Method and device for measuring and calculating pledge rate of digital assets

Abstract

The invention discloses a method and device for measuring and calculating the pledge rate of digital assets, and the method comprises the steps: extracting the target features of the digital assets from the data of the digital assets, and enabling the target features to comprise the attributes, which affect the pledge rate, of the digital assets; using an LSTM model to predict the pledge rate of the digital assets based on the target features of the digital assets, the training set data of the LSTM model is determined based on attributes of a plurality of sample digital assets, which influence pledge rates, historical net values of the plurality of sample digital assets and overdue records of the plurality of sample digital assets, and hyper-parameters of the LSTM model are determined based on a WOA algorithm; and writing the predicted pledge rate into an intelligent contract, and performing pledge or depledge on the digital assets by a block chain system based on the pledge rate.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a method and device for calculating the pledge rate of digital assets. Background Art

[0002] Traditional pledged credit products require customers to pledge time deposits, large-denomination certificates of deposit or structured deposits in banks. The types of pledged collateral are single and customers have few choices, which reduces customer stickiness. For example, if a customer has a large deposit in a bank, but at a certain time the customer urgently needs a sum of money, since pledged credit products generally require that the loan interest rate must not be lower than the highest return on the collateral, the customer will be more likely to withdraw the large deposit instead of using it as collateral for a loan, which leads to the outflow of bank funds and customer loss. Therefore, it is very necessary to expand the types of collateral.

[0003] With the rise of the digital economy and the popularization of blockchain technology, it has become feasible to use digital assets as collateral. Blockchain originated from Bitcoin, which proposed a decentralized deposit book technology for recording Bitcoin transactions. Unlike the traditional centralized banking system, blockchain uses cryptography and consensus mechanisms to improve the security and transparency of transactions. With the gradual deepening of blockchain technology research and the introduction of the concept of smart contracts, blockchain can be used to execute contracts and transactions. Existing blockchains have a good application in supply chain financing. With the core enterprises on the alliance chain as the center, the assets of upstream and downstream enterprises in the supply chain are digitized and standardized, thereby reducing the operating costs of the entire supply chain. Multi-level suppliers can perform identity authentication, asset registration, and circulation on the chain, banks can lock account funds on the chain, and both parties can edit and execute smart contracts, so that data preservation and traceability can be achieved during the entire transaction.

[0004] However, the expansion of collateral will inevitably lead to increased risks. In order to reduce risks, the most effective way is to give a suitable pledge rate to digital assets. In view of this, it is urgent to provide a method for calculating the pledge rate of digital assets that can provide a relatively accurate calculation. Summary of the invention

[0005] The present application provides a method and device for calculating the pledge rate of digital assets, which can provide a relatively accurate pledge rate. The technical solution is as follows.

[0006] In a first aspect, a method for calculating the pledge rate of a digital asset is provided, the method comprising:

[0007] Extracting target features of the digital asset from the data of the digital asset, wherein the target features include attributes of the digital asset itself that affect the pledge rate;

[0008] The LSTM model is used to predict the pledge rate of the digital asset based on the target features of the digital asset, wherein the training set data of the LSTM model is determined based on the attributes of the multiple sample digital assets that affect the pledge rate, the historical net value of the multiple sample digital assets, and the overdue records of the multiple sample digital assets, and the hyperparameters of the LSTM model are determined based on the WOA algorithm;

[0009] The predicted pledge rate is written into the smart contract, and the blockchain system performs pledge or unpledge of the digital assets based on the pledge rate.

[0010] In some implementations, extracting the target feature of the digitized asset from the data of the digitized asset includes:

[0011] Extract the balance sheet, income statement, cash flow, industry, operating conditions, employee status, credit history and capital structure at multiple times from the digital asset data to obtain the first eigenvector;

[0012] Determining, based on the historical net value of the digital asset at the multiple times, the correlation between the multiple times and the pledge rate prediction;

[0013] According to the correlation between the multiple moments and the pledge rate prediction, k key moments are determined from the multiple moments, and the correlation between the k key moments and the pledge rate prediction is ranked in the top k positions among the multiple moments;

[0014] Extracting the balance sheet, income statement, cash flow, industry, operating conditions, employee conditions, credit history and capital structure of the digital asset at the k key moments from the data of the digital asset to obtain a second eigenvector;

[0015] Using a convolutional neural network to reduce the dimension of the second eigenvector to obtain a third eigenvector;

[0016] The first feature vector and the third feature vector are concatenated to obtain a target feature of the digitized asset.

[0017] In some embodiments, the training set data of the LSTM model includes labels. When there is an overdue in the overdue record of a sample digital asset, the label corresponding to the sample digital asset is the product of the historical net value of the sample digital asset and the target coefficient. When there is no overdue in the overdue record of a sample digital asset, the label corresponding to the sample digital asset is the historical net value of the sample digital asset, and the target coefficient is obtained by iterative updating in the LSTM model.

[0018] In some embodiments, the hyperparameters of the LSTM model include the number of iterations, the learning rate, the number of neurons in the first hidden layer, and the number of neurons in the second hidden layer of the LSTM model. The process of determining the hyperparameters of the LSTM model includes:

[0019] Randomly generate multiple populations, each population corresponding to a parameter value of a hyperparameter of the LSTM model;

[0020] The LSTM model is set using the parameter values ​​of the hyperparameters corresponding to each of the populations;

[0021] The pledge rate predicted by the LSTM model and the actual pledge rate are used as fitness, the fitness corresponding to each of the populations is determined, the minimum fitness is determined from the fitness corresponding to each of the populations, and the parameter value of the hyperparameter corresponding to the minimum fitness is used as the optimal hyperparameter of this iteration;

[0022] Compare the optimal hyperparameters of this iteration with the global optimal hyperparameters in the historical iteration process. If the model prediction effect corresponding to the optimal hyperparameters of this iteration is better than the model prediction effect corresponding to the global optimal hyperparameters, replace the global optimal hyperparameters with the optimal hyperparameters of this iteration;

[0023] Using the whale algorithm to update the parameter value of the hyperparameter corresponding to each of the populations, and using the updated parameter value of the hyperparameter to set the LSTM model;

[0024] Repeat the steps of hyperparameter comparison, hyperparameter replacement, and parameter value update until the number of iterations reaches the set maximum number of iterations.

[0025] In some implementations, the updating of the parameter value of the hyperparameter corresponding to each of the populations using the whale algorithm includes:

[0026] Generate random numbers;

[0027] If the random number is less than the random number threshold and the length of the control parameter vector is less than the threshold, the parameter value of the hyperparameter is determined by shrinking and wrapping;

[0028] If the random number is less than the random number threshold, and the length of the control parameter vector is greater than or equal to the threshold, the parameter value of the hyperparameter is determined by searching;

[0029] If the random number is greater than or equal to the random number threshold, the parameter value of the hyperparameter is determined in a spiral manner.

[0030] In a second aspect, a pledge rate calculation device for digital assets is provided, the device comprising:

[0031] A feature extraction unit, used to extract target features of the digital asset from the data of the digital asset, wherein the target features include the attributes of the digital asset itself that affect the pledge rate;

[0032] A prediction unit, configured to predict the pledge rate of the digital asset based on the target features of the digital asset using an LSTM model, wherein the training set data of the LSTM model is determined based on the attributes of a plurality of sample digital assets that affect the pledge rate, the historical net values ​​of the plurality of sample digital assets, and the overdue records of the plurality of sample digital assets, and the hyperparameters of the LSTM model are determined based on the WOA algorithm;

[0033] A writing unit is used to write the predicted pledge rate into the smart contract, and the blockchain system performs pledge or unpledge on the digital asset based on the pledge rate.

[0034] In some embodiments, the feature extraction unit is used to extract the balance sheet, income statement, cash flow, industry, operating conditions, employee status, credit history and capital structure at multiple moments from the data of the digital asset to obtain a first feature vector; determine the correlation between the multiple moments and the pledge rate prediction based on the historical net value of the digital asset at the multiple moments; determine k key moments from the multiple moments based on the correlation between the multiple moments and the pledge rate prediction, and the correlation between the k key moments and the pledge rate prediction ranks in the top k among the multiple moments; extract the balance sheet, income statement, cash flow, industry, operating conditions, employee status, credit history and capital structure of the digital asset at the k key moments from the data of the digital asset to obtain a second feature vector; use a convolutional neural network to reduce the dimension of the second feature vector to obtain a third feature vector; and concatenate the first feature vector and the third feature vector to obtain the target feature of the digital asset.

[0035] In some embodiments, the training set data of the LSTM model includes labels. When there is an overdue in the overdue record of a sample digital asset, the label corresponding to the sample digital asset is the product of the historical net value of the sample digital asset and the target coefficient. When there is no overdue in the overdue record of a sample digital asset, the label corresponding to the sample digital asset is the historical net value of the sample digital asset, and the target coefficient is obtained by iterative updating in the LSTM model.

[0036] In some embodiments, the hyperparameters of the LSTM model include the number of iterations of the LSTM model, the learning rate, the number of neurons in the first hidden layer, and the number of neurons in the second hidden layer. The device also includes: a parameter determination unit, which is used to randomly generate multiple populations, each of which corresponds to a parameter value of a hyperparameter of the LSTM model; the LSTM model is set using the parameter values ​​of the hyperparameters corresponding to each of the populations; the pledge rate predicted by the LSTM model and the actual pledge rate are used as fitness, the fitness corresponding to each of the populations is determined, the minimum fitness is determined from the fitness corresponding to each of the populations, and the minimum fitness is set The parameter value of the hyperparameter corresponding to the degree is used as the optimal hyperparameter of this iteration; the optimal hyperparameter of this iteration is compared with the global optimal hyperparameter in the historical iteration process. If the model prediction effect corresponding to the optimal hyperparameter of this iteration is better than the model prediction effect corresponding to the global optimal hyperparameter, the optimal hyperparameter of this iteration is used to replace the global optimal hyperparameter; the parameter value of the hyperparameter corresponding to each of the populations is updated by using the whale algorithm, and the parameter value of the updated hyperparameter is used to set the LSTM model; the steps of hyperparameter comparison, hyperparameter replacement and parameter value update are repeated until the number of iterations reaches the set maximum number of iterations.

[0037] In some embodiments, the parameter determination unit is used to generate a random number; if the random number is less than a random number threshold and the length of the control parameter vector is less than the threshold, the parameter value of the hyperparameter is determined by a shrinking and wrapping method; if the random number is less than the random number threshold and the length of the control parameter vector is greater than or equal to the threshold, the parameter value of the hyperparameter is determined by a search method; if the random number is greater than or equal to the random number threshold, the parameter value of the hyperparameter is determined by a spiral method.

[0038] In a third aspect, a computer device is provided, the computer device comprising a processor, the processor being coupled to a memory, the memory storing at least one computer program instruction, the at least one computer program instruction being loaded and executed by the processor, so that the computer device implements the method provided in the first aspect or any optional manner of the first aspect. The specific details of the computer device provided in the third aspect can be found in the first aspect or any optional manner of the first aspect, and will not be repeated here.

[0039] In a fourth aspect, a computer-readable storage medium is provided, wherein at least one instruction is stored in the storage medium. When the instruction is executed on a computer, the computer executes the method provided by the first aspect or any optional manner of the first aspect.

[0040] In a fifth aspect, a computer program product is provided, which includes one or more computer program instructions. When the computer program instructions are loaded and executed by a computer, the computer executes the method provided by the first aspect or any optional manner of the first aspect.

[0041] The embodiment of the present application provides a method for calculating the pledge rate of digital assets of on-chain customers based on the WOA-LSTM algorithm. The method determines the pledge rate of digital assets based on the LSTM model and determines the hyperparameters of the LSTM model based on the WOA algorithm, which can obtain a more accurate, reliable and feasible pledge rate for each type of collateral of different customers. In addition, by allowing each enterprise to be on-chain, the enterprise can use digital assets as collateral, and at the same time, the pledge process is automated through smart contracts, and the pledge rates of different digital assets of each customer are calculated through the WOA-LSTM model, so that the enterprise can realize the liquidity conversion of assets. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a flow chart of a method for calculating the pledge rate of a digital asset provided in an embodiment of the present application;

[0043] Figure 2 This is a flow chart of a method for predicting digital assets based on an LSTM model provided in an embodiment of the present application;

[0044] Figure 3 It is a structural diagram of an LSTM model provided in an embodiment of the present application;

[0045] Figure 4 It is a flowchart of optimizing the hyperparameters of the LSTM model based on the WOA algorithm provided in an embodiment of the present application;

[0046] Figure 5 It is a flow chart of a method for calculating the pledge rate of a digital asset provided in an embodiment of the present application;

[0047] Figure 6 It is a structural schematic diagram of a pledge rate calculation device for digital assets provided in an embodiment of the present application. DETAILED DESCRIPTION

[0048] In order to make the objectives, technical solutions and advantages of the present application clearer, the implementation methods of the present application will be further described in detail below with reference to the accompanying drawings.

[0049] The following is an explanation of some terminology concepts involved in the embodiments of the present application.

[0050] Digital assets: Digital assets are assets that exist in digital form, usually based on blockchain technology, and have certain value and functions. Common digital assets include: Bitcoin, Ethereum, stablecoins (legal currencies, gold and other cryptocurrencies), tokenized assets (assets that digitize and tokenize traditional assets through blockchain technology, such as stocks, bonds, real estate, etc.), etc.

[0051] Blockchain: Blockchain is a decentralized distributed ledger that is tamper-proof, secure and reliable. It combines distributed storage, peer-to-peer transmission, consensus mechanism, cryptography and other technologies to record transactions and information through a growing data block chain to ensure data security and transparency.

[0052] Smart Contract: A smart contract is a computer program that automatically executes contract terms. It runs on a blockchain platform and can execute and manage contracts directly between buyers and sellers without an intermediary. The execution of smart contracts is automatic and relies on preset conditions to trigger. Once these conditions are triggered, the smart contract will automatically execute the corresponding contract terms according to the rules defined when it was written.

[0053] LSTM model: A variant model based on RNN, which solves the problem of gradient disappearance during long sequence training and can predict subsequent time series based on historical time series. It can pass information that appeared at an earlier time very well through four gates and automatically learn which information to retain and which to forget.

[0054] WOA-LSTM model: The LSTM algorithm is optimized through the WOA optimization algorithm. The WOA optimization algorithm can optimize LSTM through global search, parameter optimization, accelerated convergence, etc., making LSTM's prediction of the pledge rate more accurate and reliable.

[0055] Loss function: In machine learning, the difference between the model's predicted value for a single sample and the true value is usually called loss. The smaller the loss, the better the model.

[0056] The following is an example of an application scenario of the embodiment of the present application.

[0057] In the context of digital transformation, digital assets have become an indispensable part of corporate assets, and the demand for flexible use of funds by enterprises is increasing. However, in the traditional credit model, enterprises need to provide physical collateral or guarantees, and the procedures are cumbersome and restricted. It is difficult for enterprises to fully utilize their digital assets under the traditional financial system. Existing pledged credit can only be pledged with deposits in banks, and the types of digital assets are severely limited, with low flexibility.

[0058] In order to solve the pain points in the above-mentioned traditional pledge credit, improve the efficiency and flexibility of corporate credit, and enable digital assets to play an important role in corporate credit, the embodiment of the present application proposes a method for calculating the pledge rate of digital assets of chain customers based on the WOA-LSTM algorithm, combining blockchain technology, smart contracts and digital assets to provide legal persons with more diverse, flexible and efficient loan methods. The increase in the types of digital assets will make it difficult to define the pledge rate of each asset. Therefore, the embodiment of the present application also proposes a WOA-LSTM pledge rate calculation method. Compared with the original solution of fixed pledge rate in pledge credit, the WOA-LSTM model can perform data analysis on different digital assets of different customers, and obtain a more accurate, reliable and feasible pledge rate for each digital asset of different customers. By allowing each enterprise to go on the chain, the enterprise can use digital assets as digital assets, and at the same time, through the smart contract to automate the pledge process, the pledge rate of different digital assets of each customer is calculated through the WOA-LSTM model, so that the enterprise can realize the liquidity conversion of assets. This method expands the existing business model of pledged credit, allowing enterprises to choose digital assets more flexibly and significantly increase customer stickiness.

[0059] Considering that blockchain has actual product implementation and application in credit, the method for calculating the pledge rate of digital assets of customers on the chain based on the WOA-LSTM algorithm proposed in the embodiment of the present application is reasonable and feasible. The expansion of digital assets will inevitably lead to an increase in risk. In order to reduce the risk, the most effective way is to give a suitable pledge rate. Therefore, the embodiment of the present application also proposes a pledge rate calculation method based on WOA-LSTM, which takes into account the customer's financial status, historical loan records, pledge types and other attributes, predicts the pledge rates of different digital assets of customers through LSTM, and uses the WOA algorithm to initialize and optimize the initial parameters of the LSTM model.

[0060] The embodiments of the present application broaden the types of digital assets, allowing customers to choose more digital assets when they are engaged in pledge credit. For example, real assets (such as real estate, artworks, and precious metals) can be tokenized and used as digital assets on the chain, or customers can be allowed to use high-value NFTs as digital assets, such as digital art, virtual real estate, etc. At the same time, through the pledge rate calculation method of WOA-LSTM, a suitable pledge rate is given, reducing the risk of bad debt loans.

[0061] Please refer to the attached Figure 1 , attached Figure 1 A method for calculating the pledge rate of a digital asset provided in an embodiment of the present application mainly includes the following steps.

[0062] Step 1, data collection: obtain the historical basic data of customers and the data of digital assets. The data of digital assets includes price, trading volume, volatility and historical data.

[0063] Step 2, data preprocessing: clean, denoise and normalize the collected data to improve the quality and consistency of the data.

[0064] Step 3, extract the target features of the digital asset from the preprocessed data. The target features include the properties of the digital asset itself that affect the pledge rate. The target features of the digital asset are input into the LSTM model, the pledge rate is predicted using the LSTM model, the predicted pledge rate is adjusted in combination with other factors (such as market conditions), the final pledge rate is determined, and the WOA algorithm is used to optimize the parameters and structure of the LSTM model to accelerate convergence and improve performance.

[0065] Step 4, smart contract execution: The predicted pledge rate is written into the smart contract, and the blockchain system automatically pledges or unpledges the digital assets based on the pledge rate to determine the final pledge rate level.

[0066] Please refer to Figure 2 , Figure 2 This is a flow chart of a method for predicting digital assets based on an LSTM model provided in an embodiment of the present application. The LSTM model prediction process includes the following steps.

[0067] Step 1: Extract balance sheets, income statements, cash flows, industries, operating conditions, employee conditions, credit history, and capital structures at multiple times from the digital asset data as the first eigenvector.

[0068] The data of digital assets is the original input data. The data of digital assets include indicators such as price, trading volume, volatility, position distribution, and on-chain activity.

[0069] Step 2: Determine the correlation between the multiple moments and the pledge rate prediction based on the historical net value of the digital asset at the multiple moments; determine k key moments from the multiple moments based on the correlation between the multiple moments and the pledge rate prediction; and determine the feature set k*m (hereinafter referred to as the second feature vector) that affects the current digital asset pledge rate at each historical key moment.

[0070] k refers to the number of key moments selected from the digital asset history data. Key moments are representative or critical time points in the digital asset history data in the form of time series. By selecting the first k key moments, we try to capture the key changes or features in the digital asset history.

[0071] m is the characteristic dimension corresponding to each key moment. This usually refers to multiple characteristic variables related to the digital asset at that key moment, such as price, trading volume, market volatility, etc.

[0072] The k*m feature set is a matrix, where k rows correspond to the selected key moments (each key moment corresponds to a time point) and m columns correspond to the feature dimensions of each key moment (describing multiple attributes of each key moment).

[0073] Therefore, the k*m feature set represents that, at the selected k key moments, it contains m feature information corresponding to each key moment, and finally forms a feature matrix containing the relevant information of the key moments as the input of the LSTM model.

[0074] The attributes of digital assets that affect the pledge rate are a large-scale feature set, and the key moments of digital assets are to select the most representative and predictive moments (key moments) from these attribute data for model training and prediction. By extracting key moments, the LSTM model can be more focused on these key moments that affect the pledge rate, thereby improving the prediction accuracy and efficiency of the model. Key moment correlation: refers to the correlation between the selected key moments and the current pledge rate prediction target, that is, how the characteristics of these key moments affect the prediction of the pledge rate. The more correlated the key moments are, the greater the influence on the prediction.

[0075] Step 3: Through convolutional neural network processing, the k*m feature set (second eigenvector) obtained in step 2 is reduced in dimension to obtain a vector (x1, x2, x3, ..., xi) (hereinafter referred to as the third eigenvector).

[0076] Step 4: Concatenate the features determined by the digital asset itself (the first feature vector) with the features obtained in step 3 (the third feature vector), and combine the features using tools such as feature tool (multiply two or more features or calculate the Cartesian product to form a synthetic feature).

[0077] Step 5: Determine the labels of the training data and divide the training data into training set, validation set, and prediction set according to the ratio of 6:2:2.

[0078] Step 6: Select the model and adjust the labels of customers who have overdue payments through iteration.

[0079] Training set data determination

[0080] The training set data includes training set features and training set labels, where the training set features include the attributes of the digital assets themselves that affect the pledge rate and the factors affecting the pledge rate of the customer due to the historical net value of the digital assets.

[0081] (1) The properties of digital assets that affect the pledge rate

[0082] The information table of digital assets includes balance sheet, profit and loss statement, cash flow, industry, operating status, employee status, credit history, capital structure, etc. For each pledge rate calculation of each customer, this solution forms a feature vector (timestamp, balance sheet, profit, cash flow, industry, operating status, employee status, ...).

[0083] (2) Impact of the historical net value of digital assets

[0084] This plan believes that the historical net value of digital assets will also affect the pledge rate of the digital assets. For example, the historical net value of a customer's digital assets was 200,000 three years ago, 250,000 two years ago, and 220,000 one year ago. If a customer wants to pledge a loan, the historical net value has an important impact on the pledge rate.

[0085] Therefore, this solution will determine the correlation between the multiple moments and the pledge rate prediction through the historical net value of the digital asset at the multiple moments, sort the multiple moments according to the correlation, select the first k key moments from the multiple moments, determine the characteristics of the digital asset at the first k key moments, and obtain a k*m feature set (second feature vector), and use a convolutional neural network to reduce the dimension of the k*m feature set (second feature vector): the screened k*m feature set is input into the convolutional neural network (CNN) for extraction and dimension reduction, and a 1*s feature set is obtained. After processing, the original two-dimensional data set is abstracted into features with higher information content (third feature vector). The reduced-dimensional features (third feature vector) and the original (timestamp, assets and liabilities, profits, cash flow, industry, operating conditions, employee conditions, ...) feature vectors (first feature vectors) are spliced ​​together to obtain the final feature set of the pledge rate of a single digital asset. (hereinafter referred to as "factors affecting the historical net value of digital assets")

[0086] Training set labels

[0087] For each historical net value of each sample digital asset, this solution marks it based on the overdue records of the sample digital asset.

[0088] Label is the real result or target value used to train the model in supervised learning. In this solution, the label represents the final credit limit of each digital asset. For the historical net value of each digital asset, the setting of the label is based on the credit status (whether it is overdue) and the historical net value of the digital asset. When there is an overdue record in a sample digital asset, the final credit limit (label) of the sample digital asset is set to the historical net value of the sample digital asset *σ; when there is no overdue, the final credit limit (label) of the sample digital asset is set to the historical net value of the sample digital asset. Through several rounds of iterations in the model, σ is iteratively updated to find the optimal σ value.

[0089] Among them, σ is an adjustment coefficient, and the initial value of σ is 0.8. This coefficient reflects the impact of overdue risk on the credit limit. σ is used to quantify the degree of impact of overdue on the credit limit.

[0090] Training set partitioning

[0091] The entire training set is divided into three parts, of which the first 60% is used as the training set for model training; 20% is used as the validation set to help with subsequent model parameter adjustment; and the last 20% is used as the prediction set to evaluate the performance of the model.

[0092] Feature Engineering

[0093] (1) Data preprocessing

[0094] The missing numerical feature values ​​of a single sample digital asset are filled with the average value of the historical feature data of the sample digital asset.

[0095] The bisection type features are converted into numerical features through one-hot encoding.

[0096] For numerical features with a large difference between the maximum and minimum values, the features are mapped to [0, 1] through normalization.

[0097] (2) Feature combination

[0098] Combine some known features that can be combined to produce special meanings, such as combining the profit and cash flow attributes of the past five years to obtain the company's development trend;

[0099] The feature tool is used to multiply or Cartesian product the features to obtain a brute force feature set, and the feature_importance function provided by machine learning is used to find the top K important features.

[0100] Model selection and parameter adjustment

[0101] (1) Model selection

[0102] The historical net value of digital assets is regarded as a time series, and LSTM is used as a training model to better achieve the goal.

[0103] Please refer to Figure 3 , Figure 3 : This is a structural diagram of an LSTM model provided by an embodiment of the present application. In the LSTM model, the forget gate (ft) is used to decide which information to retain and discard. The information from the previous hidden state and the current input information output a value between 0 and 1 after passing through the forget gate. The closer the information is to 1, the more it should be retained; the input gate (it) updates the cell state. The cell state (ct) records the current state after discarding unnecessary information and adding new information discovered. The output gate (ot) passes the hidden state and cell state to the next time step.

[0104] The data that is finally input into the model of this scheme is a two-dimensional array. The first dimension is the time dimension, which considers the impact of historical net value on digital assets. It is possible that a customer has a high net value in some months and a low net value in some months. The annual net value of each digital asset is a row, sorted in chronological order of the net value. The second dimension is the characteristic data of the single pledge rate calculation of a single digital asset, that is, the characteristic vector obtained above (timestamp, assets and liabilities, profit, cash flow, industry, operating conditions, employee conditions..., related customer influencing factor 1, related customer influencing factor 2...).

[0105] WOA Optimization Algorithm Optimization Process

[0106] The Whale Optimization Algorithm (WOA) is a method for optimizing specific problems by simulating the process of whale hunting. During the hunting process, the position of the prey is the best position, and the whale changes its position by contracting, circling and spiraling. The whale's circling and contracting process is as follows:

[0107] D=|C·X * (t)-X(t)|; Formula (1)

[0108] X(t+1)=X * (t)-A·d; Formula (2)

[0109] Where t is the current iteration number; X is the position vector; X* is the optimal position vector, and D is the parameter for measuring the position; A and C represent two control parameter vectors, which can be calculated by the following formulas (3) and (4).

[0110] A=2a·ra; Formula (3)

[0111] C = 2·r; Formula (4)

[0112] Among them, r is an arbitrary vector between 0 and 1, and the value of a is related to t and tMaxlter, and it keeps decreasing with the increase of the number of iterations. This is the embodiment of shrinking and wrapping behavior. The mathematical formula involved can be written as:

[0113]

[0114] Among them, tMaxlter represents the maximum number of iterations.

[0115] The formula for the whale to update its position in a spiral manner is as follows:

[0116] D′=|X * (t)-X(t)| Formula (6)

[0117] x(t+1)=D′·e bl ·cos(2πl)+X * (t) Formula (7)

[0118] Where D represents the distance between two positions; l is an arbitrary number between -1 and -2; b is a constant, which is normally a natural number 1. The specific calculation formula is:

[0119]

[0120] Since whales simultaneously perform shrinking, circling and spiraling to update their positions during the hunting process, the probability of shrinking, circling and spiraling is 1 / 2, that is:

[0121]

[0122] Among them, p∈[0,1].

[0123] In the process of hunting, whales will continuously search in the area where they are located to update their own position in order to reach the global optimal position. The mathematical formula needed in this process is:

[0124] D=|C·X rand -X| Formula (10)

[0125] X(t+1)=X rand -A·D Formula (11)

[0126] Among them, Xrand is the position vector, randomly selected from the current generation.

[0127] The way the whale updates its position is determined by the control parameter vector |A|. When the control parameter vector |A|>=1, the global optimal hyperparameter is sought through search. When the control parameter vector |A|<1, the shrinking surround method is used to seek the global optimal hyperparameter.

[0128] Please refer to Figure 4 , Figure 4 This is a flow chart of optimizing the hyperparameters of an LSTM model based on a WOA algorithm provided in an embodiment of the present application. The flow chart of optimizing the hyperparameters of an LSTM model based on a WOA algorithm includes the following steps.

[0129] Step 1: Initialize the dimension, number of iterations, and population of the LSTM model, and confirm the number of iterations, learning rate, number of neurons in the first hidden layer, and number of neurons in the second hidden layer of the LSTM algorithm.

[0130] Step 2: Randomly generate 10 populations and set the LSTM model with the hyperparameters corresponding to each population.

[0131] Step 3: Take the mean square error between the predicted pledge rate of the LSTM model and the actual pledge rate of the LSTM model as the fitness, calculate the fitness corresponding to each population, take the hyperparameter corresponding to the minimum fitness of each population as the optimal hyperparameter of this iteration, and compare it with the global optimal hyperparameter. If the optimal hyperparameter in this iteration makes the prediction effect of the LSTM model better than the prediction effect of the LSTM model under the global optimal hyperparameter, then use the optimal hyperparameter in this iteration to replace the global optimal hyperparameter.

[0132] Step 4: Start iteration and use the whale algorithm to update the four hyperparameters corresponding to the population.

[0133] Step 5: Repeat steps 3 and 4 until the number of iterations reaches the set maximum number of iterations.

[0134] Step 6: Output the number of iterations, learning rate, number of neurons in the first hidden layer, and number of neurons in the second hidden layer corresponding to the optimal hyperparameters as the hyperparameters of the optimized LSTM model.

[0135] Pledge rate calculation method

[0136] Based on some historical basic data of customers in our bank, some collected customer data and digital asset data, the LSTM model is used as input data. At the same time, in order to make the LSTM model pay more attention to global data and obtain a more accurate pledge rate, the WOA optimization algorithm is used to optimize the number of iterations, learning rate, number of neurons in the first hidden layer and number of neurons in the second hidden layer of the entire LSTM model when the LSTM model is iterated. Finally, the WOA-LSTM model is used to predict the pledge rate of the customer under different digital assets as the pledge rate of the customer's current financing.

[0137] The embodiment of the present application applies digital assets to pledge credit as digital assets, broadening the traditional pledge credit model.

[0138] The embodiments of the present application use smart contracts to lock and release digital assets, and issue and repay loan funds, thereby ensuring the security and reliability of transactions and reducing intermediary and transaction costs.

[0139] The embodiment of the present application utilizes the LSTM model to calculate the pledge rate of the customer for different digital assets based on the historical net value of different digital assets and the customer's own situation.

[0140] The embodiment of the present application uses the WOA optimization algorithm to optimize the LSTM training process, which can more accurately evaluate the credit of customers.

[0141] Figure 5 : is a flow chart of a method for calculating the pledge rate of a digital asset provided in an embodiment of the present application, the method comprising:

[0142] S510, extracting target features of the digital asset from the data of the digital asset, where the target features include attributes of the digital asset itself that affect the pledge rate;

[0143] S520, using an LSTM model to predict the pledge rate of the digital asset based on the target features of the digital asset, wherein the training set data of the LSTM model is determined based on the attributes of the multiple sample digital assets that affect the pledge rate, the historical net value of the multiple sample digital assets, and the overdue records of the multiple sample digital assets, and the hyperparameters of the LSTM model are determined based on the WOA algorithm;

[0144] S530, writing the predicted pledge rate into the smart contract, and the blockchain system performs pledge or unpledge on the digital assets based on the pledge rate.

[0145] In some implementations, extracting target features of a digitized asset from data of the digitized asset includes:

[0146] Extract the balance sheet, income statement, cash flow, industry, operating conditions, employee status, credit history and capital structure at multiple times from the digital asset data to obtain the first eigenvector;

[0147] Determine the correlation between multiple moments and the pledge rate prediction based on the historical net value of the digital asset at multiple moments;

[0148] According to the correlation between multiple moments and the pledge rate prediction, k key moments are determined from the multiple moments, and the correlation between the k key moments and the pledge rate prediction is ranked in the top k among the multiple moments;

[0149] Extract the balance sheet, income statement, cash flow, industry, operating conditions, employee status, credit history and capital structure of the digital asset at k key moments from the digital asset data to obtain the second eigenvector;

[0150] Using a convolutional neural network to reduce the dimension of the second eigenvector to obtain a third eigenvector;

[0151] The first feature vector and the third feature vector are concatenated to obtain the target feature of the digital asset.

[0152] In some embodiments, the training set data of the LSTM model includes labels. When there is an overdue record in a sample digital asset, the label corresponding to the sample digital asset is the product of the historical net value of the sample digital asset and the target coefficient. When there is no overdue record in a sample digital asset, the label corresponding to the sample digital asset is the historical net value of the sample digital asset, and the target coefficient is obtained by iterative updating in the LSTM model.

[0153] In some embodiments, the hyperparameters of the LSTM model include the number of iterations of the LSTM model, the learning rate, the number of neurons in the first hidden layer, and the number of neurons in the second hidden layer. The process of determining the hyperparameters of the LSTM model includes:

[0154] Randomly generate multiple populations, each population corresponds to a parameter value of the hyperparameter of the LSTM model;

[0155] Set up the LSTM model using the hyperparameter values ​​corresponding to each population;

[0156] The pledge rate predicted by the LSTM model and the actual pledge rate are used as fitness, the fitness corresponding to each population is determined, the minimum fitness is determined from the fitness corresponding to each population, and the parameter value of the hyperparameter corresponding to the minimum fitness is used as the optimal hyperparameter of this iteration;

[0157] Compare the optimal hyperparameters of this iteration with the global optimal hyperparameters in the historical iteration process. If the model prediction effect corresponding to the optimal hyperparameters of this iteration is better than the model prediction effect corresponding to the global optimal hyperparameters, use the optimal hyperparameters of this iteration to replace the global optimal hyperparameters.

[0158] Use the whale algorithm to update the parameter values ​​of the hyperparameters corresponding to each population, and use the updated parameter values ​​of the hyperparameters to set the LSTM model;

[0159] Repeat the steps of hyperparameter comparison, hyperparameter replacement, and parameter value update until the number of iterations reaches the set maximum number of iterations.

[0160] In some implementations, the parameter value of the hyperparameter corresponding to each population is updated using the whale algorithm, including:

[0161] Generate random numbers;

[0162] If the random number is less than the random number threshold and the length of the control parameter vector is less than the threshold, the parameter value of the hyperparameter is determined by shrinking and wrapping;

[0163] If the random number is less than the random number threshold, and the length of the control parameter vector is greater than or equal to the threshold, the parameter value of the hyperparameter is determined by searching;

[0164] If the random number is greater than or equal to the random number threshold, the parameter value of the hyperparameter is determined in a spiral manner.

[0165] Figure 6 : is a structural schematic diagram of a pledge rate calculation device 600 for digital assets provided in an embodiment of the present application, and the device 600 includes:

[0166] A feature extraction unit 610 is used to extract target features of the digital asset from the data of the digital asset, where the target features include the attributes of the digital asset itself that affect the pledge rate;

[0167] A prediction unit 620 is used to predict the pledge rate of the digital asset based on the target features of the digital asset using an LSTM model, wherein the training set data of the LSTM model is determined based on the attributes of the multiple sample digital assets that affect the pledge rate, the historical net value of the multiple sample digital assets, and the overdue records of the multiple sample digital assets, and the hyperparameters of the LSTM model are determined based on the WOA algorithm;

[0168] The writing unit 630 is used to write the predicted pledge rate into the smart contract, and the blockchain system performs pledge or unpledge on the digital assets based on the pledge rate.

[0169] In some embodiments, the feature extraction unit 610 is used to extract the balance sheet, income statement, cash flow, industry, operating conditions, employee status, credit history and capital structure at multiple moments from the data of the digital asset to obtain a first feature vector; determine the correlation between the multiple moments and the pledge rate prediction based on the historical net value of the digital asset at multiple moments; determine k key moments from the multiple moments based on the correlation between the multiple moments and the pledge rate prediction, and the correlation between the k key moments and the pledge rate prediction ranks in the top k among the multiple moments; extract the balance sheet, income statement, cash flow, industry, operating conditions, employee status, credit history and capital structure of the digital asset at k key moments from the data of the digital asset to obtain a second feature vector; use a convolutional neural network to reduce the dimension of the second feature vector to obtain a third feature vector; and concatenate the first feature vector and the third feature vector to obtain the target feature of the digital asset.

[0170] In some embodiments, the training set data of the LSTM model includes labels. When there is an overdue record in a sample digital asset, the label corresponding to the sample digital asset is the product of the historical net value of the sample digital asset and the target coefficient. When there is no overdue record in a sample digital asset, the label corresponding to the sample digital asset is the historical net value of the sample digital asset, and the target coefficient is obtained by iterative updating in the LSTM model.

[0171] In some embodiments, the hyperparameters of the LSTM model include the number of iterations of the LSTM model, the learning rate, the number of neurons in the first hidden layer, and the number of neurons in the second hidden layer. The device also includes: a parameter determination unit, which is used to randomly generate multiple populations, each population corresponds to a parameter value of a hyperparameter of the LSTM model; the parameter value of the hyperparameter corresponding to each population is used to set the LSTM model; the pledge rate predicted by the LSTM model and the actual pledge rate are used as fitness, the fitness corresponding to each population is determined, the minimum fitness is determined from the fitness corresponding to each population, and the parameter value of the hyperparameter corresponding to the minimum fitness is used as the optimal hyperparameter of this iteration; the optimal hyperparameter of this iteration is compared with the global optimal hyperparameter in the historical iteration process. If the model prediction effect corresponding to the optimal hyperparameter of this iteration is better than the model prediction effect corresponding to the global optimal hyperparameter, the optimal hyperparameter of this iteration is used to replace the global optimal hyperparameter; the parameter value of the hyperparameter corresponding to each population is updated using the whale algorithm, and the parameter value of the updated hyperparameter is used to set the LSTM model; the steps of hyperparameter comparison, hyperparameter replacement, and parameter value update are repeated until the number of iterations reaches the set maximum number of iterations.

[0172] In some embodiments, the parameter determination unit is used to generate a random number; if the random number is less than a random number threshold and the length of the control parameter vector is less than the threshold, a shrinking and wrapping method is used to determine the parameter value of the hyperparameter; if the random number is less than the random number threshold and the length of the control parameter vector is greater than or equal to the threshold, a search method is used to determine the parameter value of the hyperparameter; if the random number is greater than or equal to the random number threshold, a spiral method is used to determine the parameter value of the hyperparameter.

[0173] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

[0174] A refers to B, which means that A is the same as B or A is a simple variant of B.

[0175] The terms "first" and "second" in the description and claims of the embodiments of the present application are used to distinguish different objects, rather than to describe a specific order of objects, and cannot be understood as indicating or implying relative importance. For example, the first eigenvector and the second eigenvector are used to distinguish different eigenvectors, rather than to describe a specific order of eigenvectors, and cannot be understood as meaning that the first eigenvector is more important than the second eigenvector.

[0176] In the embodiments of the present application, unless otherwise specified, "at least one" means one or more, and "a plurality" means two or more. For example, a plurality of feature vectors means two or more feature vectors.

[0177] The above embodiments can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in accordance with the embodiment of the present application is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions may be transmitted from a website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode to another website site, computer, server or data center. The computer-readable storage medium may be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more available media integrated. The available medium may be a magnetic medium, (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state hard disk SolidState Disk (SSD)), etc.

[0178] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for calculating the pledge rate of digital assets, characterized in that: The method comprises: Extracting target features of the digital asset from the data of the digital asset, wherein the target features include attributes of the digital asset itself that affect the pledge rate; The LSTM model is used to predict the pledge rate of the digital asset based on the target features of the digital asset, wherein the training set data of the LSTM model is determined based on the attributes of the multiple sample digital assets that affect the pledge rate, the historical net value of the multiple sample digital assets, and the overdue records of the multiple sample digital assets, and the hyperparameters of the LSTM model are determined based on the WOA algorithm; The predicted pledge rate is written into the smart contract, and the blockchain system performs pledge or unpledge of the digital assets based on the pledge rate.

2. The method according to claim 1, characterized in that The step of extracting target features of the digital asset from the data of the digital asset comprises: Extract the balance sheet, income statement, cash flow, industry, operating conditions, employee status, credit history and capital structure at multiple times from the digital asset data to obtain the first eigenvector; Determining, based on the historical net value of the digital asset at the multiple times, the correlation between the multiple times and the pledge rate prediction; According to the correlation between the multiple moments and the pledge rate prediction, k key moments are determined from the multiple moments, and the correlation between the k key moments and the pledge rate prediction is ranked in the top k positions among the multiple moments; Extracting the balance sheet, income statement, cash flow, industry, operating conditions, employee conditions, credit history and capital structure of the digital asset at the k key moments from the data of the digital asset to obtain a second eigenvector; Using a convolutional neural network to reduce the dimension of the second eigenvector to obtain a third eigenvector; The first feature vector and the third feature vector are concatenated to obtain a target feature of the digitized asset.

3. The method according to claim 1, characterized in that The training set data of the LSTM model includes labels. When there is an overdue record in a sample digital asset, the label corresponding to the sample digital asset is the product of the historical net value of the sample digital asset and the target coefficient. When there is no overdue record in a sample digital asset, the label corresponding to the sample digital asset is the historical net value of the sample digital asset, and the target coefficient is obtained by iterative updating in the LSTM model.

4. The method according to claim 1, characterized in that: The hyperparameters of the LSTM model include the number of iterations, the learning rate, the number of neurons in the first hidden layer, and the number of neurons in the second hidden layer of the LSTM model. The process of determining the hyperparameters of the LSTM model includes: Randomly generate multiple populations, each population corresponding to a parameter value of a hyperparameter of the LSTM model; The LSTM model is set using the parameter values ​​of the hyperparameters corresponding to each of the populations; The pledge rate predicted by the LSTM model and the actual pledge rate are used as fitness, the fitness corresponding to each of the populations is determined, the minimum fitness is determined from the fitness corresponding to each of the populations, and the parameter value of the hyperparameter corresponding to the minimum fitness is used as the optimal hyperparameter of this iteration; Compare the optimal hyperparameters of this iteration with the global optimal hyperparameters in the historical iteration process. If the model prediction effect corresponding to the optimal hyperparameters of this iteration is better than the model prediction effect corresponding to the global optimal hyperparameters, replace the global optimal hyperparameters with the optimal hyperparameters of this iteration; Using the whale algorithm to update the parameter value of the hyperparameter corresponding to each of the populations, and using the updated parameter value of the hyperparameter to set the LSTM model; Repeat the steps of hyperparameter comparison, hyperparameter replacement, and parameter value update until the number of iterations reaches the set maximum number of iterations.

5. The method according to claim 1, characterized in that The method of updating the parameter value of the hyperparameter corresponding to each population by using the whale algorithm includes: Generate random numbers; If the random number is less than the random number threshold and the length of the control parameter vector is less than the threshold, the parameter value of the hyperparameter is determined by shrinking and wrapping; If the random number is less than the random number threshold, and the length of the control parameter vector is greater than or equal to the threshold, the parameter value of the hyperparameter is determined by searching; If the random number is greater than or equal to the random number threshold, the parameter value of the hyperparameter is determined in a spiral manner.

6. A pledge rate calculation device for digital assets, characterized in that: The device comprises: A feature extraction unit, used to extract target features of the digital asset from the data of the digital asset, wherein the target features include the attributes of the digital asset itself that affect the pledge rate; A prediction unit, configured to predict the pledge rate of the digital asset based on the target features of the digital asset using an LSTM model, wherein the training set data of the LSTM model is determined based on the attributes of a plurality of sample digital assets that affect the pledge rate, the historical net values ​​of the plurality of sample digital assets, and the overdue records of the plurality of sample digital assets, and the hyperparameters of the LSTM model are determined based on the WOA algorithm; A writing unit is used to write the predicted pledge rate into the smart contract, and the blockchain system performs pledge or unpledge on the digital asset based on the pledge rate.

7. The device according to claim 6, characterized in that The feature extraction unit is used to extract the balance sheet, income statement, cash flow, industry, operating conditions, employee status, credit history and capital structure at multiple times from the data of the digital asset to obtain a first feature vector; and determine the correlation between the multiple times and the pledge rate prediction according to the historical net value of the digital asset at the multiple times; According to the correlation between the multiple moments and the pledge rate prediction, k key moments are determined from the multiple moments, and the correlation between the k key moments and the pledge rate prediction is ranked in the top k among the multiple moments; the balance sheet, income statement, cash flow, industry, operating conditions, employee status, credit history and capital structure of the digital asset at the k key moments are extracted from the data of the digital asset to obtain a second eigenvector; a convolutional neural network is used to reduce the dimension of the second eigenvector to obtain a third eigenvector; the first eigenvector and the third eigenvector are concatenated to obtain the target feature of the digital asset.

8. The device according to claim 6, characterized in that The training set data of the LSTM model includes labels. When there is an overdue record in a sample digital asset, the label corresponding to the sample digital asset is the product of the historical net value of the sample digital asset and the target coefficient. When there is no overdue record in a sample digital asset, the label corresponding to the sample digital asset is the historical net value of the sample digital asset, and the target coefficient is obtained by iterative updating in the LSTM model.

9. The device according to claim 6, characterized in that The hyperparameters of the LSTM model include the number of iterations of the LSTM model, the learning rate, the number of neurons in the first hidden layer, and the number of neurons in the second hidden layer. The device also includes: a parameter determination unit, which is used to randomly generate a plurality of populations, each of which corresponds to a parameter value of a hyperparameter of the LSTM model; respectively use the parameter values ​​of the hyperparameters corresponding to each of the populations to set the LSTM model; use the pledge rate predicted by the LSTM model and the actual pledge rate as fitness, determine the fitness corresponding to each of the populations, determine the minimum fitness from the fitness corresponding to each of the populations, and use the hyperparameter corresponding to the minimum fitness as the fitness; The parameter value of the parameter is used as the optimal hyperparameter of this iteration; the optimal hyperparameter of this iteration is compared with the global optimal hyperparameter in the historical iteration process. If the model prediction effect corresponding to the optimal hyperparameter of this iteration is better than the model prediction effect corresponding to the global optimal hyperparameter, the optimal hyperparameter of this iteration is used to replace the global optimal hyperparameter; the parameter value of the hyperparameter corresponding to each of the populations is updated by using the whale algorithm, and the parameter value of the updated hyperparameter is used to set the LSTM model; the steps of hyperparameter comparison, hyperparameter replacement and parameter value update are repeated until the number of iterations reaches the set maximum number of iterations.

10. The device according to claim 6, characterized in that The parameter determination unit is used to generate a random number; if the random number is less than the random number threshold, and the length of the control parameter vector is less than the threshold, a shrinking wrap method is used to determine the parameter value of the hyperparameter; If the random number is less than the random number threshold and the length of the control parameter vector is greater than or equal to the threshold, the parameter value of the hyperparameter is determined by searching; if the random number is greater than or equal to the random number threshold, the parameter value of the hyperparameter is determined by spiral method.

Images