Exchange rate profit and loss prediction model training and exchange rate profit and loss prediction method

Abstract

The invention relates to the technical field of data processing, and particularly provides an exchange rate profit and loss prediction model training method and an exchange rate profit and loss prediction method. According to the exchange rate profit and loss prediction model training method, exchange rate index data is expanded based on a generative adversarial network; inputting the expanded exchange rate index data into a to-be-trained exchange rate profit and loss prediction model to obtain an output exchange rate profit and loss prediction result; and according to the exchange rate profit and loss prediction result and the exchange rate profit and loss data, adjusting internal parameters of the to-be-trained exchange rate profit and loss prediction model to obtain a trained exchange rate profit and loss prediction model. According to the exchange rate profit and loss prediction model training method, the obtained exchange rate index data related to the target project is expanded based on the generative adversarial network, so that more diversified model training data can be obtained, and the accuracy of the obtained fluctuation factors influencing the exchange rate and the accuracy of the exchange rate profit and loss prediction are improved; and the possibility of overfitting of the obtained exchange rate profit and loss prediction model can be reduced.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to exchange rate profit and loss prediction model training and exchange rate profit and loss prediction method. Background Art

[0002] With the acceleration of globalization, more and more companies are involved in overseas engineering projects. However, the economic environment faced by overseas engineering projects is complex and changeable, especially exchange rate fluctuations, which will have a direct impact on the profits and losses of the projects.

[0003] When predicting the exchange rate profit and loss of a project, the existing technology is usually based on traditional time series analysis or simple machine learning models. However, due to the complex and changing economic environment and limited historical data, the existing models cannot accurately obtain the factors that affect exchange rate fluctuations, and thus the accuracy of exchange rate profit and loss prediction is poor. Summary of the invention

[0004] In view of this, the purpose of the embodiments of the present application is to provide an exchange rate profit and loss prediction model training and an exchange rate profit and loss prediction method, so as to solve the technical problem that the existing exchange rate profit and loss prediction model cannot accurately obtain the factors affecting the exchange rate fluctuation and the accuracy of the exchange rate profit and loss prediction is poor.

[0005] In a first aspect, an embodiment of the present application provides a method for training an exchange rate profit and loss prediction model, the method comprising:

[0006] Based on the generative adversarial network, the acquired exchange rate indicator data related to the target project is expanded;

[0007] Inputting the expanded exchange rate index data into the exchange rate profit and loss prediction model to be trained, and obtaining the exchange rate profit and loss prediction result output by the exchange rate profit and loss prediction model to be trained;

[0008] According to the exchange rate profit and loss prediction result and the exchange rate profit and loss data corresponding to the expanded exchange rate indicator data, the internal parameters of the exchange rate profit and loss prediction model to be trained are adjusted to obtain a trained exchange rate profit and loss prediction model.

[0009] In the above implementation process, the exchange rate profit and loss prediction model training method expands the exchange rate index data related to the target project based on the generative adversarial network; inputs the expanded exchange rate index data into the exchange rate profit and loss prediction model to be trained, and obtains the exchange rate profit and loss prediction result output by the exchange rate profit and loss prediction model to be trained; according to the exchange rate profit and loss prediction result and the exchange rate profit and loss data corresponding to the expanded exchange rate index data, adjusts the internal parameters of the exchange rate profit and loss prediction model to be trained, and obtains the trained exchange rate profit and loss prediction model. The exchange rate profit and loss prediction model training method expands the exchange rate index data related to the target project through the generative adversarial network, and can obtain more diversified model training data, improve the accuracy of the acquired exchange rate fluctuation factors, and can also reduce the possibility of overfitting of the obtained exchange rate profit and loss prediction model, thereby improving the accuracy of the exchange rate profit and loss prediction. It solves the technical problem that the existing exchange rate profit and loss prediction model cannot accurately obtain the factors affecting the exchange rate fluctuation and has poor accuracy in the exchange rate profit and loss prediction.

[0010] Optionally, in an embodiment of the present application, the exchange rate profit and loss prediction model to be trained includes: a feature extraction module and a profit and loss prediction module; the inputting of the expanded exchange rate index data into the exchange rate profit and loss prediction model to be trained to obtain the exchange rate profit and loss prediction result output by the exchange rate profit and loss prediction model to be trained includes: inputting the expanded exchange rate index data into the exchange rate profit and loss prediction model to be trained, extracting indicator feature data in the expanded exchange rate index data based on the feature extraction module; and determining the exchange rate profit and loss prediction result corresponding to the expanded exchange rate index data based on the indicator feature data by the profit and loss prediction module.

[0011] In the above implementation process, the feature extraction module can extract the indicator feature data related to the exchange rate profit and loss prediction in the expanded exchange rate indicator data (that is, the factors affecting the exchange rate fluctuation), and determine the exchange rate profit and loss prediction result based on the indicator feature data based on the profit and loss prediction module.

[0012] Optionally, in an embodiment of the present application, the training process of the profit and loss prediction module includes: converting the indicator feature data into a quantum state to obtain quantum feature data; adjusting the entanglement and phase between the quantum feature data based on a quantum logic gate; wherein, the parameter configuration of the quantum logic gate is determined based on the current internal parameters of the profit and loss prediction module; measuring the adjusted quantum feature data, and determining the exchange rate profit and loss prediction result corresponding to the expanded exchange rate indicator data based on the measurement result; adjusting the internal parameters of the profit and loss prediction module based on the exchange rate profit and loss prediction result and the exchange rate profit and loss data corresponding to the expanded exchange rate indicator data until the training conditions of the profit and loss prediction module are met.

[0013] In the above implementation process, the indicator feature data is converted into quantum state, and the training process of the profit and loss prediction model is completed based on quantum computing, and the exchange rate profit and loss prediction result corresponding to the expanded exchange rate indicator data is determined through quantum measurement. Quantum measurement can effectively utilize the superposition and entanglement characteristics of quantum states, significantly improve the processing ability of the profit and loss prediction module for high-dimensional feature space, and the accuracy and controllability of the profit and loss prediction module during the training process, so as to further improve the accuracy of exchange rate profit and loss prediction.

[0014] Optionally, in an embodiment of the present application, the training process of the feature extraction module includes: generating an initialization population; wherein the initialization population includes multiple individuals, each of the individuals corresponds to a set of module parameters of the feature extraction module; according to the module parameters, calculating the feature loss value corresponding to each of the individuals; according to the feature loss value, selecting candidate individuals that meet the candidate conditions from the initialization population; based on the candidate individuals, new individuals are generated through crossover operations and mutation operations until the training conditions of the feature extraction module are met; wherein, during the training process of the feature extraction module, the probability of the crossover operation and / or the mutation operation is adjusted according to the changing trend of the feature loss value.

[0015] In the above implementation process, the probability of crossover operation and / or mutation operation is adjusted through the changing trend of feature loss value, so that the training process of feature extraction module can more flexibly adapt to different exchange rate indicator data distribution conditions, thereby improving the generalization ability and training efficiency of feature extraction module.

[0016] Optionally, in an embodiment of the present application, the generative adversarial network includes a generator and a discriminator; the training process of the generative adversarial network includes: based on the generator receiving a noise signal, and based on the noise signal, the exchange rate index data and the exchange rate profit and loss data corresponding to the exchange rate index data, generating simulated training data; based on the optimal transmission theory, the simulated training data and the discriminant function of the discriminator, determining the generation loss function value of the generator; based on the optimal transmission theory, the simulated training data, the exchange rate index data and the exchange rate profit and loss data corresponding to the exchange rate index data, determining the discriminant loss function value of the discriminator; adjusting the internal parameters of the generator based on the generation loss function value, and adjusting the internal parameters of the discriminator based on the discriminant loss function value until the training conditions of the generative adversarial network are met.

[0017] Optionally, in an embodiment of the present application, the updating of the internal parameters of the generator based on the generation loss function value includes: updating the internal parameters of the generator based on the generation loss function value obtained in the current training process, the generation loss function value obtained in the previous N training processes, and the internal parameters of the generator after the previous N training updates; wherein N is a positive integer.

[0018] In the above implementation process, in the process of updating the internal parameters of the generator, the generation loss function value obtained in the current training process, the generation loss function value obtained in the previous N training processes and the internal parameters of the generator after the previous N training updates are comprehensively considered; that is, based on the multi-step backtracking update method, the internal parameters of the generator are updated, which improves the effectiveness and stability of the generator parameter update.

[0019] Optionally, in an embodiment of the present application, the generation loss function value includes a transmission loss value and an entropy regularization term; the generation loss function value of the generator is determined based on the optimal transmission theory, the simulated training data, and the discriminant function of the discriminator, including: determining the transmission loss value based on the optimal transmission theory, the simulated training data, and the discriminant function of the discriminator; determining the entropy regularization term based on the entropy regularization term of the simulated training data.

[0020] In the above implementation process, since the generated loss function value includes the transmission loss value and the entropy regularization term, the entropy regularization term can encourage the generator to explore the characteristics of minority groups and improve the feature coverage and diversity of the generated simulated training data.

[0021] Optionally, in an embodiment of the present application, before the exchange rate profit and loss prediction result corresponding to the expanded exchange rate indicator data is determined based on the profit and loss prediction module according to the indicator characteristic data, the method also includes: performing dimensionality reduction processing on the indicator characteristic data based on a feature dimensionality reduction module to obtain dimensionality reduction feature data; the exchange rate profit and loss prediction result corresponding to the expanded exchange rate indicator data is determined based on the profit and loss prediction module according to the indicator characteristic data, including: determining the exchange rate profit and loss prediction result corresponding to the expanded exchange rate indicator data based on the profit and loss prediction module according to the dimensionality reduction feature data.

[0022] In the above implementation process, the feature dimension reduction module is used to reduce the dimension of the indicator feature data, which helps to remove redundant features and reduce the correlation between features, thereby improving the generalization ability and prediction accuracy of the model. It can also reduce the complexity of feature data, speed up the model training speed and exchange rate profit and loss prediction speed.

[0023] Optionally, in an embodiment of the present application, the dimensionality reduction feature module includes an encoder and a decoder; the training process of the dimensionality reduction feature module includes: based on the encoder, mapping the indicator feature data to a dimensionality reduction feature space of a preset dimension to obtain dimensional feature data; based on the decoder, reconstructing the dimensional feature data back to the original feature space to obtain reconstructed feature data; calculating the dimensionality reduction loss function value of the dimensionality reduction feature module based on the reconstructed feature data and the indicator feature data; and adjusting the internal parameters of the dimensionality reduction feature module based on the dimensionality reduction loss function value until the training conditions of the dimensionality reduction feature module are met.

[0024] In the above implementation process, the internal parameters of the dimension reduction feature module are adjusted by calculating the dimensionality reduction loss function value based on the reconstructed feature data and the indicator feature data to strengthen the important features and weaken the secondary features. This enables the dimension reduction feature module to retain the necessary feature data while realizing the dimensionality reduction processing of the indicator feature data; thereby ensuring the accuracy of the model's exchange rate profit and loss prediction while accelerating the model's training speed and exchange rate profit and loss prediction efficiency.

[0025] In a second aspect, an embodiment of the present application provides a method for predicting exchange rate gains and losses, the method comprising:

[0026] Obtain exchange rate indicator data related to the target project;

[0027] The exchange rate indicator data is input into the exchange rate profit and loss prediction model to obtain the exchange rate profit and loss prediction result output by the model; wherein the exchange rate profit and loss prediction model is configured to be obtained by training based on the exchange rate profit and loss prediction model training method as described in any one of the first aspects.

[0028] In a third aspect, an embodiment of the present application provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the method described in any one of the first aspect or the second aspect above.

[0029] In a fourth aspect, an embodiment of the present application further provides an electronic device; the electronic device includes:

[0030] Memory;

[0031] processor;

[0032] The memory stores a computer program executable by the processor, and when the computer program is executed by the processor, the method described in any one of the first aspect or the second aspect is performed.

[0033] In a fifth aspect, an embodiment of the present application further provides a computer-readable storage medium, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described in any one of the first aspects is executed.

[0034] The beneficial effects of the present application include at least: the exchange rate profit and loss prediction model training method expands the exchange rate index data related to the target project based on the generative adversarial network; inputs the expanded exchange rate index data into the exchange rate profit and loss prediction model to be trained, and obtains the exchange rate profit and loss prediction result output by the exchange rate profit and loss prediction model to be trained; according to the exchange rate profit and loss prediction result and the exchange rate profit and loss data corresponding to the expanded exchange rate index data, adjusts the internal parameters of the exchange rate profit and loss prediction model to be trained, and obtains the trained exchange rate profit and loss prediction model. The exchange rate profit and loss prediction model training method expands the exchange rate index data related to the target project through the generative adversarial network, can obtain more diversified model training data, improve the accuracy of the acquired exchange rate fluctuation factors, and can also reduce the possibility of overfitting of the obtained exchange rate profit and loss prediction model, thereby improving the accuracy of the exchange rate profit and loss prediction. It solves the technical problem that the existing exchange rate profit and loss prediction model cannot accurately obtain the factors affecting the exchange rate fluctuation and has poor accuracy in the exchange rate profit and loss prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments of the present application will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0036] Figure 1 A flow chart of a method for training an exchange rate profit and loss prediction model provided in an embodiment of the present application;

[0037] Figure 2 A schematic diagram of a training process of a profit and loss prediction module provided in an embodiment of the present application;

[0038] Figure 3 A schematic diagram of a training process of a feature extraction module provided in an embodiment of the present application;

[0039] Figure 4 A schematic diagram of a training process of a generative adversarial network provided in an embodiment of the present application;

[0040] Figure 5 A schematic diagram of a training process of a dimensionality reduction feature module provided in an embodiment of the present application;

[0041] Figure 6 A flow chart of a method for forecasting exchange rate gains and losses provided in an embodiment of the present application. DETAILED DESCRIPTION

[0042] The following embodiments of the technical solution of the present application are described in detail in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present application, and are therefore only used as examples, and cannot be used to limit the scope of protection of the present application.

[0043] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by technicians in the technical field to which this application belongs; the terms used herein are only for the purpose of describing specific embodiments and are not intended to limit this application.

[0044] In the description of the embodiments of the present application, the technical terms "first", "second", etc. are only used to distinguish different objects, and cannot be understood as indicating or implying relative importance or implicitly indicating the number, specific order or primary and secondary relationship of the indicated technical features. In the description of the embodiments of the present application, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined.

[0045] See also Figure 1 The flow chart of a method for training a model for predicting exchange rate gains and losses is shown in an embodiment of the present application. The method for training a model for predicting exchange rate gains and losses may include the following steps:

[0046] S101, expanding the acquired exchange rate index data related to the target project based on a generative adversarial network;

[0047] S102, inputting the expanded exchange rate index data into the exchange rate profit and loss prediction model to be trained, and obtaining the exchange rate profit and loss prediction result output by the exchange rate profit and loss prediction model to be trained;

[0048] S103. According to the exchange rate profit and loss prediction result and the exchange rate profit and loss data corresponding to the expanded exchange rate indicator data, internal parameters of the exchange rate profit and loss prediction model to be trained are adjusted to obtain a trained exchange rate profit and loss prediction model.

[0049] Wherein, in step S101, the generative adversarial network may include a generator and a discriminator; based on the generator, the exchange rate index expansion data is generated, and based on the discriminator, the exchange rate index data is distinguished from the exchange rate index expansion data generated by the generator. The target project may be a project involving exchange rate settlement. The exchange rate index data may include: foreign exchange rate fluctuation index, crude oil price index, political stability index of the country where the project is located, economic stability index of the country where the project is located (for example, consumer price index CPI and producer price index PPI), uncertainty index of trade policy, economic growth rate of the country where the project is located, engineering material cost index, international freight cost, completion percentage of engineering projects, stability index of the labor market, environmental risk assessment index, etc. Simulation index data for simulating exchange rate index data may be generated based on the generative adversarial network, and the expansion of the exchange rate index data may be realized based on the generated simulation index data.

[0050] Wherein, in step S102, the exchange rate profit and loss prediction model to be trained can be implemented by a classifier (e.g., a decision tree, a support vector machine, or a logistic regression model, etc.) and / or a neural network model. The exchange rate profit and loss prediction results can include profit prediction results and loss prediction results, and can also include multiple profit level prediction results and multiple loss level prediction results.

[0051] Among them, in step S103, the exchange rate profit and loss data corresponding to the expanded exchange rate indicator data may include profit and loss, and may also include multiple profit levels and loss levels; the exchange rate profit and loss data may be adjusted according to the actual project forecasting requirements. The above exchange rate profit and loss data may be obtained by marking the expanded exchange rate indicator data with profit and loss. The specific setting method of the exchange rate profit and loss prediction result may be adjusted according to the marked exchange rate profit and loss data. The exchange rate profit and loss prediction result output by the exchange rate profit and loss prediction model to be trained may be obtained, and the internal parameters of the model may be adjusted according to the difference between the exchange rate profit and loss prediction result and the "exchange rate profit and loss data corresponding to the expanded exchange rate indicator data" to obtain a trained exchange rate profit and loss prediction model.

[0052] It can be seen that the exchange rate profit and loss prediction model training method can obtain more diversified model training data by expanding the exchange rate indicator data related to the target project through the generative adversarial network, improve the accuracy of the acquired factors affecting exchange rate fluctuations, and reduce the possibility of overfitting of the acquired exchange rate profit and loss prediction model, thereby improving the accuracy of exchange rate profit and loss prediction. The technical problem that the existing exchange rate profit and loss prediction model cannot accurately obtain the factors affecting exchange rate fluctuations and has poor accuracy in exchange rate profit and loss prediction is solved.

[0053] In some optional embodiments, the exchange rate profit and loss prediction model to be trained includes: a feature extraction module and a profit and loss prediction module; the inputting of the expanded exchange rate index data into the exchange rate profit and loss prediction model to be trained to obtain the exchange rate profit and loss prediction result output by the exchange rate profit and loss prediction model to be trained includes: inputting the expanded exchange rate index data into the exchange rate profit and loss prediction model to be trained, extracting the index feature data in the expanded exchange rate index data based on the feature extraction module; and determining the exchange rate profit and loss prediction result corresponding to the expanded exchange rate index data based on the index feature data by the profit and loss prediction module.

[0054] Among them, the feature extraction module can be implemented by a convolutional neural network or an autoencoder. The profit and loss prediction module can be implemented by a classifier (for example, a decision tree, a support vector machine or a logistic regression model, etc.) or a neural network model. The feature extraction module can extract the indicator feature data related to the exchange rate profit and loss prediction (that is, the factors affecting the exchange rate fluctuation) in the expanded exchange rate indicator data, and determine the exchange rate profit and loss prediction result based on the indicator feature data based on the profit and loss prediction module.

[0055] Please refer to Figure 2 , Figure 2 A schematic diagram of the training process of a profit and loss prediction module provided in an embodiment of the present application. Among them, the profit and loss prediction module can be implemented by a support vector machine, and specifically uses the support vector machine algorithm of quantum measurement as a classification algorithm. In some optional embodiments, the training process of the profit and loss prediction module includes: S201, converting the indicator feature data into a quantum state to obtain quantum feature data; S202, adjusting the entanglement and phase between the quantum feature data based on a quantum logic gate; wherein, the parameter configuration of the quantum logic gate is determined based on the current internal parameters of the profit and loss prediction module; S203, measuring the adjusted quantum feature data, and determining the exchange rate profit and loss prediction result corresponding to the expanded exchange rate indicator data based on the measurement result; S204, based on the exchange rate profit and loss prediction result and the exchange rate profit and loss data corresponding to the expanded exchange rate indicator data, adjusting the internal parameters of the profit and loss prediction module until the training conditions of the profit and loss prediction module are met.

[0056] Before training the profit and loss prediction module, the parameters of the profit and loss prediction module can be converted to quantum states through initialization. During the training of the profit and loss prediction module, each qubit represents a feature dimension, and feature conversion and nonlinear mapping are achieved through quantum gate operations. For example, the superposition of the quantum state can be initialized based on the Hadamard gate. Initialize the parameters of the profit and loss prediction module; where ψ u represents the initialized quantum state; Indicates the application of Hadamard gate to each quantum bit; n u represents the number of qubits; represents the initial state of all quantum bits; represents the initial quantum state.

[0057] In step S201, the indicator characteristic data may be converted using a set quantum coding strategy to obtain quantum characteristic data. For example, the indicator characteristic data may be converted into a probability amplitude of a quantum bit based on amplitude coding, and as well as Convert the index characteristic data into quantum state; where x u Represents indicator characteristic data (usually expressed in vector form); x u，i Represents the indicator characteristic data x u The i-th element of u，i Represents the normalized indicator feature data x u The i-th element (α u，i represents the amplitude of the corresponding quantum state, and it is necessary to ensure that the total probability amplitude of the quantum state is 1); |i> represents the i-th quantum ground state; represents the first quantum state.

[0058] In step S202, the nonlinear characteristic transformation can be implemented based on quantum logic gates. For example, CNOT gates and parameterized quantum gates ( 如 Revolving Door y ) to adjust the entanglement and phase between quantum characteristic data. That is, based on as well as To adjust the entanglement and phase between quantum bits; where U u (θ u ) is given by the rotation angle θ u Parameterized sequence of quantum gate operations; θ u is the rotation angle, representing the parameter of the profit and loss prediction module; m u is the number of gates applied, CNOT j,k is the controlled NOT gate applied to the j-th qubit and the k-th qubit; R y () and R z () are the revolving doors corresponding to the y-axis and z-axis respectively, R y () and R z () can be adjusted according to the angle and Rotate to adjust the phase and amplitude of the quantum bit state; and is the angle parameter of the revolving door corresponding to the y-axis and z-axis; θ u，kis the kth dimension parameter corresponding to the rotation angle; is the second quantum state; i* is an imaginary unit. Based on the rotating gate, the entanglement and nonlinearity of the quantum state can be increased.

[0059] In step S203, based on The adjusted quantum characteristic data is measured, and the exchange rate profit and loss prediction result is obtained based on the measurement result; where P(u|x u ) is a given input x u The probability that the post-quantum state is |1> (that is, the probability that the quantum bit is in the |1> state) is used to characterize the probability that the classification decision corresponds to different categories; y u is the class label (can be +1 or -1); Tr[] is the trace operation of the matrix; is the generalized inverse matrix of the second quantum state. During the measurement process, the collapse of the quantum state provides information about the decision boundary. Using this information, the decision boundary of the profit and loss prediction module (support vector machine) is optimized by adjusting the quantum state, and the exchange rate profit and loss prediction result is obtained by measuring the quantum bit.

[0060] In step S204, based on Calculate the loss function of the profit and loss prediction module (support vector machine); where, is the loss function of the quantum support vector machine. And the internal parameters of the profit and loss prediction module are adjusted based on the calculated loss function until the training conditions of the profit and loss prediction module are met. The classification accuracy of the model can be refined through iterative optimization of the quantum state. After each iteration, the parameters and configuration of the quantum logic gate (that is, the internal parameters of the profit and loss prediction module) are adjusted according to the measurement results to gradually reduce the classification errors. The training conditions of the profit and loss prediction module may include a maximum number of iterations (which may be one thousand times or other reasonable values) or a loss function threshold. When the number of adjustments to the internal parameters of the profit and loss prediction module reaches the maximum number of iterations, or the calculated loss function of the quantum support vector machine is less than the loss function threshold, it is considered that the training conditions of the profit and loss prediction module are met.

[0061] In some optional embodiments, it can be based on Update the parameters and configuration of the quantum logic gate; where η u is the learning rate of the quantum support vector machine, which can be 0.01 or other reasonable values; u is the actual classification label (corresponding to the exchange rate profit and loss forecast result); is the gradient of the quantum support vector machine parameters; are the parameters of the quantum support vector machine at the t+1th iteration; are the parameters of the quantum support vector machine at the tth iteration.

[0062] Please refer to Figure 3 , Figure 3 A schematic diagram of the training process of a feature extraction module provided for an embodiment of the present application. The feature extraction module can be implemented by a multi-layer fully connected neural network. In some optional embodiments, the training process of the feature extraction module may include: S301, generating an initialization population; wherein the initialization population includes multiple individuals, each of which corresponds to a set of module parameters of the feature extraction module; S302, calculating the feature loss value corresponding to each of the individuals according to the module parameters; S303, selecting candidate individuals that meet the candidate conditions from the initialization population according to the feature loss value; S304, generating new individuals based on the candidate individuals and through crossover operations and mutation operations until the training conditions of the feature extraction module are met; wherein, during the training process of the feature extraction module, the probability of the crossover operation and / or the mutation operation is adjusted according to the changing trend of the feature loss value.

[0063] In step S301, taking the feature extraction module implemented by a multi-layer fully connected neural network as an example, an initial population can be generated in the initialization stage according to the bionic algorithm initialization method. as well as Generate an initial population; where W pi is the weight of the neural network (feature extraction module) corresponding to the i-th individual; b pi is the bias of the neural network (feature extraction module) corresponding to the i-th individual; Represents the weight matrix of the i-th individual in the initial state; represents the bias of the i-th individual in the initial state; σ 2 Represents the initialization variance, which can be 0.01 or other reasonable values; Indicates that the mean is 0 and the variance is σ 2 The normal distribution of .

[0064] In step S302, for each individual in the population, the input training data is processed using its corresponding neural network configuration, the output of the feature extraction module is calculated, and the individual performance is evaluated according to a preset loss function. The loss of the i-th individual on the training data set can be calculated based on the weight and bias corresponding to the i-th individual. Specifically, Calculate the feature loss value corresponding to the i-th individual; where Lpi is the feature loss value of the neural network (feature extraction module) corresponding to the i-th individual; mps represents the number of samples input in the current batch; l() represents the composite loss function; fsig() represents the neural network model function (model function of the feature extraction module); Represents the characteristics of the jth sample; represents the label of the jth sample.

[0065] In some optional embodiments, a composite loss function including a regularization term can be used as the feature loss value corresponding to the individual, that is, based on as well as Calculate the feature loss value corresponding to the i-th individual; where MSE() is the mean square error function; Reg(W pi ) is the regularization term; λ ps is the regularization parameter; W pi，k is the kth weight of the neural network (feature extraction module) corresponding to the ith individual.

[0066] In step S303, the number of selected candidate individuals can be adjusted according to actual conditions. Calculate the fitness of the i-th individual and use the individual fitness as the probability of the individual being selected as a candidate individual; where P select (i) represents the probability of the i-th individual being selected; γ pse is a parameter that controls the selection pressure and can be 2 or other reasonable values; L pk is the loss of the neural network corresponding to the kth individual; N p is the population size. Individuals whose probability of being selected is greater than a preset probability value (which may be 0.8 or other reasonable values) may be selected as candidate individuals.

[0067] In step S304, the crossover operation allows two excellent individuals to exchange some genes to generate new offspring; the mutation operation randomly changes some genes in the individuals to increase the diversity of the population. pi =α pcs W p1 +(1-α pcs )W p2 and b′ pi =α pcs b p1 +(1-α pcs )b p2 To realize the crossover operation between the two selected candidate individuals; where α pcs is the crossover rate (can be 0.3 or other reasonable value); W p1 is the weight of the neural network corresponding to the first individual selected; b p1 is the bias of the neural network corresponding to the first individual selected; W p2 is the weight of the neural network corresponding to the second individual selected; b p2is the bias of the neural network corresponding to the second individual selected; W′ pi is the weight of the neural network corresponding to the individual after the crossover operation; b′ pi is the bias of the neural network corresponding to the individual after the crossover operation. as well as Realize the mutation operation of the individuals generated by the crossover operation; where τ 2 represents the variance of the variation (can be 0.04 or other reasonable values), W″ pi is the weight of the neural network corresponding to the individual after the mutation operation, b″ pi is the bias of the neural network corresponding to the individual after the mutation operation. The individual weights and / or biases generated by the crossover operation can be randomly perturbed by a small amplitude through the mutation operation to further increase the diversity of the individuals. The probability of the crossover operation can be adjusted by adjusting the number of selected candidate individuals or the crossover rate; the probability of the mutation operation can be adjusted by adjusting the number of selected candidate individuals or the variation variance. The training conditions of the feature extraction module may include a maximum number of iterations (which may be one thousand times or other reasonable values) or a feature loss threshold. When the number of operations of the crossover operation and the mutation operation reaches the maximum number of iterations, or the feature loss value of the feature extraction module calculated is less than the feature loss threshold, it is considered that the training conditions of the feature extraction module are met. Specifically, after a new individual is generated through the crossover operation and the mutation operation, the feature loss value corresponding to the new individual can be calculated again; if the feature loss value corresponding to the new individual is less than the feature loss threshold, it is considered that the training conditions of the feature extraction module are met. If the feature loss value corresponding to the new individual is greater than or equal to the feature loss threshold, it is considered that the training conditions of the feature extraction module are not met; at this time, based on the feature loss value, candidate individuals that meet the candidate conditions can be selected from the optimized population (including the original individuals and the newly generated individuals); again based on the selected candidate individuals, new individuals are generated through crossover operations and mutation operations until the training conditions of the feature extraction module are met.

[0068] Please refer to Figure 4 , Figure 4A schematic diagram of a training process of a generative adversarial network provided for an embodiment of the present application. In some optional embodiments, the generative adversarial network includes a generator and a discriminator; the training process of the generative adversarial network includes: S401, based on the generator receiving a noise signal, and based on the noise signal, the exchange rate index data and the exchange rate profit and loss data corresponding to the exchange rate index data, generating simulated training data; S402, based on the optimal transmission theory, the simulated training data and the discriminant function of the discriminator, determining the generation loss function value of the generator; S403, based on the optimal transmission theory, the simulated training data, the exchange rate index data and the exchange rate profit and loss data corresponding to the exchange rate index data, determining the discriminant loss function value of the discriminator; S404, adjusting the internal parameters of the generator based on the generation loss function value, and adjusting the internal parameters of the discriminator based on the discriminant loss function value until the training conditions of the generative adversarial network are met.

[0069] Before training the generative adversarial network, the parameters of the generator and the discriminator can be initialized. Specifically, Implement the initialization of the generator based on To achieve the initialization of the discriminator; where G c is the generator; D c is the discriminator; is the weight of the generator; is the weight of the discriminator; ~ means obeying a specific distribution; represents a normal distribution with mean 0 and variance 0.02.

[0070] In step S401, the noise signal may be a random noise signal, and simulated training data may be generated based on the random noise signal and in combination with conditional coding (exchange rate index data and exchange rate profit and loss data corresponding to the exchange rate index data). The goal of the generator is to imitate the statistical characteristics of the real exchange rate index data and the corresponding relationship between the "real exchange rate index data and exchange rate profit and loss data" as much as possible. as well as Generate simulated training data; where z c From uniform distribution The noise signal sampled in represents the uniform distribution in the interval (-1, 1), λ c Encode conditions, such as regional economic indicators or historical exchange rate data; is the generated simulated training data; G c () represents a generator function.

[0071] In step S402, the stability and convergence speed of the generative adversarial network can be enhanced based on the optimal transmission theory. Calculate the loss function value; where D c () is the discriminant function of the discriminator; Represents the generated loss function value.

[0072] In step S403, the task of the discriminator is to distinguish the data generated by the generator from the data in the real data set. The discriminator can gradually improve its discrimination ability by learning to identify the difference between the generated data and the real data. as well as Calculate the discriminant loss function value; where x c Represents the real exchange rate index data and the exchange rate profit and loss data corresponding to the exchange rate index data; represents the simulated training data produced by the generator; is the discriminant loss function value; f c () is an auxiliary function; x represents the input of the auxiliary function; max() is the maximum value function; min() is the minimum value function. The auxiliary function can be used to calculate the standardized loss. The loss calculation based on progressive standardization helps the discriminator improve its sensitivity when dealing with unbalanced data.

[0073] In step S404, the training conditions of the generative adversarial network may include a maximum number of iterations (which may be one thousand or other reasonable values) or a loss value condition. The training conditions of the generative adversarial network may be considered to be met when the number of parameter updates of the generator and the discriminator reaches the maximum number of iterations, or when the calculated generation loss function value and the discriminant loss function value meet the loss value condition.

[0074] In some optional embodiments, the updating of the internal parameters of the generator based on the generation loss function value includes: updating the internal parameters of the generator based on the generation loss function value obtained in the current training process, the generation loss function value obtained in the previous N training processes, and the internal parameters of the generator after the previous N training updates; wherein N is a positive integer.

[0075] Among them, it can be based on as well as Update the internal parameters of the generator; where, and Respectively represent the generator parameters at the tth and t+1th iterations; η gan is the learning rate of the generative adversarial network; is the loss function of the generator at the tth iteration; Indicates about The gradient of c is a dynamic weight adjusted based on the Sigmoid function; k c is a tuning parameter, which can be 0.3 or other reasonable values; Tc is the total number of training cycles; is the attenuation factor, which can be attenuated by simulated annealing strategy to adjust the influence of historical gradients; is the generator loss function value of the ti-th iteration; int(t) is the current iteration number; h c (t) is the historical gradient influence function; δ c (t) is the dynamic adjustment factor of interactive learning.

[0076] In some optional embodiments, the interactive learning mechanism can be used to monitor the response of the discriminator to the generated data in real time, and adjust the output strategy of the generator to more effectively simulate the complexity of "exchange rate indicator data and exchange rate profit and loss data corresponding to the exchange rate indicator data". That is, the discriminator adjusts the dynamic adjustment factor of interactive learning by identifying the performance difference between generated data and real data; where, and They represent the loss value of the discriminator for the real data in the tth training cycle and the loss value of the generated data in the tth training cycle respectively; σ c It is the standard deviation parameter used to adjust the dynamic response sensitivity, which can be 0.01 or other reasonable values; tanh() is the hyperbolic tangent function.

[0077] In some optional embodiments, the generation loss function value includes a transmission loss value and an entropy regularization term; determining the generation loss function value of the generator based on the optimal transmission theory, the simulated training data, and the discriminant function of the discriminator includes: determining the transmission loss value based on the optimal transmission theory, the simulated training data, and the discriminant function of the discriminator; determining the entropy regularization term based on the entropy regularization term of the simulated training data.

[0078] Among them, it can be based on Determine the transmission loss value. Based on Determine the entropy regularization term; where γ c is a hyperparameter that controls the influence of the regularization term and can be 0.1 or other reasonable values; is a regular term based on sample entropy. Calculate the generated loss function value

[0079] In some optional embodiments, before the exchange rate profit and loss prediction result corresponding to the expanded exchange rate indicator data is determined based on the profit and loss prediction module according to the indicator characteristic data, the method further includes: performing dimensionality reduction processing on the indicator characteristic data based on a feature dimensionality reduction module to obtain dimensionality reduction characteristic data; the exchange rate profit and loss prediction result corresponding to the expanded exchange rate indicator data is determined based on the profit and loss prediction module according to the indicator characteristic data, including: determining the exchange rate profit and loss prediction result corresponding to the expanded exchange rate indicator data based on the profit and loss prediction module according to the dimensionality reduction characteristic data.

[0080] Among them, the feature dimension reduction module can be implemented based on principal component analysis, recursive feature elimination or sparse regularization feature dimension reduction methods. The feature dimension reduction module can reduce the dimension of the indicator feature data to help remove redundant features and reduce the correlation between features, thereby improving the generalization ability and prediction accuracy of the model. It can also reduce the complexity of feature data and speed up the model training speed and exchange rate profit and loss prediction speed.

[0081] Please refer to Figure 5 , Figure 5 A schematic diagram of the training process of a dimensionality reduction feature module provided in an embodiment of the present application. In some optional embodiments, an autoencoding neural network based on feature refinement can be used as a feature dimensionality reduction module. The dimensionality reduction feature module can specifically include an encoder and a decoder; the training process of the dimensionality reduction feature module includes: S501, mapping the indicator feature data to a dimensionality reduction feature space of a preset dimension based on the encoder to obtain dimensionality feature data; S502, reconstructing the dimensionality feature data back to the original feature space based on the decoder to obtain reconstructed feature data; S503, calculating the dimensionality reduction loss function value of the dimensionality reduction feature module based on the reconstructed feature data and the indicator feature data; S504, adjusting the internal parameters of the dimensionality reduction feature module based on the dimensionality reduction loss function value until the training conditions of the dimensionality reduction feature module are met.

[0082] Among them, in steps S501-S504, the encoder can be based on Z r =Sig enc (W r X r +b r ) maps the indicator feature data to the reduced-dimensional feature space of the preset dimension; where X r Indicates indicator characteristic data; Z r Represents dimensional feature data; W r is the weight matrix of the encoder; b r is the bias vector of the encoder; Sig enc () is the multi-layer Sigmoid activation function of the encoder. The decoder can be based on X′r =Sig dec (W′ r Z r +b′ r ) reconstructs the dimensional feature data back to the original feature space; where X′ r is the reconstructed high-dimensional data, W′ r is the weight matrix of the decoder, b′ r is the bias vector of the decoder, Sig dec () is the multi-layer Sigmoid activation function of the decoder. The above-mentioned dimensionality reduction loss function value can be calculated based on the reconstruction error loss function. The training conditions of the dimensionality reduction feature module may include a maximum number of iterations (which may be one thousand times or other reasonable values) or a dimensionality reduction loss value condition. When the number of parameter updates of the dimensionality reduction feature module reaches the maximum number of iterations, or the calculated dimensionality reduction loss function value satisfies the dimensionality reduction loss value condition, it is considered that the training conditions of the dimensionality reduction feature module are met.

[0083] In some optional embodiments, the dimension reduction feature module may further include a feature adjustment unit, which is used to generate feature weights according to the feature importance in the current feature space. When the module is initialized, all features may be given the same initial weight so that they can be gradually adjusted according to feature contributions in subsequent steps. Specifically, based on A r =diag(α r ) generates feature weights; where A r is the initial weight matrix; α r is the feature weight vector; α r Each element α in r,i Initialized to the same value, indicating that all features have the same importance in the initial stage; diag() is a function used to extract the diagonal elements of the matrix. Further, based on Z′ r =A r Z r Calculate the adjusted feature data; where Z′ r is the feature representation after feature weight adjustment. In this case, the decoder can be based on X′ r =Sig dec (W′ r Z′ r +b′ r ) reconstructs the dimensional feature data back to the original feature space. Update the feature weights of the feature adjustment unit; where, Represents the feature weight of the t+1th iteration; represents the feature weight of the tth iteration; η ris the learning rate of the autoencoder neural network (that is, the dimension reduction feature module), which can be 0.01 or other reasonable values; L r () is the dimensionality reduction loss function value of the dimensionality reduction feature module; Y r is the label data; Represents the gradient of the dimensionality reduction loss function value with respect to the feature weight.

[0084] Please refer to Figure 6 , Figure 6 A flow chart of a method for forecasting exchange rate gains and losses provided in an embodiment of the present application. The method for forecasting exchange rate gains and losses includes:

[0085] S601, obtaining exchange rate index data related to the target project;

[0086] S602. Input the exchange rate indicator data into an exchange rate profit and loss prediction model to obtain an exchange rate profit and loss prediction result output by the model; wherein the exchange rate profit and loss prediction model is configured to be obtained by training based on the exchange rate profit and loss prediction model training method as described in any one of the first aspects.

[0087] Among them, the exchange rate index data related to the target project can be obtained through automated web crawler technology. The specific implementation of the exchange rate profit and loss prediction method can refer to the relevant description of the exchange rate profit and loss prediction model training method above. To avoid repetition, some detailed descriptions are omitted here.

[0088] An embodiment of the present application further provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the method described in any one of the first aspect or the second aspect above.

[0089] An embodiment of the present application also provides an electronic device; the electronic device includes: a memory; a processor; a computer program executable by the processor is stored on the memory, and when the computer program is executed by the processor, the method described in any one of the first aspect or the second aspect is executed.

[0090] An embodiment of the present application further provides a computer-readable storage medium, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described in the first aspect or the second aspect above is executed.

[0091] Among them, the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable red-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0092] In several embodiments provided by the embodiments of the present application, it should be understood that the disclosed device / system and method can also be implemented in other ways. The device embodiments described above are merely schematic. For example, the flowcharts and block diagrams in the accompanying drawings show the possible architecture, functions and operations of the devices, methods and computer program products according to the multiple embodiments of the embodiments of the present application. In this regard, each box in the flowchart or block diagram can represent a module, a program segment or a part of the code, and a part of the module, program segment or code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order than the order marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, and the combination of boxes in the block diagram and / or flowchart can be implemented with a dedicated hardware-based system that performs a specified function or action, or can be implemented with a combination of dedicated hardware and computer instructions.

[0093] In addition, the functional modules in each embodiment of the present application can be integrated together to form an independent part, or each module can exist separately, or two or more modules can be integrated to form an independent part.

[0094] The above description is only an optional implementation manner of the embodiments of the present application, but the protection scope of the embodiments of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or replacements within the technical scope disclosed in the embodiments of the present application, which should be covered within the protection scope of the embodiments of the present application.

Claims

1. A method for training an exchange rate profit and loss prediction model, characterized in that: The method comprises: Based on the generative adversarial network, the acquired exchange rate indicator data related to the target project is expanded; Inputting the expanded exchange rate index data into the exchange rate profit and loss prediction model to be trained, and obtaining the exchange rate profit and loss prediction result output by the exchange rate profit and loss prediction model to be trained; According to the exchange rate profit and loss prediction result and the exchange rate profit and loss data corresponding to the expanded exchange rate indicator data, the internal parameters of the exchange rate profit and loss prediction model to be trained are adjusted to obtain a trained exchange rate profit and loss prediction model.

2. The method according to claim 1, characterized in that in, The exchange rate profit and loss prediction model to be trained includes: a feature extraction module and a profit and loss prediction module; The step of inputting the expanded exchange rate index data into the exchange rate profit and loss prediction model to be trained, and obtaining the exchange rate profit and loss prediction result output by the exchange rate profit and loss prediction model to be trained, comprises: Inputting the expanded exchange rate index data into the exchange rate profit and loss prediction model to be trained, and extracting index feature data from the expanded exchange rate index data based on the feature extraction module; Based on the profit and loss prediction module, the exchange rate profit and loss prediction result corresponding to the expanded exchange rate indicator data is determined according to the indicator characteristic data.

3. The method according to claim 2, characterized in that in, The training process of the profit and loss prediction module includes: Converting the indicator characteristic data into a quantum state to obtain quantum characteristic data; Adjusting the entanglement and phase between the quantum feature data based on a quantum logic gate; wherein the parameter configuration of the quantum logic gate is determined based on the current internal parameters of the profit and loss prediction module; Measuring the adjusted quantum characteristic data, and determining the exchange rate profit and loss prediction result corresponding to the expanded exchange rate index data based on the measurement result; Based on the exchange rate profit and loss prediction result and the exchange rate profit and loss data corresponding to the expanded exchange rate indicator data, internal parameters of the profit and loss prediction module are adjusted until the training conditions of the profit and loss prediction module are met.

4. The method according to claim 2, characterized in that: in, The training process of the feature extraction module includes: Generate an initialization population; wherein the initialization population includes a plurality of individuals, each of the individuals corresponds to a set of module parameters of the feature extraction module; Calculating the feature loss value corresponding to each of the individuals according to the module parameters; According to the feature loss value, selecting a candidate individual that meets the candidate condition from the initialization population; Based on the candidate individuals, new individuals are generated through crossover operations and mutation operations until the training conditions of the feature extraction module are met; Wherein, during the training process of the feature extraction module, the probability of the crossover operation and / or the mutation operation is adjusted according to the changing trend of the feature loss value.

5. The method according to claim 1, characterized in that in, The generative adversarial network includes a generator and a discriminator; The training process of the generative adversarial network includes: Receiving a noise signal based on the generator, and generating simulated training data based on the noise signal, the exchange rate indicator data, and the exchange rate profit and loss data corresponding to the exchange rate indicator data; Determine the generation loss function value of the generator based on the optimal transmission theory, the simulated training data and the discriminant function of the discriminator; Determining a discriminant loss function value of the discriminator based on the optimal transmission theory, the simulated training data, the exchange rate index data, and the exchange rate profit and loss data corresponding to the exchange rate index data; The internal parameters of the generator are adjusted based on the generation loss function value, and the internal parameters of the discriminator are adjusted based on the discriminant loss function value until the training conditions of the generative adversarial network are met.

6. The method according to claim 5, characterized in that The updating of the internal parameters of the generator based on the generated loss function value comprises: Update the internal parameters of the generator based on the generation loss function value obtained in the current training process, the generation loss function values ​​obtained in the previous N training processes, and the internal parameters of the generator after the previous N training updates; where N is a positive integer.

7. The method according to claim 5, characterized in that in, The generated loss function value includes a transmission loss value and an entropy regularization term; The step of determining the generation loss function value of the generator based on the optimal transmission theory, the simulated training data, and the discriminant function of the discriminator includes: Determining the transmission loss value based on the optimal transmission theory, the simulated training data, and the discriminant function of the discriminator; The entropy regularization term is determined based on a regularization term of the entropy of the simulated training data.

8. The method according to claim 2, characterized in that: Before determining the exchange rate profit and loss prediction result corresponding to the expanded exchange rate indicator data based on the profit and loss prediction module according to the indicator feature data, the method further includes: Performing dimensionality reduction processing on the indicator feature data based on a feature dimensionality reduction module to obtain dimensionality reduction feature data; The step of determining the exchange rate profit and loss prediction result corresponding to the expanded exchange rate indicator data based on the profit and loss prediction module according to the indicator feature data includes: Based on the profit and loss prediction module, the exchange rate profit and loss prediction result corresponding to the expanded exchange rate indicator data is determined according to the dimension reduction feature data.

9. The method according to claim 8, characterized in that in, The dimension reduction feature module includes an encoder and a decoder; The training process of the dimension reduction feature module includes: Mapping the indicator feature data to a reduced-dimensional feature space of a preset dimension based on the encoder to obtain dimensional feature data; Reconstructing the dimensional feature data back to the original feature space based on the decoder to obtain reconstructed feature data; Calculating a dimensionality reduction loss function value of the dimensionality reduction feature module according to the reconstructed feature data and the indicator feature data; The internal parameters of the dimensionality reduction feature module are adjusted based on the dimensionality reduction loss function value until the training conditions of the dimensionality reduction feature module are met.

10. A method for forecasting exchange rate gains and losses, characterized in that: The method comprises: Obtain exchange rate indicator data related to the target project; The exchange rate indicator data is input into the exchange rate profit and loss prediction model to obtain the exchange rate profit and loss prediction result output by the model; wherein the exchange rate profit and loss prediction model is configured to be obtained by training based on the exchange rate profit and loss prediction model training method according to any one of claims 1 to 9.

Images