JPH07296057A - Method for monitoring financial market - Google Patents

Abstract

PURPOSE:To provide a function prompting attention by notifying a user of the difficulty of prediction when it is expected, by monitoring a financial market. CONSTITUTION:A trend extraction 102 takes out from economic time series data 101 its time series of the trend component. Trend prediction 103 predicts a future trend. A monitor 104 calculates the alienation between the result of trend prediction 103 and the economic time series data realizing value of original so as to continue monitoring how this alienation changes along with the passage of time. When the monitoring means detects the drastic change of alienation, a notifying means notifies 105 the user of it.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a decision support system for financial transactions.

[0002]

2. Description of the Related Art In Japanese Unexamined Patent Publication No. 5-135060, "Price formation simulation method for trading market"
Modeling the mechanism of price formation and the initial price of the product,
A method of simulating price formation with the number of investors as a parameter is shown. This simulation method has three types of investors: a physical investor, a futures investor, and an arbitrage investor, and each investor decides an order according to a trading behavior model based on the current prices of the real and futures. Upon receipt of this order, the price of the spot and futures is determined according to the pricing model. The above process of ordering and price determination is repeated.

According to this simulation method,
(1) It is possible to analyze the mechanism of price formation by simulating with a model that is as close as possible to the actual market. (2) When establishing a new trading market, it is necessary to check in advance whether the market system will function as intended. I can confirm
(3) It is said that it can be used as a training system for investment methods and investment strategies.

Japanese Patent Application No. 3-187254, "Information Processing Device Using Fractal Dimension", discloses a technique for predicting the accuracy of time series prediction using fractal dimension and generating an abnormal alarm signal. . Here, the complexity of time-series fluctuations is measured by a scale called fractal dimension, and the prediction accuracy can be known in advance because the prediction accuracy is lower as the value of the fractal dimension increases. Moreover, since the value of the fractal dimension changes with the passage of time, it is possible to know the change in the phase of the time series fluctuation.

[0005]

A price prediction function is one of the functions of a decision support system for financial transactions.
For example, stock price prediction. Although errors are inevitable in forecasting, the difficulty of forecasting varies greatly depending on the market situation, and the magnitude of forecasting errors is not always constant. Therefore, it is important to diagnose the current situation and know the difficulty of prediction.

It is an object of the present invention to provide a method for monitoring market conditions and notifying a user when it is predicted that it is difficult to make a prediction and calling attention to the user.

[0007]

The financial transaction market monitoring apparatus according to the present invention receives economic time series data as input. This device is a means for extracting a trend component that is a global fluctuation from the input time series, a means for predicting a future value of the trend component, and a means for monitoring the time change of this prediction result,
It has means for notifying the user of this change when the preset condition is met. It is also possible to have a neural network that inputs the time change amount of the prediction result and outputs the degree of change in the magnitude of the change.

[0008]

In the present invention, the economic time series data is regarded as consisting of the trend component and the irregular fluctuation component. The trend component mentioned here is a moving average of the original time series or a low frequency component. The trend extracting means extracts the time series of the trend component from the input economic time series data. The trend predicting means predicts a future value of the trend component using the extracted time series of the trend. The monitoring unit temporally changes the deviation between the predicted value of the trend component and the actual value of the original time series, changes with time of the deviation between the predicted value of the trend component and the actual value of the trend component, and the unit of the predicted value of the trend component. The change over time in the amount of change per hour is monitored as an indicator of the state of the trading market. If these indicators do not change significantly, the market is stable and the prediction results of economic time series data can be expected to have relatively high accuracy. However, when these indicators change drastically, it is considered that the market is unstable, that is, irregular fluctuations are large, and the reliability of the forecast results becomes low. When the monitoring means detects a drastic change in the index, the reporting means reports it to the user.

Further, the monitoring means has a neural network which inputs the above-mentioned index indicating the state of the trading market and outputs the degree of change of the time series fluctuation, and reports a signal according to the output of the neural network. Send to the means.

Further, the financial transaction market monitoring apparatus according to the present invention always receives the latest economic information and continues to monitor the time series. Then, when the way of changing the prediction result satisfies the abnormal condition, the user is immediately notified by a method according to the condition.

Further, the financial transaction market monitoring system according to the present invention has a function of monitoring the prediction error of the trend component, and if the error increases, the reliability of the prediction value is low and the user is warned. Function as a trend prediction method.

[0012]

EXAMPLE An example of the present invention will be described. This embodiment can be implemented in a general-purpose computer equipped with a storage device, a numerical operation device, an input / output device, and a workstation. The storage device includes a removable external storage device. The input / output device includes a display device, a keyboard, and a communication device. FIG. 1 shows the flow of processing.

The input is an economic time series 101 such as stock prices. For simplicity in the following description, it is assumed that the present apparatus inputs only one time series without inputting other economic factors. For example, when monitoring changes in the stock price of company A, only the past history of the stock price of company A is input. However, it is easy to extend this so that multiple time series are input. In the following explanation, the scalar quantity that represents the time series value at each time is replaced with a vector quantity that has the values at the same time of various economic factor time series as components, and if necessary, the coefficients can be replaced with vectors or matrices. Good. This is for example
This is a case of using not only the past history of the stock price of the company A but also the past history of economic factors such as stock prices and interest rates of other companies closely related to the same industry in monitoring the change in the stock price of the company A.

Generally, the economic time series data can be regarded as composed of a trend component which changes relatively gently and largely with the passage of time and an irregular fluctuation component.

[0015]

[Equation 1] (Economic time series) = (Time series of trend component) + (Time series of irregular fluctuation component) (Equation 1) The trend extraction means 102 extracts the trend component from the input time series.

One method of extracting the trend component is to find the moving average of the original time series and regard it as the trend component. Here, two types of moving averages are used. Equations 2 and 3 are expressions for obtaining them.

[0017]

[Equation 2]

[0018]

[Equation 3]

T (t) and TT (t) are the values of the moving average at time t, y (t) is the actual value at the time t of the original economic time series, and w is the time width for obtaining the moving average. So w>
It is 0. Note that t is a discrete value. In number 3, TT
Since the actual value after the time t in the original time series is also used to obtain (t), in reality, TT (t) cannot be obtained at the time t. On the other hand, in Equation 2, although T (t) can be obtained at the time t, there is a time lag with respect to the original time series {y}. FIG. 2 shows an example of the original time series {y} and the moving average {TT} of Equation 3. The solid line shows {y} and the broken line shows {TT}. Thus, when y (t) is obtained, {T
T} can only be obtained up to TT (tw).

Another method of extracting the trend component is to extract a low frequency component from the original time series using a frequency filter and regard this as the trend component.
FIG. 3 shows a process of extracting low frequency components using a frequency filter. First, in the cutout process 301, a time series having an appropriate length is cut out from the original time series. The first-stage graph 311 in FIG. 3 is a cut-out time series. Next, Fourier transform 302 is applied to this time series to obtain a spectrum. The second graph 312 in FIG. 3 is the obtained spectrum,
The horizontal axis is frequency. Next, the high frequency component cut 303 is performed. The frequency at the boundary between the high frequency component and the low frequency component will be referred to as the cutting frequency. Then, the high frequency component is removed by setting the absolute value of the component having a frequency higher than the cutting frequency to 0. The graph 313 in the third row of FIG. 3 is the spectrum from which high frequency components have been removed, and the points indicated by triangles are the cutting frequencies. Inverse Fourier transform 30 is applied to the remaining low frequency components.
By applying 4, a time series low frequency component is obtained. The graph 314 in the fourth row of FIG. 3 is the obtained low frequency component.

FIG. 4 is an example of the original time series and the low frequency components extracted by the frequency filter. The solid line shows the original time series, and the broken line shows the low frequency component. In this embodiment, the trend component is extracted using the moving average or the frequency filter as described above. However, there is no general definition of the trend component relating to an arbitrary time series, and it is up to the user to decide what is regarded as the trend component of the original time series.

In the trend predicting means 103, the time (t-
Realization value y (t-1) of time series {y} obtained in 1),
Based on y (t-2), y (t-3), ..., Values T0 (t), T0 (t after time t of the time series {T0} of the trend component
Predict 1), .... Note that T0 (t), T0 (t +
1), ... Predicted values are T0 * (t), T0 * (t + 1) ,.
Will be expressed as As a simple example, here time t-
Based on the realization values of the time series {y} obtained in 1, T0 (t) one time ahead is predicted. As a prediction method, a regression model shown in Formula 4 is used.

[0023]

## EQU00004 ## T0 * (t) = f [x (t-1), x (t-2), x (t-3), ...] (Equation 4) f [] is a prediction function. x (t-1), x (t-2),
x (t−3), ... Are realization values of the time series {x} until time (t−1). {X} is a time series obtained from the realized values of the original economic time series {y}.

As shown in FIG. 5, the trend predicting means is composed of three processes of a model determining section 501, an estimating section 502 and a predicting section 503. In the model determination unit 501, what kind of function is to be used for the prediction function f [] of Equation 4, {T0},
Decide what kind of time series {x} should be. These are determined by the user specifying in advance. Various time series obtained from the original economic time series {y} can be used for {x}. Also, the prediction function f
[] Can also use various linear and non-linear functions.

For example, the time series of the trend to be predicted {T
0} is a moving average time series {T} obtained by Equation 2, and the time series {x} is also a trend time series {T},
The prediction model is as shown in Equation 5.

[0026]

[Equation 5] T * (t) = f [T (t-1), T (t-2), T (t-3), ...] (Equation 5) As a specific example of the prediction model of Equation 5. , F [] are simple linear functions, and a prediction model as shown in Expression 6 is created.

[0027]

## EQU6 ## T * (t) = a (0) + a (1) .T (t-1) + ... + a (p) .T (tp) ... (Equation 6) This is a normal AR (Autoregressive) model, p is AR
The order of the model, a (0), a (1), ..., A (p), are the coefficients of the AR model. ARIMA instead of AR model
A (autoregressive integrated moving average) model or the like can also be used.

Further, a non-linear prediction function can be realized by using a multi-layer neural network having three or more layers as shown in FIG. In this case, T (t
-1), ..., T (t-pp) is an input and T * (t) is an output, which is a multilayer neural network. However, p
p is the number of input nodes 603 of the neural network.

Time series {TT} of the moving average obtained by Equation 3
Is regarded as the time series of the trend to be predicted, and the time series {x}
Assuming that the same time series is {TT}, as shown in Equation 3, at time t−1, time series after time t−w {TT}
Since the realization value of is not obtained, the prediction model is as shown in Expression 7.

[0030]

TT * (t) = f [TT (t-w-1), TT (t-w-2), TT (t-w-3), ...]] (Equation 7) In this case, As shown in FIG. 7, the latest value TT (t-1),
.., TT (tw) cannot be used to calculate TT * (t). In FIG. 7, the solid line shows the actual value of the original time series {y}, and the broken line shows the moving average {T}.
Is the realization value of T}. Also, the prediction model corresponding to equation 6 is as in equation 8.

[0031]

TT * (t) = b (0) + b (1) .TT (tw-1) + ... + b (q) .TT (tw-q) (Equation 8) where , W is the time width of the moving average in Equation 3, b
(0), b (1), ..., B (q) are coefficients of the prediction function. Also, a multilayer neural network can be used for the prediction function. In this case, TT (tw-1), ...,
The multi-layer neural network has TT (tw-qq) as input and TT * (t) as output. However, qq is the number of input nodes of the neural network.

In the time series {x} in the prediction function of equation 4,
There is also a method of directly using the original economic time series {y} instead of the trend time series. In this case, the prediction model is as shown in Equations 9 and 10.

[0033]

[Equation 9] T * (t) = f [y (t-1), y (t-2), y (t-3), ...] ... (Equation 9)

[0034]

TT * (t) = f [y (t-1), y (t-2), y (t-3), ...] (Equation 10) The prediction models of Equations 5 and 7 are Both are time series {T},
Using the past realization value of {TT}, the future prediction value of the same time series is obtained. On the other hand, the prediction models of Expressions 9 and 10 have a difference in that the past realization values of the time series {y} are used to obtain future prediction values of other time series {T} and {TT}. Equation 9 is {T} from the original economic time series {y}, and Equation 10
Is a model for predicting {TT} from {y}.
Further, when a linear function is used as the prediction function similarly to the equations 6 and 8, the equations 9 and 10 become the equations 11 and 12, respectively.

[0035]

[Equation 11] T * (t) = c (0) + c (1) .y (t-1) + ... + c (r) .y (tr) ... (Equation 11) c (0), c (1), ..., C (r) are the coefficients of the prediction function.

[0036]

TT * (t) = d (0) + d (1) .y (t-1) + ... + d (s) .y (ts) ... (Equation 12) d (0), d (1), ..., D (s) are coefficients of the prediction function. Instead of a linear function as a prediction function, a non-linear function by a multi-layered neural network can also be used.

The estimation unit 502 estimates the parameters of the prediction function determined by the model determination unit 501. The parameters mentioned here are the coefficients of each term on the right side in the case of the linear functions of the equations 6, 8, 11, and 12, and are the coupling coefficients between the calculation units in the case of the multilayer neural network of FIG. . When predicting the value of the trend component at time t, the estimated value of the parameter is calculated using the realized value of the original economic time series {y} obtained by time (t-1). Of the realized values of the original economic time series {y}, the realized value time series of length L immediately before time t, that is, the time (t-
It is assumed that the realization value time series from L) to time (t-1) is used for parameter estimation. The length L is preset by the user. The specific estimation method differs depending on the prediction model, and there are some estimation methods depending on the prediction model. In the case of linear functions such as Equation 6, Equation 8, Equation 11, and Equation 12, the least squares method is typical. In the least squares method, a parameter that minimizes the sum of squares of prediction errors is obtained. Taking the model for predicting TT * (t) from {TT} shown in Expression 8 as an example, the prediction error is expressed by Expression 13.

[0038]

Err (i) = TT (i) −TT * (i) = TT (i) − {b (0) + b (1) · TT (i−w−1) + ... + b (q) TT (i−w−q)} (Equation 13) where err (i) is the prediction error at time i. The sum of squares of the prediction error is represented by ERROR in Expression 14.

[0039]

[Equation 14]

However, t0 = t-L.

FIG. 8 shows the time relationship between t, w, q and t0. In FIG. 8, the solid line shows the actual value of the original time series {y}, and the broken line shows the moving average {T}.
Is the realization value of T}. Then, b (0), b (1), ..., B (q) that minimizes ERROR in Equation 14 are obtained.

The error back-propagation method is well known as an estimation method for a multilayer neural network. When the error back-propagation method is used, the coupling coefficient 604 of the neural network is calculated by the steepest descent method using the ERROR of Expression 14 as the objective function.

A prediction function can be obtained by applying the estimated amount thus obtained to each parameter.

The predicting unit 503 applies a time series realization value to the prediction function obtained by the estimating unit, and calculates the predicted value T0 * (t) of the trend component.

In the trend predicting means, the above three processes of model determination, estimation and prediction are performed according to the following procedure.

[Processing of trend predicting means] Step 1: Model determination Step 2: Set length L of time series used for parameter estimation Step 3: Set prediction start time t Step 4: Time t-L to t Parameter estimation is performed using a time series of realized values up to -1 Step 5: Calculate predicted value T0 * (t) Step 6: Set t ← t + 1 and return to Step 4 Forecast (Step 5)
By continuing the processing of time t, t + 1, t + 2, ..., Time series of predicted values {T0 * (t), T0 * (t + 1),
T0 * (t + 2), ...}.

The monitoring means inputs the original economic time series {y}, the time series {T0} of the realization value of the trend component, and the time series {T0 *} of the predicted value of the trend component, and the magnitude of the fluctuation of the time series. Detect changes in. The indicators that show the state of the trading market include
Use the following:

(1) Amount of change in trend realization value (2) Amount of change in trend prediction value (3) Amount of change in difference between trend realization value and prediction value (4) Trend realization value and original time series Change amount of difference in realization value of (5) Change amount of difference between predicted value of trend and realization value of original time series (1) is the amount of change ΔT0 (t) in realization value of trend component It is regarded as an abnormal fluctuation. ΔT0
(t) is expressed by Equation 15.

[0049]

ΔT0 (t) = T0 (t) −T0 (t−1) (Equation 15) where T0 (t) and T0 (t−1) are times t and t, respectively.
It is the realized value of the trend component at (t-1). Then, when ΔT0 (t) satisfies a preset abnormal condition, the monitoring means sends a signal to the notification means.

(2) is the amount of change Δ in the predicted value of the trend component
The abnormality of T0 * (t) is regarded as a time-series variation abnormality. ΔT0 * (t) is expressed by Expression 16.

[0051]

ΔT0 * (t) = T0 * (t) −T0 * (t−1) (Equation 16) Here, T0 * (t) and T0 * (t−1) are times t and t, respectively. It is the predicted value of the trend component at t-1. Then, when ΔT0 * (t) satisfies a preset abnormal condition, the monitoring means sends a signal to the notification means.

In (3), the abnormality of the change amount ΔE0 (t) of the difference between the realization value of the trend component and the predicted value is regarded as the abnormality of the time series fluctuation. ΔE0 (t) is expressed by Equation 17.

[0053]

ΔE0 (t) = E0 (t) −E0 (t−1) = {T0 * (t) −T0 (t)} − {T0 * (t−1) −T0 (t−1)} (Equation 17) Here, T0 (t) and T0 (t-1) are times t and t, respectively.
Realization value of the trend component at (t-1), T0 * (t),
T0 * (t-1) is the predicted value of the trend component at times t and (t-1), respectively. Then, when ΔE0 (t) satisfies the preset abnormal condition, the monitoring means sends a signal to the notification means.

In (4), the abnormality in the variation ΔD0 (t) of the difference between the realization value of the trend component and the original economic time series is regarded as the abnormality in the time series fluctuation. ΔD0 (t) is expressed by Equation 18.

[0055]

ΔD0 (t) = D0 (t) −D0 (t−1) = {T0 (t) −y (t)} − {T0 (t−1) −y (t−1)} ( (Equation 18) Here, T0 (t) and T0 (t-1) are times t and t, respectively.
Realized value of the trend component at (t-1), y (t), y
(t-1) are the realized values of the original economic time series at the times t and (t-1), respectively. Then, when ΔD0 (t) satisfies the preset abnormal condition, the monitoring means sends a signal to the notification means.

In (5), the abnormality of the change amount ΔD * 0 (t) of the difference between the predicted value of the trend component and the original economic time series is regarded as the abnormality of the time series fluctuation. ΔD0 * (t) is expressed by Equation 19.

[0057]

ΔD0 * (t) = D0 * (t) −D0 * (t−1) = {T0 * (t) −y (t)} − {T0 * (t−1) −y (t−) 1)} (Equation 19) Here, T0 * (t) and T0 * (t-1) are prediction values of the trend component at time t and (t-1), y (t), and
y (t-1) are the realization values of the original economic time series at times t and (t-1), respectively. Then, when ΔD0 * (t) satisfies the preset abnormal condition, the monitoring means sends a signal to the notification means.

The amount of change from (1) to (5) above is generally expressed as Δ (t). The weighted average W [Δ] (t) of the amount of change Δ (t) during a certain period, which is represented by Equation 20, is not the value of the amount of change Δ (t) at only one time point, and is the magnitude of the time-series variation. Alternatively, a signal may be sent to the notification means when W [Δ] (t) satisfies a preset abnormal condition.

[0059]

[Equation 20]

Here, w (i) is a preset weight, and u is the length of time for averaging. The weight becomes w (i) as it goes back in the past, that is, as i increases.
To be small. Or simply all i
With respect to w (i) = 1 / u.

Furthermore, the change amount of the change amount Δ (t), that is,

[0062]

[Expression 21] ΔΔ (t) = Δ (t) −Δ (t−1) (Expression 21) is used as an index, and when ΔΔ (t) satisfies a preset abnormal condition, the monitoring means is a reporting means. Alternatively, a method of sending a signal to

There are several abnormal conditions, that is, conditions under which a signal is sent from the monitoring means to the notification means. The simplest condition is that the index being monitored exceeds or falls below a preset threshold. There is also a condition that the index continues to be above or below the threshold for a certain period of time. There is not only one condition,
It is also possible to combine the conditions relating to the indexes of (1) to (5) and use the logical product or logical sum of them as the condition. Prepare several types of signals to be sent from the monitoring means to the reporting means,
It is also possible to select the type of signal according to how the trading market changes. For example, the degree of importance according to the value of each index is sent from the monitoring means to the notification means.

Also, input an index indicating the state of the trading market,
A multi-layered neural network that outputs the degree of change in the trading market situation may be used. This is shown in FIG.
V in FIG. 9 is an output value of the neural network, which is close to 1 when the change in the phase of the trading market is large, and is close to 0 when it is not so. Do.

The notification means notifies the user that the signal has been received from the monitoring means. If the signal has a degree of importance, it can be notified by a method according to the degree of importance. For example, if the degree of importance is high, even if the user is doing other work, it will interrupt and report. If the degree of importance is low, it will be reported only when there is an inquiry from the user. To do.

[0066]

According to the present invention, the user can know the change of the situation of the trading market and can be useful for decision making. Regarding time-series prediction, the user can know when it enters a phase where high prediction accuracy cannot be expected, so avoid the situation of making inappropriate decisions by using prediction results with low accuracy. You can

[Brief description of drawings]

FIG. 1 is a flowchart of processing performed in an embodiment of the present invention.

FIG. 2 is a graph of an economic time series and its moving average.

FIG. 3 is a flowchart of processing of a frequency filter.

FIG. 4 is a graph of an economic time series and its low frequency components.

FIG. 5 is a flowchart of processing performed inside the trend predicting unit.

FIG. 6 is an explanatory diagram of a multilayer neural network that realizes a prediction function.

FIG. 7 is an explanatory diagram showing a time relationship between data used for time series prediction of a moving average.

FIG. 8 is an explanatory diagram showing a time relationship between data used for parameter estimation of a prediction function.

FIG. 9 is an explanatory diagram of a multilayer neural network used as a monitoring unit.

[Explanation of symbols]

101 ... Economic time series, 102 ... Trend extraction means, 10
3 ... Trend predicting means, 104 ... Monitoring means, 105 ... Reporting means.

Claims (8)

[Claims] Claims: 1. Inputting economic time series data, extracting trend components from the input time series, predicting future values of the trend, and the time of deviation between this predicted value and the actual value of the input time series. A method for monitoring a financial transaction market, characterized by detecting a change in the magnitude of fluctuation by monitoring the change and notifying the user of the change. 2. The financial transaction market monitoring method according to claim 1, wherein a change in the magnitude of the fluctuation is detected by monitoring the time change of the deviation between the predicted value of the trend and the realized value of the trend. 3. The financial transaction market monitoring method according to claim 1, wherein a change in the magnitude of fluctuation is detected by monitoring the amount of change in the predicted value of the trend per unit time. 4. The financial transaction market monitoring method according to claim 1, wherein a moving average of the input time series is calculated and used as a trend component of the original time series. 5. The financial transaction market monitoring method according to claim 1, wherein a low frequency component is extracted from the input time series by using a frequency filter and is used as a trend component of the original time series. 6. The time variation of the deviation between the predicted value of the trend and the actual value of the original input time series, or the temporal variation of the deviation of the predicted value of the trend and the actual value of the trend, or It has a neural network that inputs the amount of change in the predicted value of the trend per unit time and outputs the magnitude of the time series fluctuation, and the user can change the magnitude of the time series fluctuation according to the output of the neural network. Financial transaction market monitoring method to notify. 7. The method according to claim 1, wherein the economic information is always automatically received, the economic time series is continuously monitored, and when a change in the magnitude of the time series fluctuation is detected, various types are detected according to the magnitude of the change. The financial transaction market monitoring method for notifying users by the method of. 8. A method of inspecting a time change of a deviation between an input realization value of a time series and a predicted value of a trend component of the time series, and notifying a user of that when a prediction error increases. Characteristic time series prediction method.