).
Furthermore, the input unit 13 and the display unit 14 may be integrally formed, such as a touch panel display.

In the stock price evaluation device 1, the control unit 10 and the image processing unit 11 may be integrally formed, such as in a CPU with a built-in GPU, a chip-on-module package, or an SOC (System On a Chip).
Furthermore, the control unit 10 and the image processing unit 11 may have built-in RAM, ROM, flash memory, or the like.
Furthermore, the stock price evaluation device 1 may be connected to a printer or the like that prints out various evaluation results and the like.

Next, the functional configuration of the stock price evaluation device 1 will be described.
The control unit 10 of the stock price evaluation device 1 includes a machine learning unit 100, an intrinsic value calculation unit 110, a value difference calculation unit 120, and an asset management unit 130.
The storage unit 12 stores financial asset forecast data 200 , model data 210 , intrinsic value data 220 , monitoring data 230 , management data 240 , a stock price evaluation program 300 , a stock price monitoring program 310 , and an asset management program 320 .

The machine learning unit 100 performs machine learning on the relationship between financial information and asset prices of multiple companies over multiple periods, and estimates a dynamics function common to the multiple companies.
In this embodiment, the machine learning unit 100 may use, in machine learning, performance forecasts by management or analysts indicated in the financial results summary.
Furthermore, in this embodiment, the dynamics function may be an extension of the OVM model (Ohlson Valuation Model) described in Non-Patent Document 1 to a non-linear model.

The intrinsic value calculation unit 110 calculates the intrinsic value of the asset using the dynamics function estimated by the machine learning unit 100.

The value difference calculation unit 120 calculates the difference between the intrinsic value calculated by the intrinsic value calculation unit 110 and the asset price as herd psychology.
Additionally, in this embodiment, the value difference calculation unit 120 can also detect abnormal stock prices based on crowd psychology. In this case, the value difference calculation unit 120 can draw a list of crowd psychology based on the period and type of stock, and generate monitoring data 230.

The asset management unit 130 identifies undervalued stocks and overvalued stocks based on the herd psychology calculated by the value difference calculation unit 120.
Specifically, in this embodiment, the asset management unit 130 can manage a portfolio by going long on undervalued stocks and short on overvalued stocks.

Financial asset forecast data 200 is data on stock prices, financial information used for corporate valuation (financial data, panel data), and other data obtained from so-called cloud servers via an external network. In this embodiment, financial asset forecast data 200 also includes performance forecasts by management or analysts as shown in the financial results summary.

The model data 210 is data of a model learned by machine learning.
Specifically, in this embodiment, data of a model calculated as a dynamics function F in machine learning may be stored.
Furthermore, the model data 210 may be retrained by acquiring actual market data, as described below.

The intrinsic value data 220 is data that indicates the calculated intrinsic value of an asset and crowd psychology.
Specifically, in this embodiment, the intrinsic value data 220 may include data on theoretical stock prices for each company.
Furthermore, the intrinsic value data 220 may include, as herd psychology, a PIR index that indicates the dissociation between the current stock price and the theoretical stock price, which will be described in detail in an example below.
In these cases, the intrinsic value data 220 may be a database that stores the current stock price, theoretical stock price, and PIR index for each stock.

The monitoring data 230 is image data or database data that shows a list of crowd psychology based on a period and a type of stock.
The monitoring data 230 may be viewable and searchable, for example, from a web browser or a dedicated program.

Management data 240 is data on each stock used in portfolio management by the asset management unit 130. This management data 240 includes, for example, each stock for short and long management, as described below, and may be dynamically set based on actual market data.

The stock price valuation program 300 is an app primarily used for machine learning, intrinsic value calculation, and crowd psychology calculation. The stock price valuation program 300 may be run periodically using, for example, a task manager or "CRON."

The stock price monitoring program 310 is an application that detects abnormal stock prices using herd psychology and displays a list of herd psychology. The stock price monitoring program 310 may be executed periodically in conjunction with the execution of the stock price evaluation program 300, or may be executed on demand at the user's request. In addition, if an abnormal stock price is detected, a warning or the like may be sent to the user.

The asset management program 320 is an app for asset management. The asset management program 320 may also be run periodically or at the user's request for stock price evaluation.

Here, the control unit 10 of the stock price evaluation device 1 is made to function as a machine learning unit 100, an intrinsic value calculation unit 110, a value difference calculation unit 120, and an asset management unit 130 by executing the stock price evaluation program 300, the stock price monitoring program 310, and the asset management program 320 stored in the memory unit 12.
In addition, the image processing unit 11 can be used as an accelerator or the like when performing various calculations in the machine learning unit 100, the intrinsic value calculation unit 110, the value difference calculation unit 120, and the asset management unit 130.
Furthermore, each unit of the above-described stock price evaluation device 1 serves as a hardware resource for executing the stock price evaluation method, stock price monitoring method, and asset management method of the present invention.
Note that a part or any combination of the above-described functional configurations may be configured in terms of hardware or circuits using ICs, programmable logic, FPGAs, or the like.

[Stock Price Valuation Process by Stock Price Valuation Device 1]
Next, the stock price evaluation process performed by the stock price evaluation device 1 according to the embodiment of the present invention will be described with reference to FIGS.
In the stock price evaluation process of this embodiment, machine learning is used to separate theoretical stock prices from herd psychology. That is, the current stock price is calculated by adding the true value (fundamental value) that is the theoretical stock price to the value of herd psychology. Then, the deviation between this theoretical stock price and the actual stock price is monitored, stocks with abnormal stock prices are monitored, and risk assessment of the entire stock market, portfolio management, etc. are performed.
In the stock price evaluation process of this embodiment, the control unit 10 mainly executes the programs stored in the storage unit 12 in cooperation with the other units, using hardware resources.
The stock price evaluation process will be described in detail below for each step with reference to the flowchart of FIG.

(Step S101)
First, the machine learning unit 100 performs machine learning processing.
The machine learning unit 100 obtains financial information such as net profits and forecast dividends for multiple periods for multiple companies, as well as asset prices such as current stock prices, from a server on the cloud via an external network, and stores them in financial asset forecast data 200.
Then, the machine learning unit 100 performs machine learning of the relationship between financial information and asset prices using this financial asset forecast data 200. During this machine learning, the machine learning unit 100 may use an accelerator of the image processing unit 11 or the like.
As a result, the machine learning unit 100 estimates a dynamics function common to multiple companies and stores it in the model data 210.

3, stock prices deviate from true value (fundamental value) because they are influenced by the herd psychology of investors. The machine learning unit 100 can generate a model that outputs intrinsic value by performing machine learning on financial information and asset prices of multiple companies regardless of the time period.
In this case, in order to improve the learning accuracy, the machine learning unit 100 also performs machine learning using the performance forecast by the management or the performance forecast by the analyst, which is shown in the financial results summary. This enables the machine learning unit 100 to efficiently learn a model that uses observable asset prices as the objective variable, as shown in the examples described later.

Specifically, the machine learning unit 100 performs machine learning on a model that is a nonlinear extension of the multiple regression type OVM (Olson model) described in Non-Patent Document 1. This machine learning can be performed using, for example, artificial neural networks with various layer structures, kernel machines, decision trees, Bayesian networks, other Bayesian statistical methods, statistical methods such as regression and multiple regression, and various other artificial intelligence and statistical methods.

(Step S102)
Next, the intrinsic value calculation unit 110 performs an intrinsic value calculation process.
The intrinsic value calculation unit 110 reads out the machine-learned model data 210 and inputs the financial information of each company into the calculated dynamics function to calculate the intrinsic value of the asset.

3, in this embodiment, the intrinsic value calculation unit 110 inputs the financial information of the company into the calculated dynamics function, thereby removing psychological biases such as noise and outputting only the intrinsic value (true value, fundamental value) of the asset. The intrinsic value calculation unit 110 can regard this intrinsic value of the company as the "theoretical stock price."
The intrinsic value calculation unit 110 stores the intrinsic value (theoretical stock price) calculated for each company (stock) in the intrinsic value data 220 .

(Step S103)
Next, the value difference calculation unit 120 performs a value difference calculation process.
The value difference calculation unit 120 calculates the difference between the intrinsic value calculated by the intrinsic value calculation unit 110 and the asset price as herd psychology.

As shown in Figure 3, in this embodiment, the value difference calculation unit 120 can calculate the herd psychology (psychological bias) itself by calculating the difference between the asset price, which is the current stock price (realized stock price), and the intrinsic value of the asset, which is the calculated theoretical stock price, for each company (stock), i.e., the "deviation" from the intrinsic value.

In this embodiment, the value difference calculation unit 120 can calculate this herd psychology as PIR (Price Intrinsic-Value Ratio), which will be described in detail in the examples below. PIR is a deviation rate similar to PBR (Price Book-Value Ratio), and is calculated by regressing the herd psychology of each stock with the average stock price such as TOPIX (TOPIX) and calculating the "residual." PIR makes it possible to remove the influence of the entire market and the constant undervaluation and overvaluation specific to each stock.
The value difference calculation unit 120 stores the PIR index calculated for each company (stock) in the intrinsic value data 220.

At this time, the value difference calculation unit 120 can also detect abnormal stock prices based on the calculated crowd psychology.
For example, the value difference calculation unit 120 may determine that a stock price is abnormal when the deviation rate in the PIR is outside a specific range from the standard deviation. Alternatively, the value difference calculation unit 120 may detect this when a stock is identified as an "undervalued stock" or an "overvalued stock" as described below. That is, for example, as shown in the following example, the value difference calculation unit 120 may determine that a stock whose PIR is selected within the upper or lower 20% range among all stocks for which a PIR has been calculated is an abnormal stock price.
When such an abnormal stock price is detected, the value difference calculation unit 120 can also warn the user about the stock (company) in question.

(Step S104)
Next, the value difference calculation unit 120 performs a list drawing process.
The value difference calculation unit 120 displays a list of crowd psychology as monitoring data 230 based on the period and type of stock.
An example of this tabulated monitoring data 230 is shown in Figure 4. In this example, the vertical axis represents stocks categorized by sector, and the horizontal axis represents each period. The graph and white circles in the figure represent an example of ETF purchases by the Bank of Japan.

In this way, the value difference calculation unit 120 can perform risk prediction management by visualizing crowd psychology. That is, it is possible to get a comprehensive view of whether each company's stock price is "overvalued" or "undervalued." Specifically, in this example of monitoring data 230, the more companies show synchronized overvaluation and undervaluation, as seen in the striped areas, the more likely it is that a bubble due to changes in the market environment is suspected.
The value difference calculation unit 120 can also warn the user about abnormal judgments such as "suspected bubble."

Specifically, when plotted in the same manner as the example shown in Figure 4, when the price falls from the year-to-date high on February 16 to 27,448 yen on May 13, the black stripes of the abnormality pattern can indicate a "sign" of a fall. The value difference calculation unit 120 can issue a warning of this.
In this case, the value difference calculation unit 120 can determine from the correlation between the white circles and the striped patterns in the figure that the Bank of Japan's ETF purchases are stimulating herd psychology, and can also point this out.
Although Figure 4 shows the herd psychology in black and white, it is possible to clearly express this herd psychology using colors such as red (overpriced) to blue (underpriced) or green (overpriced) to red (underpriced).

(Step S105)
Next, the asset management unit 130 performs asset management processing.
The asset management unit 130 selects stocks for asset management based on the strength of the divergence between the intrinsic value stored in the intrinsic value data 220 and herd psychology, and stores the results in the management data 240. In other words, the asset management unit 130 identifies undervalued stocks and overvalued stocks.
Then, the asset management unit 130 refers to the management data 240 and manages the portfolio by going long (buying) on undervalued stocks and short (selling short) on overvalued stocks.
Specifically, the asset management unit 130 identifies undervalued and overvalued stocks based on the difference between the intrinsic value and the asset price, that is, the strength of herd psychology, which is the divergence between the theoretical stock price and the current stock price.

More specifically, the asset management department 130 divides all stocks into quintiles (every 20%) based on the PIR of each company (stock). Then, the asset management department 130 goes long on the first quintile (undervalued stocks) and short on the fifth quintile (overvalued stocks). This makes it possible to efficiently obtain spread returns, as will be shown in the examples below.

Thereafter, the value difference calculation unit 120 calculates the theoretical stock price every time new financial information is released. As a result, the asset management unit 130 updates the stocks that make up the portfolio and stores them in the management data 240. In other words, machine learning re-training and portfolio rebalancing are performed every specific period.
This completes the stock price evaluation process according to the embodiment of the present invention.

The above configuration can provide the following effects.
Numerous financial market anomalies that cannot be explained by the EMH (Efficient Market Hypothesis) or the CAPM (Capital Asset Pricing Model), such as the value stock effect and the small-cap effect, have been reported in the past, and the risk premium hypothesis and mispricing hypothesis are representative of the mechanisms by which these anomalies arise.
Of these, the mispricing hypothesis is supported by behavioral economics and recognizes the possibility that the price of a financial asset may diverge from its intrinsic value. This divergence (mispricing) occurs due to the psychological biases of investors. Because modeling human psychology is difficult, fundamental analysis models intrinsic value. Representative deductive models include the DDM (Dividend Discount Model) and the RIM (Residual Income Model), but because both include growth forecasts related to corporate finances, intrinsic value has not been uniquely determined.
Therefore, it is possible to consider inductive modeling from data. For example, the OVM model described in Non-Patent Document 1 exists, but it makes numerous assumptions and the intrinsic value, which is the objective variable, is unobservable.

In contrast, the stock price evaluation device 1 according to an embodiment of the present invention is characterized by comprising a machine learning unit 100 that performs machine learning on the relationship between asset prices and financial information of multiple companies over multiple time periods and estimates a dynamics function common to the multiple companies, an intrinsic value calculation unit 110 that calculates the intrinsic value of the asset using the dynamics function estimated by the machine learning unit 100, and a value difference calculation unit 120 that calculates the difference between the intrinsic value calculated by the intrinsic value calculation unit 110 and the asset price as a herd psychology.

By configuring it in this way, it is possible to calculate the intrinsic asset value based on the divergence between the current stock price, which indicates the company's value, and the asset price, and evaluate the difference from the asset price. In other words, machine learning can extract herd psychology, which is mispricing caused by the psychology of collective investors. This allows for more accurate evaluation of theoretical stock prices.

Furthermore, as global interest rates remain low for a prolonged period, investing in risky assets has become as important as lending for regional financial institutions, primarily banks.
On the other hand, the Financial Services Agency is calling on financial institutions to improve their risk management, for example by carefully selecting the risks they should take, monitoring the amount of risk, and controlling the amount of risk.
In contrast, the stock price evaluation device 1 according to this embodiment can efficiently use machine learning to enhance risk management, thereby improving the efficiency of risk management and portfolio management by using it in combination with the specialized knowledge of experienced people.

Furthermore, in the stock price evaluation device 1 according to the embodiment of the present invention, the machine learning unit 100 can use, in machine learning, performance forecasts by management or analysts indicated in the financial results summary.
This configuration allows for the use of expert knowledge and tacit knowledge from the parties involved in addition to objective observations of financial information. In other words, it creates room for not only data-driven, objective AI estimation but also subjective human expertise. This is expected to have the effect of flexibly expanding the framework of machine learning, which tends to be rigid.

In recent years, asset management has become no exception, with the increasing use of AI and machine learning. However, these AI and machine learning technologies have presented the so-called "black box" problem, where the results are difficult to interpret.
In this regard, the conventional OVM model described in Non-Patent Document 1 places importance on interpretability, but includes many constraints such as the assumption of linearity, etc. For this reason, it has not been possible to accurately calculate the original theoretical stock price.

In contrast to this, in the stock price evaluation device 1 according to the embodiment of the present invention, the dynamics function is characterized by being an extension of the OVM model (Ohlson Valuation Model) to a nonlinear model.
With this configuration, the OVM model, a corporate valuation model based on corporate financial information, can be expanded with machine learning. This reduces the shortcomings of conventional corporate valuation models and enables more accurate calculation of intrinsic value. In other words, compared to conventional theoretical stock price models, the versatility of the model is enhanced by introducing nonlinearity through machine learning, making it possible to calculate more realistic theoretical stock prices.
Furthermore, the dynamics function according to this embodiment is less of a black box, as is the case with "AI investments" that utilize alternative data. This increases explainability and transparency, making it easier for users to accept, like undervalued stock investments, and enabling the creation of publicly offered funds.

In contrast to this, in the stock price evaluation device 1 according to this embodiment, the value difference calculation unit 120 is characterized in that it detects abnormal stock prices based on crowd psychology.
With this configuration, it is possible to constantly monitor the deviation from the current stock price, thereby making it possible to evaluate the risk of the entire stock market. In other words, the stock price evaluation device 1 according to this embodiment makes it possible to continuously monitor a huge number of investment stocks in real time at high speed, in large quantities, automatically, objectively, and stably. This makes it possible to notify users of abnormal stock prices and obtain buy/sell instructions.

Furthermore, in the stock price evaluation device 1 according to the embodiment of the present invention, the value difference calculation unit 120 is characterized in that it draws a list of crowd psychology based on the period and the type of stock.
This configuration makes it possible to use it to determine investment positions in asset management. In other words, it acts as a "protective gear" or "umbrella" for asset management, helping to assess market conditions and make buying and selling decisions. Specifically, as shown in the example of monitoring data 230 in Figure 4, it makes it possible to observe the market environment objectively and in real time, like a "nowcast" on a weather radar.

Furthermore, the stock price evaluation device 1 according to the embodiment of the present invention is characterized by further comprising an asset management unit 130 that identifies undervalued stocks and overvalued stocks based on the crowd psychology calculated by the value difference calculation unit 120.
By configuring it in this way, it can be used to judge whether a stock is undervalued or overvalued, just like PBR (price-to-book ratio) or V/P ratio (value-to-price ratio). This makes it possible to build a portfolio in a direction that corrects mispricing (herd psychology) and observe subsequent investment performance. It also makes it possible to distribute undervalued and overvalued stocks through financial information distribution vendors.

Furthermore, in recent years, empirical analysis has shown that the value stock effect has disappeared, pointing out the ineffectiveness of this type of investment style. This is because, although identifying value stocks relied on corporate valuation models, it was difficult to calculate true corporate value using conventional models.

In contrast, in the stock price evaluation device 1 according to an embodiment of the present invention, the asset management unit 130 is characterized by managing a portfolio by going long on undervalued stocks and short on overvalued stocks.

By configuring it in this way, it is possible to significantly improve the "return/risk ratio" compared to the past. In other words, considering the efficiency of financial markets, the greater the deviation, the more likely it is that a correction to the fair price will occur, and by managing a portfolio using this correction process, it is possible to obtain good investment performance, as will be shown in the examples below.
Specifically, by utilizing the corporate valuation model according to this embodiment, which is more accurate than conventional models, in portfolio management, it is possible to confirm profitability that exceeds the risk premium through statistical testing.
As a result, although the financial industry has pointed out a slump in "undervalued stock investment" in recent years, the operation according to this embodiment can demonstrate its effectiveness.

Other Embodiments
In the above embodiment, an example in which a PC or the like is used alone as the stock price evaluation device 1 has been described.
However, the stock price evaluation device 1 may be configured as a server-client type system using a plurality of servers, general-purpose computers, etc., which are accessed from terminals such as PCs and smartphones.

Furthermore, information such as theoretical stock prices, undervalued stocks, and overvalued stocks calculated by the stock price evaluation device 1 according to this embodiment can be distributed by a financial information distribution vendor to terminals such as PCs and smartphones in addition to financial information. Here, the stock price evaluation device 1 is not limited to Japanese stocks, and can be applied to global stocks including U.S. stocks.
This configuration makes it possible to accommodate a variety of configurations.

Furthermore, in the above embodiment, the OVM model is used as the original model and is nonlinearly extended, but other types of enterprise valuation models may also be used.
Furthermore, it is possible to introduce dummy variables for industries or specific companies into the model, and to change the frequency of machine learning retraining and portfolio rebalancing to monthly, quarterly, semi-annually, annually, etc.

In the above embodiment, an example has been described in which the current stock price is calculated as the asset price and the theoretical stock price is calculated as the asset value.
However, the stock price evaluation device according to the present invention can be similarly applied to the evaluation, monitoring, and management of bonds, FX (Foreign Exchange), commodity investments (commodity futures trading), and other assets in addition to stock prices.
By configuring it in this way, it is possible to conduct transactions based on the discrepancy between asset prices and intrinsic value, and it is expected that assets will increase.

Next, the present invention will be described in more detail based on examples, but the following specific examples do not limit the present invention.

(Corporate valuation model)
Let _Vt be the intrinsic value of the stock price _Pt and _dt be the dividend per share. The rate of return (cost of capital) expected by investors is calculated using the following formula (1):

Here, _Et represents the net income for the period t. Assuming that there is no mispricing due to the efficiency of financial markets, _Pt = _Vt , and therefore the DDM (Dividend Discount Model) can be obtained by the following equation (2).

Furthermore, let _Bt be the net assets per share, _xt be the net income per share, _and ^xat = _xt - _rBt-1 be the residual income. By applying the clean surplus relationship ( _Bt = Bt _-1 + _xt - _dt ) to equation (2), the RIM (residual income model) can be obtained as shown in equation (3) below.

Since this model involves predictions for an infinite future, some assumptions are required about the dynamics of time evolution. Therefore, according to Non-Patent Document 1, the linear information model (LIM) of the following equations (4) and (5) is assumed for the time evolution of the residual profit x ^a _t :

From equation (4), we obtain v _t = E _t [x ^a _t+1 ] - ωx ^a _t , and by substituting this into the RIM in equation (3) and rearranging, we obtain the OVM (Ohlson model) in equation (6) below.

Here, the OVM coefficients are expressed by the following equations (7) to (10).

These are estimated as ω, γ, and r using LIM or CAPM, and substituted into the above equations (7) to (10) to obtain the estimated intrinsic value ^V _t (here, hereinafter, the "^" next to an English letter is a hat above the English letter, and the same notation will be used hereinafter.) Alternatively, the partial regression coefficient is estimated using the following equation (11), which is a multiple regression equation for the observable P _t .

Then, an estimated value ^V _t of V _t is obtained by the following equation (12).

Furthermore, by examining the difference between ^V _t and P _t , it is possible to determine whether the asset price P _t is undervalued or overvalued.
In order to adjust for differences in scale due to company size, equation (11) may be divided by the previous month's stock price _Pt-1 , or the PBR (Price-Book-Value Ratio) divided by the net assets _Bt may be used as the objective variable. In this case, the following equation (13) is obtained.

In either case, it is preferable to use the stock analyst's forecast (analyst forecast) for E _t [x _t+1 ]. For example, in Japan, management's own forecasts are published in quarterly financial results briefs. This is available to anyone and is useful information that is not subordinate to analyst forecasts, so it is preferable to use management forecasts in this embodiment.

(Nonlinear modeling of OVM using machine learning)
The corporate valuation model in the above-described embodiment makes simple assumptions for theoretical derivation, prioritizing deductive interpretation over correspondence to more complex reality. However, because financial data is generally non-stationary, estimating the LIM coefficients in equations (4) and (5) is difficult. The time evolution dynamics of residual profits is not linear like LIM, but is likely to be nonlinear. This is because if residual profits are positive, the business expands, and if negative, the business contracts. Therefore, the persistence coefficient ω is considered to be a function of residual profits f( ^xat ₎ , rather than a constant.

In this embodiment, priority is given to correspondence with reality, and the OVM is expanded by introducing machine learning.
However, since the financial information used for corporate valuation changes only a few times a year, it is difficult to secure a sufficient amount of data for machine learning. Also, although the corporate valuation model assumes that each coefficient is constant, in reality, there is a possibility that it may change over time. Therefore, in this example, the use of old data is reduced.

Therefore, it is preferable to increase the amount of data by using panel data that includes not only the time series direction but also the cross-sectional direction.In addition, the more samples there are, the closer E[V _t ] = P _t - E[ε _t ] becomes to P _t , which improves the validity of the coefficient estimate in equation (11).
Here, the model coefficient θ is considered to differ for each individual company i. In particular, the cost of capital r may be estimated by the capital asset pricing model (CAPM) shown in the following equation (14). Note that when risk premiums other than market risk are taken into consideration, a multi-factor model including more risk factors may be used.

Here, _rf is the risk-free interest rate, _Et [ _rm,t+1 ] is the expected market return, and βi _,t is the market beta of company i. This market beta is a value that indicates sensitivity to market risk factors. In other words, even with conventional OVM, the partial regression coefficient θ (Equation (7) to Equation (10)) involving r differs for each company i, so coefficient estimation by cross-sectional regression is not possible. Therefore, in this embodiment, characteristics unique to company i are used as explanatory variables, and F, the dynamics common to all companies, is estimated by cross-sectional machine learning.
Therefore, equation (13) may be replaced with the following equation (15).

This F is estimated using machine learning such as a decision tree, as described below. This estimated value of F is hereafter referred to as "^F" (dynamics function). ^F estimated by this machine learning is used to calculate the value of the following equation (16).

Here, t represents the monthly rate, x _i,qt represents the net profit at the end of the previous quarter (actual value per share), d _i,yt represents the dividend at the end of the previous fiscal year (actual value per share), E _t [x _i,qt+1 ] represents the forecast net profit at the end of the current quarter (management forecast per share), and E _t [d _i,yt+1 ] represents the forecast dividend at the end of the current fiscal year (management forecast per share).
Furthermore, as variables related to the cost of capital r _i,t , market beta β _i,t and volatility σ _i,t were calculated from the monthly returns of company i over the most recent 36 months, and the expected market return E _t [r _m,t+1 ] was calculated using the average monthly returns of TOPIX over the most recent 36 months. Note that there is a time lag of up to two months until financial information after settlement is made public, and it takes further time for this to be incorporated into stock prices, so the stock price P _i,t in equation (15) uses the realized value three months later. On the other hand, since there is no time lag for β _i,t , σ _i,t , and E _t [r _m,t+1 ], the values for the same month as P _i,t were used.

For machine learning, XGBoost decision trees were used, with 1000 decision trees, a learning rate of 0.03, and a loss function of RMSE, with other hyperparameters set to default values. Because stock prices P _i,t fluctuate monthly, the learning period was the most recent 36 months (excluding the current month), and F was retrained every month. Financial information from three months prior was input into the estimated F to estimate the firm value ^V _i,t .
The above method was verified to work. The analysis period was from January 31, 2008 to June 30, 2021. The target company i was selected based on the following criteria:

-The company is listed on the first section of the Tokyo Stock Exchange.
- General business companies (excluding banks, securities, insurance, and other financial services).
-Net assets are always positive.
- Quarterly or full-year financial data is not missing for two or more consecutive periods.

If financial data was missing for only one period, it was supplemented with the value from the previous period. Financial and stock price data were obtained from Reinitiv Eikon (<URL="https://www.refinitiv.com/").

(result)
Figure 5 shows the correlation between the estimated intrinsic value ^V _i,t and the asset price P _i,t . The correlation roughly fits a straight line, with the deviations considered to be mispricing due to investor sentiment. Therefore, the deviation rate ξ _i,t was calculated using the following formula (17).

The results are shown in Figure 6. Investor sentiment (psychological bias) oscillated around 0, confirming steady center reversion. However, depending on the industry or the trends of the company itself, the market price P _i,t may constantly deviate from the intrinsic value. In this example, constant deviations are corrected by subtracting the moving average value for each company i. The calculation formula is shown in Equation (18) below. Note that improvements can also be made by applying dummy variables to machine learning.

Here, this corrected deviation rate is designated as ξ ^† _i,t . In this example, M = 36 (36 months), and overvalued and undervalued stocks are distinguished based on ξ ^† _i,t . Due to its similarity to PBR, in this example, ξ ^† _i,t is called the PIR (Price Intrinsic-Value Ratio) indicator.
The corrected ξ ^† _i,t is shown in Figure 7. The average value was almost the same as in Figure 6, but the standard deviation was reduced. In other words, it was possible to consistently correct undervalued stocks (ξ _i,t < 0) or overvalued stocks (ξ _i,t > 0).

Figure 8 shows the feature importance of explanatory variables. The importance of expected values (especially expected dividends, Et[d _i,yt+1 ]) is particularly notable. On the other hand, the importance of expected net income, Et[x _i,qt+1 ], is low, but this is likely due to overlapping information with expected dividends. Furthermore, the importance of the realized divides (x _i,qt , d _i,yt ) used in OVM was high, followed by the volatility related to capital cost (Sigma of return rates) (σ _i,t ), market beta (Market beta of CAPM) (β _i,t ), and expected market return (Expected net income) E _t [r _m,t+1 ].

(Verification of validity through portfolio management)
Asset prices P do not necessarily reflect intrinsic value V, and temporary mispricing may occur. As shown in Figure 6, in such cases, asset prices are expected to move in a way that corrects them to intrinsic value due to the efficiency of financial markets. For this reason, we verified the validity of the PIR index shown in equation (18) above through portfolio management that utilizes this correction process.
The portfolio management methodology is as follows:

- Machine learning retraining and PIR calculation are performed at the end of each month.
-Divide investment target stocks into 5 categories based on PIR.
- Short the first quintile (overvalued stocks) and go long the fifth quintile (undervalued stocks).
・All buying and selling ratios will be equal weighted.
-Rebuild your portfolio every month by repeating the above at the end of the following month.

Figure 9 shows a cumulative graph of the return r _p,t (cumulative return) earned by a portfolio consisting only of long (purchases) (long portfolio) or a portfolio consisting only of short (short sales) (short portfolio). In Figure 9, the straight line indicates the cumulative value of the 0.5% monthly management fee. As a result, the long portfolio exceeded the benchmark TOPIX, and it can be said that the value (undervalued) stock effect was extracted. On the other hand, the short portfolio, which allows the selling of growth stocks, tends to be more susceptible to trends over time.

Figure 10 shows the cumulative active returns of long portfolios by quantile. In Figure 10, long portfolios are constructed for each quantile, from the 5th quantile (cheapest 20%, 5th quartile) to the 1st quantile (highest valuation, 1st quartile), and the active returns (r _p,t - r _m,t ) excluding TOPIX returns are totaled. Profitability is arranged in quantile order, confirming the validity of stock evaluation using the PIR index.

Figure 11 shows the cumulative return of a long-short portfolio, where the top 20% (1st quantile) is short and the bottom 20% (5th quantile) is long, as a graph of the PIR index of equation (18).
"PIR of Eq. (18)" in FIG. 11 is a graph of the cumulative return of this example.
As a comparative example, a long-short portfolio based on the conventional PBR index is shown in the following equation (19). This is "PBR of Eq. (19)" in FIG. 11.

Furthermore, as a comparative example, similar to equation (18), an example in which the average value for M=36 (36 months) is removed to correct for persistent mispricing is shown in equation (20) below. This is the "PBR of Eq. (20)" in Figure 11.

Here, η _i,t is the PBR index, and η ^† _i,t is the corrected value.
As a result, the PIR index according to this example achieved a stably higher return than the PBR index and TOPIX, and the risk-return ratio (mean/std) was also significantly improved. In other words, it is believed that the PIR index was able to extract market mispricing as expected.

Next, to verify whether the mispricing extracted by the PIR index contains anomalies that cannot be explained by the risk premium, we performed a regression analysis of the returns r _p,t obtained by each portfolio using control variables as shown in equation (21) below.

where r _f,t is the risk-free interest rate, f ^Mkt-RF _t is the market factor return minus the risk-free interest rate, f ^SMB _t is the size factor return, and f ^HML _t is the value factor return, all of which were obtained from K. R. French's website (URL = "https://mba.tuck.dartmouth.edu/pages/faculty/ken.french/data_library.html (last accessed: October 1, 2021)">).
The coefficient test results are shown in Table 1 below.

In Table 1, a significance test was conducted using a three-factor model. The "*" and "**" next to the coefficients indicate p<0.05 and p<0.001, respectively.

As a result, we were able to confirm strong significance, particularly in the α of the long-short portfolio, and the existence of an anomaly that cannot be explained by the risk premium in equation (21). Naturally, there remains the possibility that premiums due to risk factors other than equation (21) may be included, but it is highly likely that we were able to extract mispricing due to investor psychology.

(summary)
By extending the Ohlson model in Non-Patent Document 1 using machine learning, we were able to evaluate the discrepancy between market price and intrinsic value. Furthermore, we confirmed that the returns obtained through portfolio management can be expressed in terms of alpha other than risk factors, and showed that this discrepancy may be a market mispricing.

It goes without saying that the configurations and operations of the above-described embodiments are merely examples and can be modified as appropriate without departing from the spirit of the present invention.

REFERENCE SIGNS LIST 1 Stock price evaluation device 10 Control unit 11 Image processing unit 12 Memory unit 13 Input unit 14 Display unit 15 Transmitting/receiving unit 100 Machine learning unit 110 Intrinsic value calculation unit 120 Value difference calculation unit 130 Asset management unit 200 Financial asset forecast data 210 Model data 220 Intrinsic value data 230 Monitoring data 240 Management data 300 Stock price evaluation program 310 Stock price monitoring program 320 Asset management program

Claims (11)

a machine learning unit that performs machine learning on the relationship between financial information of a plurality of companies over a plurality of periods and asset prices, and estimates a dynamics function common to the plurality of companies;
an intrinsic value calculation unit that calculates the intrinsic value of an asset using the dynamics function estimated by the machine learning unit;
a value difference calculation unit that calculates the difference between the intrinsic value calculated by the intrinsic value calculation unit and the asset price as a herd psychology ;
The value difference calculation unit draws a list of the crowd psychology based on the period and the type of issue.
A stock price evaluation device characterized by: 2. The stock price evaluation device according to claim 1, wherein the machine learning unit also uses, in the machine learning, performance forecasts by management or analysts indicated in financial statements. 3. The stock price evaluation device according to claim 1, wherein the dynamics function is an OVM model (Ohlson Valuation Model) extended to a nonlinear model. 4. The stock price evaluation device according to claim 1, wherein the value difference calculation unit detects abnormal stock prices based on the crowd psychology. 5. The stock price evaluation device according to claim 1, further comprising an asset management unit that identifies undervalued stocks and overvalued stocks based on the herd psychology calculated by the value difference calculation unit. The asset management department
6. The stock price evaluation device according to claim 5 , wherein the device manages a portfolio by going long on the undervalued stocks and short on the overvalued stocks. A stock price evaluation program executed by a stock price evaluation device,
Machine learning is performed on the relationship between financial information and asset prices of multiple companies over multiple periods, and a dynamics function common to the multiple companies is estimated;
The estimated dynamics function is used to calculate the intrinsic value of the asset.
Calculating the difference between the calculated intrinsic value and the asset price as herd psychology ;
The herd psychology is displayed based on the period and type of stock.
A stock price evaluation program characterized by: Identifying undervalued stocks and overvalued stocks based on the herd psychology calculated by the stock price evaluation program according to claim 7 ;
An asset management program characterized by managing a portfolio by going long on the identified undervalued stocks and short on the identified overvalued stocks. A stock price evaluation method executed by a stock price evaluation device, the stock price evaluation device
Machine learning is performed on the relationship between financial information and asset prices of multiple companies over multiple periods, and a dynamics function common to the multiple companies is estimated;
The estimated dynamics function is used to calculate the intrinsic value of the asset.
Calculating the difference between the calculated intrinsic value and the asset price as herd psychology ;
Illustrate the herd psychology based on time period and type of stock.
A stock price valuation method characterized by: A stock price monitoring method, comprising: monitoring abnormal stock prices based on the herd psychology calculated by the stock price evaluation method according to claim 9 . Identifying undervalued stocks and overvalued stocks based on the herd psychology calculated by the stock price evaluation method according to claim 9 ;
and managing a portfolio by going long on the identified undervalued stocks and short on the identified overvalued stocks.