JPH06139227A - Time series prediction device - Google Patents

Abstract

(57) [Summary] [Purpose] Time series forecasting that can predict market prices with high accuracy and stability, using both symbolic knowledge about economic time series and past economic time series patterns. Provide a device. [Structure] A plurality of neural networks 5, 6 and 7 for individually learning corresponding to a plurality of market types and predicting a corresponding market, respectively, and at least one general-purpose neural network learning for all market types A neural network group 3 including a network 8 is provided, and when the market type is inferred by the knowledge base inference system 1, the economic time series pattern is input to the neural network 5, 6 or 7 corresponding to the market type, The neural network is made to predict a future market price, and when the type of market price cannot be inferred by the knowledge base inference system 1, the economic time series pattern is inputted to the general-purpose neural network 8 so that the future price is inputted to the neural network. Neural network that predicts market price and predicts future market price The prediction result is output based on the output value of the work.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a time series prediction device for predicting future market prices by inputting symbolic knowledge about economic time series and past economic time series.

[0002]

2. Description of the Related Art When predicting future market prices by inputting symbolic knowledge about economic time series or past economic time series patterns, the following methods have been adopted in the prior art.

When the symbolic knowledge about the economic time series is used, the symbolic knowledge about the market price is derived by inputting the symbolic knowledge into the knowledge base inference system. For example, there is a market analysis expert system that points out the timing of bond trading based on knowledge about price movements of bonds [Reference 1, List of expert systems put into practical use in Japan, Nikkei AI separate volume 1992 winter issue].

When the past economic time series pattern is used, the market price is predicted by inputting the economic time series pattern into a neural network. For example, input time series patterns (for the past 1 year) such as yen-dollar exchange rate, producer price index, business trend leading index, etc. into the neural network, raise and lower the market price after 1-5 days, yield over the last 5 days, There is a market forecast system that predicts the peak bottom of the market [Reference 2, Nakata: Neuro-applied market forecast system that can support various market developments, see the Nikkei Intelligent System appendix 1992 summer issue].

[0005]

When the market price is inferred by the symbol knowledge and the knowledge base inference system, it is possible to derive the symbol knowledge related to the market price prediction. For example, it is possible to output the timing of buying and selling bonds, the determination of the up / down trend of the market price, the type of market price (“up market price”, “down market price”, “waiting market price”), etc. However, it is not possible to output numerical data such as yields and stock prices.

When an economic time series pattern is input to a neural network to predict a market price, not only symbol data such as a market type but also numerical data such as a yield can be output. However, in a neural network trained in correspondence with a certain market type, there is a problem that the prediction accuracy is lowered in the situation where other market types are used.
For example, the prediction accuracy of a neural network trained at a rising market is lowered when the market is lowered or in a state of entanglement [Reference 2 above].

The present invention has been made in view of the above,
The purpose is to use a symbolic knowledge about economic time series and past economic time series patterns to create a time series prediction device that can predict the market price with high accuracy and stability regardless of the market type. To provide.

[0008]

In order to achieve the above object, a time series prediction device of the present invention is a time series for predicting future market prices by inputting symbolic knowledge about economic time series and past economic time series patterns. A predicting device, when a plurality of market types are set, a knowledge-based reasoning system that infers the market type based on the symbolic knowledge, and the economic time series pattern as an input to each of the plurality of market types A group of neural networks having a plurality of neural networks that individually learn correspondingly to predict respective corresponding quotes and at least one general-purpose neural network that learns correspondingly to all quotes, and a quote based on the knowledge-based reasoning system When the type of the market price is inferred, the economic time series pattern is added to the neural network corresponding to the market type. Is input to the neural network to predict a future market price, and when the type of market price cannot be inferred by the knowledge-based inference system, the economic time series pattern is input to the general-purpose neural network to input the neural network. An input control unit for controlling the network to predict a future market price, and an output control unit for outputting a prediction result based on the output value of the neural network input with the economic time series pattern by the input control unit That is the summary.

[0009]

In the time-series prediction apparatus of the present invention, a plurality of neural networks for individually learning corresponding to a plurality of market types and predicting a corresponding market price respectively, and at least learning corresponding to all market prices. One general-purpose neural network is provided, and when the market type is inferred by the knowledge-based inference system, the economic time series pattern is input to the neural network corresponding to the market type to predict the future market price in the neural network. In addition, when the type of the market price cannot be inferred by the knowledge-based inference system, the economic time series pattern is input to the general-purpose neural network to cause the neural network to predict the future market price, and predict the future market price. Output the prediction result based on the output value of the neural network To have.

[0010]

Embodiments of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram showing the configuration of a time series prediction device according to an embodiment of the present invention. The time-series prediction device shown in the figure explains the bond futures market price prediction problem as an example, and sets three types of market prices, "up", "down", and "waiting".

In FIG. 1, 1 is a knowledge base inference system, 2 is an input control unit, 3 is a neural network group,
4 is an output control unit. Reference numerals 5 to 8 are neural networks, and reference numeral 5 is a neural network trained corresponding to the "raising market" (a neural network trained using an economic time series pattern when a market expert determines "raising market"). The same shall apply hereinafter), 6 is a neural network trained in response to a "down market", 7 is a neural network trained in response to a "waiting market", and 8 is a general-purpose trained for all markets. It is a neural network. Reference numeral 9 is a connection line for sending the inference result from the knowledge-based inference system 1 to the input control unit 2.
0 is a connection line for informing which output of the neural network is valid from the input control unit 2 to the output control unit 4.
Reference numeral 1 is a connection line for transmitting a signal from the input control unit 2 to the neural network group 3 (there are plural sets), and 12 is a connection line for transmitting a signal from the neural network group 3 to the output control unit 4 (there are a plurality of sets).

Next, the operation sequence of the time series prediction device of FIG. 1 will be described with reference to FIG.

From outside, as knowledge about economic time series,
When the symbol knowledge a “with political change” is input and the knowledge base inference system 1 has the symbol knowledge b with “if political change, then down market”, the time series prediction device of the present invention operates as follows. .

First, the symbol knowledge input from the outside and the symbol knowledge in the knowledge-based reasoning system are collated (in this case, the symbol knowledge a and the symbol knowledge b are matched and the inference of "downward quote"). The result will be obtained). The inference result of the knowledge-based inference system 1 is output to the input control unit 2 (step 110).

Secondly, the input control unit 2 inputs only the past economic time series pattern to the neural network 6 which has been trained in correspondence with the "down market". Further, the input control unit 2 notifies the output control unit 4 that the output of the neural network 6 is valid (step 120).

Thirdly, the neural network 6 executes a time series prediction process based on the input time series pattern (step 130).

Fourth, the output control section 4 outputs the output value of the neural network 6 as a prediction result (step 140).

The neural network 6 is a "down market"
Since it is a dedicated neural network trained corresponding to, the prediction accuracy is higher than the neural network 5 and the neural network 7 trained corresponding to other market prices, and the general-purpose neural network 8 that does not distinguish the market type. it is obvious.

[0020]

As described above, according to the present invention,
A knowledge-based inference system is provided, which is provided with a plurality of neural networks that individually learn corresponding to a plurality of market types and predict respective corresponding market prices and at least one general-purpose neural network that learns corresponding to all market prices. When the market type is inferred by the market type, the economic time series pattern is input to the neural network corresponding to the market type to cause the neural network to predict the future market rate, and the knowledge type inference system determines the market type. When it cannot be inferred, the economic time series pattern is input to the general-purpose neural network to cause the neural network to predict a future market price, and a prediction result based on the output value of the neural network that predicted this future market price. Since it is output, the new type according to the market type Can be predicted rate by Le network, it is possible to realize high prediction accuracy.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of a time series prediction device according to an embodiment of the present invention.

FIG. 2 is a flowchart showing an operation of the time series prediction device shown in FIG.

[Explanation of symbols]

 1 Knowledge-based inference system 2 Input control unit 3 Neural network group 4 Output control unit 5 Neural network (NN) for up market 6 Neural network for down market 7 Neural network for waiting market 8 General neural network

Claims (1)

[Claims] 1. A time series prediction device for predicting future market prices by inputting symbolic knowledge about economic time series and past economic time series patterns, and based on the symbolic knowledge when a plurality of market types are set. -Based inference system that infers the market type, and a plurality of neural networks that individually input the economic time series pattern and individually learn for each of the plurality of market types to predict each corresponding market type And a group of neural networks having at least one general-purpose neural network for learning corresponding to all quotes, and, when the quote type is inferred by the knowledge-based inference system, the neural network corresponding to the quote type becomes the economy. If you input a time series pattern and let the neural network predict future market prices, Also, when the type of market price cannot be inferred by the knowledge-based inference system, the input control for inputting the economic time series pattern to the general-purpose neural network and controlling the neural network to predict a future market price. And an output control unit that outputs a prediction result based on an output value of a neural network to which the economic time series pattern is input by the input control unit.