WO2020258994A1 - 一种节点数据预测方法及装置 - Google Patents

一种节点数据预测方法及装置 Download PDF

Info

Publication number
WO2020258994A1
WO2020258994A1 PCT/CN2020/084322 CN2020084322W WO2020258994A1 WO 2020258994 A1 WO2020258994 A1 WO 2020258994A1 CN 2020084322 W CN2020084322 W CN 2020084322W WO 2020258994 A1 WO2020258994 A1 WO 2020258994A1
Authority
WO
WIPO (PCT)
Prior art keywords
time period
data
unit time
training
node
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2020/084322
Other languages
English (en)
French (fr)
Inventor
李治
曾岩
李晶
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
WeBank Co Ltd
Original Assignee
WeBank Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by WeBank Co Ltd filed Critical WeBank Co Ltd
Publication of WO2020258994A1 publication Critical patent/WO2020258994A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING OR CALCULATING; COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/24Classification techniques
    • G06F18/241Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2411Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
    • GPHYSICS
    • G06COMPUTING OR CALCULATING; COUNTING
    • G06QINFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00Administration; Management
    • G06Q10/04Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • GPHYSICS
    • G06COMPUTING OR CALCULATING; COUNTING
    • G06QINFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q40/00Finance; Insurance; Tax strategies; Processing of corporate or income taxes
    • G06Q40/02Banking, e.g. interest calculation or account maintenance
    • GPHYSICS
    • G06COMPUTING OR CALCULATING; COUNTING
    • G06QINFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q40/00Finance; Insurance; Tax strategies; Processing of corporate or income taxes
    • G06Q40/03Credit; Loans; Processing thereof

Definitions

  • the present invention relates to the field of machine learning in Fintech, and in particular to a method and device for predicting node data.
  • Bank position is a professional term used by the banking system for the scheduling of available funds. It refers to the sum of all funds currently available to the bank, mainly including excess reserves in the central bank, net deposits of inter-bank liquidation funds, bank deposits and cash, etc. section. Every bank or securities industry has its own capital position.
  • bank position management is to minimize the opportunity cost of existing positions while maintaining liquidity.
  • Bank positions should keep the total amount of positions moderate, that is, the total amount of bank positions must be kept on a moderate scale.
  • a multiple linear regression analysis method is generally used to predict the reserved data in the node, and the weight coefficient of each factor is obtained by linear regression by selecting factors that may affect the reserved data. This method has large errors in the prediction results, and the correlation between different factors interferes with each other, which further increases the prediction error.
  • This application provides a node data prediction method and device, which are used to reduce the prediction error of node data and ensure the appropriateness of reserved data.
  • a node data prediction method provided by an embodiment of the present invention includes:
  • the reserved data in the node is determined.
  • the reference time period includes N reference unit time periods
  • the determining the reference data of the reference unit time period according to the initial data in the reference unit time period includes:
  • the method before acquiring the model parameters of the SVM model, the method further includes:
  • the training data in the historical time period is used as a parameter and input into the SVM model to determine the model parameters of the SVM model.
  • the training data in the historical time period is used as a parameter and input into the SVM model.
  • the model parameters of the SVM model are determined, the Before the reserved data in the node, it also includes:
  • training unit time period in the historical time period input the training data before the training unit time period and the model parameters into the SVM model to determine the prediction data of the training unit time period; Comparing the prediction data of the training unit time period with the training data of the training unit time period to determine the error data of the training unit time period;
  • the determining the reserved data in the node according to the predicted data includes:
  • the reserved data in the node is determined.
  • an embodiment of the present invention also provides a node data prediction device, including:
  • the acquiring unit is used to acquire the initial data of the node in the reference time period
  • a calculation unit configured to determine the reference data of the reference unit time period according to the initial data in the reference unit time period for any reference unit time period in the reference time period;
  • the obtaining unit is also used to obtain model parameters of the support vector machine SVM model
  • a prediction unit configured to input the model parameters and all reference data in the reference time period into the SVM model to calculate the predicted data in the prediction time period;
  • the control unit is configured to determine the reserved data in the node according to the predicted data.
  • the reference time period includes N reference unit time periods
  • the calculation unit is specifically used for:
  • a training unit is further included for:
  • the training data in the historical time period is used as a parameter and input into the SVM model to determine the model parameters of the SVM model.
  • an error unit is further included for:
  • training unit time period in the historical time period input the training data before the training unit time period and the model parameters into the SVM model to determine the prediction data of the training unit time period; Comparing the prediction data of the training unit time period with the training data of the training unit time period to determine the error data of the training unit time period;
  • the control unit is also used for:
  • the reserved data in the node is determined.
  • an embodiment of the present invention also provides an electronic device, including:
  • At least one processor and,
  • a memory communicatively connected with the at least one processor; wherein,
  • the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor, so that the at least one processor can execute the method for predicting node data as described above.
  • the embodiment of the present invention also provides a non-transitory computer-readable storage medium, the non-transitory computer-readable storage medium stores computer instructions, and the computer instructions are used to make the computer execute the node described above Data prediction method.
  • the embodiments of the present invention also provide a computer program product
  • the computer program product includes a computer program stored on a non-transitory computer-readable storage medium
  • the computer program includes program instructions, when the program instructions When executed by a computer, the computer is caused to execute the node data prediction method described above.
  • the initial data of the node in the reference time period is acquired, for any reference unit time period in the reference time period, the reference data of the reference time period is determined according to the initial data in the reference unit time period, so that the reference All reference data in the time period are related to time and form time series data.
  • the model parameters of the SVM model are obtained. Input the model parameters and all reference data in the reference time period into the SVM model, and calculate the forecast data in the forecast time period. Based on the predicted data, the reserved data in the node is determined.
  • the embodiment of the present invention uses the SVM model to predict the prediction time period, and can perform real-time reservation data prediction, and the determined reference data is time series data.
  • the reference data is used as the input of the SVM model with high accuracy. At the same time, only reference data is used, the amount of calculation is small, and the data source requirements are simple, which reduces the calculation pressure.
  • Figure 1 is a schematic structural diagram of a possible system architecture provided by an embodiment of the present invention.
  • FIG. 2 is a schematic flowchart of a method for predicting node data provided by an embodiment of the present invention
  • FIG. 3 is a schematic flowchart of a method for predicting node data according to a specific embodiment of the present invention
  • FIG. 4 is a schematic structural diagram of a node data prediction device provided by an embodiment of the present invention.
  • FIG. 5 is a schematic structural diagram of an electronic device provided by an embodiment of the present invention.
  • a system architecture to which the embodiment of the present invention is applicable includes a position node 101, a position management device 102, and a position prediction device 103.
  • the position node 101, the position management device 102, and the position prediction device 103 may be network devices such as computers, may be an independent device, or may be a server cluster formed by multiple servers.
  • the position node 101, the position management device 102, and the position prediction device 103 may use cloud computing technology for information processing.
  • the position prediction device 103 is an auxiliary device of the position management device 102, which is used to predict the positions of each position node 101, so as to give the position management device 102 a reference, so that the position management device 102 can adjust the position according to the real-time prediction of the position prediction device 103 Position data in node 101.
  • SVM Support Vector Machine (Support Vector Machine), a classification and regression algorithm based on machine learning, is a generalized linear classifier that performs binary classification of data in a supervised learning manner, and its decision boundary is the maximum solution to the learning sample Margin hyperplane.
  • Pattern recognition scenarios such as portrait recognition and text classification.
  • Spark A computing framework under an open source distributed platform.
  • Spark Mlib A set of machine learning development kits based on the Spark computing framework.
  • Position prediction Predict the inflow and outflow of funds from each position node, and provide a clear quantitative data reference for the business to avoid insufficient funds or excessive position arrangements.
  • the reserved data of the node in the embodiment of the present invention is introduced by taking position data as an example.
  • Time series refers to the sequence of numbers of the same statistical indicator arranged in the order of the time of occurrence.
  • the main purpose of time series analysis is to predict the future based on existing historical data. Most of the economic data are given in the form of time series.
  • the time in the time series can be year, quarter, month or any other time format.
  • an embodiment of the present invention provides a method for predicting node data. As shown in FIG. 2, the method for predicting node data provided by the embodiment of the present invention includes the following steps:
  • Step 201 Obtain initial data of a node in a reference time period.
  • the initial data in the node can be real-time incoming data or outgoing data at each time point in the reference time period in the node, and the incoming and outgoing data can be marked with a sign.
  • the initial data at a certain time point is positive. It indicates inflow data; if it is negative, it indicates outflow data.
  • the reference time period here is the time period before the prediction time period, that is, the data in the prediction time period is predicted based on the data in the reference time period.
  • Table 1 shows the real-time capital inflow data of position node 1.
  • Step 202 For any reference unit time period in the reference time period, determine the reference data of the reference unit time period according to the initial data in the reference unit time period.
  • the initial data is the fund flow data at each time point
  • the reference data is the fund flow data of the node in each reference unit time period.
  • the reference unit time period can be one minute, one quarter of an hour, one day, etc., generally as needed
  • the reference unit time period is one hour, that is, according to the inflow or outflow data of each time point in the node in each hour, the fund flow situation of the node every hour is determined.
  • Step 203 Obtain model parameters of the SVM model.
  • the SVM model can be called from the Spark Mlib package, or it can be implemented using the development language python.
  • the model parameters of the SVM model can be calculated in advance and can be obtained directly when prediction is required; or during the prediction process, the model parameters are also calculated and corrected in real time to improve the accuracy of model calculation.
  • Step 204 Input the model parameters and all reference data in the reference time period into the SVM model, and calculate the predicted data in the predicted time period.
  • the initial data of the node in the reference time period is acquired, for any reference unit time period in the reference time period, the reference data of the reference time period is determined according to the initial data in the reference unit time period, so that the reference All reference data in the time period are related to time and form time series data.
  • the model parameters of the SVM model are obtained. Input the model parameters and all reference data in the reference time period into the SVM model, and calculate the forecast data in the forecast time period. Based on the forecast data, determine the position data in the node.
  • the embodiment of the present invention uses the SVM model to predict the forecast time period, and can perform real-time position data forecasting, and the determined reference data is time series data.
  • the reference time period includes N reference unit time periods;
  • the determining the reference data of the reference unit time period according to the initial data in the reference unit time period includes:
  • the initial data at each time point in the reference unit time period is first added, and the positive and negative initial data cancel each other out.
  • Table 2 shows the fund flow data of position node 1 in each unit time period.
  • Position node period Inflow amount Position node 1 20190101 11 o'clock-12 o'clock 12323432.32 Position node 1 20190101 10:00-11:00 34534534.67 Position node 1 20190101 9:00-10:00 12312342.45 Position node 1 20190101 8:00-9:00 32132423.00 ... ... ...
  • Table 3 shows the reference data obtained after logarithmic processing in the reference time period.
  • the reference data obtained after logarithmic processing indicates the rate of change of each unit time period relative to the previous unit time period. Inputting this rate of change into the SVM algorithm model improves the accuracy of prediction.
  • the model parameters need to be trained. Before acquiring the model parameters of the SVM model, it further includes:
  • the training data in the historical time period is used as a parameter and input into the SVM model to determine the model parameters of the SVM model.
  • the change rate after logarithmic processing calculated from the initial data described above is used as the training data.
  • the training data is the node in the training unit time period.
  • the training data in each unit time period is used as the input parameter of the SVM model, and the time series data corresponding to each unit time period is used as the output parameter of the SVM model.
  • the time series data corresponding to one unit time period is the unit time
  • the training data corresponding to the preset number of unit time periods before the segment is arranged in the order of time from first to last; obviously, after obtaining a large number of corresponding relationships between the input parameters and output parameters, based on the existing SVM core
  • the function training method can obtain the model parameters of the SVM model.
  • the historical time period corresponding to the training process and the reference time period corresponding to the prediction process can be the same time period or different time periods. If the historical time period and the reference time period are different time periods, the two The time periods may or may not overlap.
  • the historical time period is 1000 hours before the current time point, and the reference time period is 999 hours before the current time point; or the historical time period is 9 am to 11 am every day from January to March 2018, reference The time period is from 9 am to 11 am every day from January to March 2019.
  • the selection of the historical time period and the reference time period is based on calculation requirements, and is not limited in the embodiment of the present invention.
  • the calculation method of the training data is similar to the calculation method of the above reference data, so I won't repeat it here.
  • the model parameters of the SVM model may be dynamic data or relatively static data. That is, when the model parameters are dynamic data, the model parameters are calculated in real time every time the position data is predicted, so that the accuracy of the model parameters is high, but it needs to be calculated in real time, and the server calculation pressure is relatively high.
  • the model parameters are relatively static data, the model parameters can be calculated and stored in advance. When position data prediction is required, they can be directly obtained from the memory. At this time, the model parameters can be performed at a relatively long period of time, such as one month or six months. Update once. In this way, the server does not need to calculate model parameters in real time, but its accuracy is low.
  • the time series data corresponding to the preset time period is the preset number of unit time periods before the unit time period.
  • the training data is arranged in the order of time from first to last) as the input parameter input through the SVM model trained above, the SVM model trained above can calculate the pair at a moment after the preset time period
  • the rate of change after digitization processing is used as prediction data.
  • Step 205 Determine the reserved data in the node according to the predicted data.
  • the reservation can be calculated after the inflow of funds and the predicted data for the preset time period are known. Data.
  • the bank node will predict the difference between the outflow and inflow funds in the node on the day based on the flow of funds and experience in the node in the previous few days, so that a certain amount of funds will be reserved in the account before the day of business. That is the position. Therefore, in the embodiment of the present invention, the bank node can determine the reserved data, that is, the position data, based on the forecast data.
  • the embodiment of the present invention compares the predicted data with the true value in the historical time to determine the error data. Said using the training data in the historical time period as a parameter and inputting it into the SVM model, after determining the model parameters of the SVM model, before determining the reserved data in the node according to the prediction data ,Also includes:
  • training unit time period For any training unit time period in the historical time period, input the training data of the training unit time period before the training unit time period and the model parameters into the SVM model to determine the training unit time period Compare the prediction data of the training unit time period with the training data of the training unit time period to determine the error data of the training unit time period;
  • the determining the reserved data in the node according to the predicted data includes:
  • the reserved data in the node is determined.
  • the position data in each training unit time period in the historical time period is predicted, and multiple forecast data in the historical time period are obtained.
  • the method of specifically calculating the prediction data in each training unit time period is similar to the method of calculating the reference data.
  • the historical time period includes N training unit time periods, which are respectively denoted as T1, T2,... TN.
  • T1 input the training data before T1 and the model parameters into the SVM model to obtain the predicted data of T1, denoted as X1; subtract X1 and the real data of T1, that is, the training data of T1, to obtain the error data of T1.
  • the error data of T2 to the error data of TN are obtained according to the above method.
  • the mean value ⁇ and variance ⁇ of the error data can be calculated as the historical error of the SVM model.
  • the calculated forecast data is a forecast value.
  • the bank node if the bank node directly equals the actual position to the forecast data, the error is not taken into account and the unexpected situation cannot be handled. Therefore, often the bank node will increase the amount of data on the basis of the forecast data, that is, the actual position data will be larger than the forecast data. Therefore, in the embodiment of the present invention, the bank node is provided with a reference value for increasing the data.
  • the error range can be obtained according to the mean and/or variance of the error data. The number of error ranges can be selected according to needs.
  • error ranges are selected, ⁇ , ⁇ + ⁇ , and ⁇ +2 ⁇ , respectively. Under demand conditions, one of these three error ranges can be selected plus forecast data as the final position data of the node.
  • the error range in the embodiment of the present invention may also be ⁇ and 2 ⁇ , or such as ⁇ , 2 ⁇ , 3 ⁇ , and 4 ⁇ , which are only examples here and are not limited.
  • Step S301 The position prediction device obtains the initial data of the position node 1 999 hours before the current time point from the position management device.
  • Step S302 For any hour, the position forecasting device adds up the initial data of the hour to obtain the weighted data of the hour, and compares it with the weighted data of the previous hour of the hour, and takes the logarithm of the ratio to obtain Reference data for the hour.
  • Step S303 The position prediction device obtains the initial data of the position node 1 within 1000 hours before the current time point from the position management device.
  • Step S304 For any hour, the position forecasting device adds up the initial data of the hour to obtain the weighted data of the hour, and compares it with the weighted data of the previous hour of the hour, and takes the logarithm of the ratio to obtain Training data for the hour.
  • Step S305 The position prediction device uses the training data of each hour in the 1000 hours before the current time point as a parameter, and inputs it into the SVM algorithm model for training, to obtain the model parameters of the SVM algorithm model.
  • the kernel function adopts Gaussian radial basis function as the SVM kernel function.
  • Step S306 The position prediction device inputs 999 reference data into the trained SVM algorithm model for calculation, and obtains the rate of change data, which is recorded as V.
  • Step S307 The position forecasting device determines the reference data within one hour before the current time point, denoted as A, and calculates the forecast data Y according to the change rate data V and the reference data A.
  • the forecast data Y is calculated according to the following formula:
  • Step S308 The position prediction device obtains the error range.
  • Step S309 The position prediction device sends the error range and the prediction data Y to the position management device.
  • Step S310 The position management device determines the position data of the position node 1 according to the error range and the forecast data.
  • the embodiment of the present invention also provides a node data prediction device, as shown in FIG. 4, including:
  • the acquiring unit 401 is configured to acquire the initial data of the node in the reference time period
  • the calculation unit 402 is configured to determine the reference data of the reference unit time period according to the initial data in the reference unit time period for any reference unit time period in the reference time period;
  • the obtaining unit 401 is also used to obtain model parameters of the support vector machine SVM model;
  • the prediction unit 403 is configured to input the model parameters and all the reference data in the reference time period into the SVM model to calculate the predicted data in the prediction time period;
  • the control unit 404 is configured to determine the reserved data in the node according to the predicted data.
  • the reference time period includes N reference unit time periods
  • the calculation unit 402 is specifically configured to:
  • a training unit 405 for:
  • the training data in the historical time period is used as a parameter and input into the SVM model to determine the model parameters of the SVM model.
  • training unit time period in the historical time period input the training data before the training unit time period and the model parameters into the SVM model to determine the prediction data of the training unit time period; Comparing the prediction data of the training unit time period with the training data of the training unit time period to determine the error data of the training unit time period;
  • the control unit 404 is also used for:
  • the reserved data in the node is determined.
  • the present invention also provides an electronic device, as shown in FIG. 5, including:
  • processor 501 Including a processor 501, a memory 502, a transceiver 503, and a bus interface 504, wherein the processor 501, the memory 502 and the transceiver 503 are connected through the bus interface 504;
  • the processor 501 is configured to read the program in the memory 502 and execute the following methods:
  • the reserved data in the node is determined.
  • the reference time period includes N reference unit time periods
  • the processor 501 is specifically configured to:
  • the processor 501 is specifically configured to:
  • the training data in the historical time period is used as a parameter and input into the SVM model to determine the model parameters of the SVM model.
  • the processor 501 is specifically configured to:
  • training unit time period in the historical time period input the training data before the training unit time period and the model parameters into the SVM model to determine the prediction data of the training unit time period; Comparing the prediction data of the training unit time period with the training data of the training unit time period to determine the error data of the training unit time period;
  • the reserved data in the node is determined.
  • the embodiments of the present invention also provide a non-transitory computer-readable storage medium, the non-transitory computer-readable storage medium stores computer instructions, and the computer instructions are used to make the computer execute the above The node data prediction method.
  • the embodiments of the present invention also provide a computer program product.
  • the computer program product includes a computer program stored on a non-transitory computer-readable storage medium.
  • the computer program includes program instructions. When the program instructions are executed by a computer, the computer executes the node data prediction method.
  • These computer program instructions can also be stored in a computer-readable memory that can guide a computer or other programmable data processing equipment to work in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture including the instruction device.
  • the device implements the functions specified in one process or multiple processes in the flowchart and/or one block or multiple blocks in the block diagram.
  • These computer program instructions can also be loaded on a computer or other programmable data processing equipment, so that a series of operation steps are executed on the computer or other programmable equipment to produce computer-implemented processing, so as to execute on the computer or other programmable equipment.
  • the instructions provide steps for implementing functions specified in a flow or multiple flows in the flowchart and/or a block or multiple blocks in the block diagram.

Landscapes

  • Engineering & Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • Theoretical Computer Science (AREA)
  • Accounting & Taxation (AREA)
  • Finance (AREA)
  • Economics (AREA)
  • Strategic Management (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Development Economics (AREA)
  • Marketing (AREA)
  • General Business, Economics & Management (AREA)
  • Human Resources & Organizations (AREA)
  • Data Mining & Analysis (AREA)
  • Technology Law (AREA)
  • Artificial Intelligence (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Quality & Reliability (AREA)
  • Tourism & Hospitality (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Game Theory and Decision Science (AREA)
  • Operations Research (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

涉及机器学习领域,尤其涉及一种节点数据预测方法及装置,用以减小节点数据的预测误差,保证预留数据的适度性。该方法包括:获取参考时间段内节点的初始数据(201);针对所述参考时间段内的任一参考单位时间段,根据所述参考单位时间段内的初始数据,确定所述参考单位时间段的参考数据(202);获取SVM模型的模型参数(203);将所述模型参数以及所述参考时间段内的所有参考数据,输入所述SVM模型,计算出预测时间段的预测数据(204);根据所述预测数据,确定所述节点中的预留数据(205)。

Description

一种节点数据预测方法及装置
相关申请的交叉引用
本申请要求在2019年06月24日提交中国专利局、申请号为201910549462.1、申请名称为“一种节点数据预测方法及装置”的中国专利申请的优先权,其全部内容通过引用结合在本申请中。
技术领域
本发明涉及金融科技(Fintech)中的机器学习领域,尤其涉及一种节点数据预测方法及装置。
背景技术
随着金融科技,尤其是互联网科技金融的不断发展,越来越多的技术(如分布式、区块链Blockchain、人工智能等)应用在金融领域,但金融业也对技术提出了更高的要求,如针对包括头寸数据在内的节点数据进行计算和预测。
头寸其实就是款项的意思,可指投资者拥有或借用的资金数量。银行头寸是银行系统对于可用资金调度的一个专业的叫法,指的是银行当前所有可以运用的资金的总和,主要包括在央行的超额准备金、存放同业清算款项净额、银行存款以及现金等部分。每个银行或者证券业都有自己的资金头寸。
银行如果头寸不能满足业务需要,出现客户挤兑或者无资金发放贷款,势必会失去客户,从而影响效益;而如果头寸相对积压,将会增加银行的资金机会成本。因此,银行头寸管理的主要任务是在保持资金流动性需要的前提下,将现有头寸的机会成本降到最低。银行头寸应保持头寸的总量适度,即银行头寸总量要保持在一个适度的规模上。
现有技术中一般利用多元线性回归分析的方法对节点中的预留数据进行预测,通过选取可能会对预留数据有影响的因素,通过线性回归得到每个因素的权重系数。这种方法预测结果误差较大,且不同因素之间会有相关性互相干扰,从而进一步增加了预测误差。
发明内容
本申请提供一种节点数据预测方法及装置,用以减小节点数据的预测误差,保证预留数据的适度性。
一方面,本发明实施例提供的一种节点数据预测方法,包括:
获取参考时间段内节点的初始数据;
针对所述参考时间段内的任一参考单位时间段,根据所述参考单位时间段内的初始数据,确定所述参考单位时间段的参考数据;
获取SVM模型的模型参数;
将所述模型参数以及所述参考时间段内的所有参考数据,输入所述SVM模型,计算出预测时间段的预测数据;
根据所述预测数据,确定所述节点中的预留数据。
一种可选的实施例中,所述参考时间段内包括N个参考单位时间段;
针对第i个参考单位时间段,所述根据所述参考单位时间段内的初始数据,确定所述参考单位时间段的参考数据,包括:
将所述参考单位时间段内的初始数据相加,确定所述参考单位时间段的加权数据;
将第i个参考单位时间段的加权数据与所述第i-1个参考时间段的加权数据相比,并将比值取对数,得到所述第i个参考单位时间段的参考数据;其中1<i≤N。
一种可选的实施例中,所述获取SVM模型的模型参数之前,还包括:
获取历史时间段内节点的初始数据;
针对所述历史时间段内的任一训练单位时间段,根据所述训练单位时间段内的初始数据,确定所述训练单位时间段的训练数据;
将所述历史时间段内的训练数据作为参数,输入所述SVM模型中,确定所述SVM模型的模型参数。
一种可选的实施例中,所述将所述历史时间段内的训练数据作为参数,输入所述SVM模型中,确定所述SVM模型的模型参数之后,所述根据所述预测数据,确定所述节点中的预留数据之前,还包括:
针对所述历史时间段内的任一训练单位时间段,将所述训练单位时间段之前的训练数据,以及所述模型参数输入所述SVM模型,确定所述训练单位时间段的预测数据;将所述训练单位时间段的预测数据与所述训练单位时间段的训练数据相比较,确定所述训练单位时间段的误差数据;
根据所述历史时间段内所有训练单位时间段的误差数据,确定所述SVM模型的历史误差;
所述根据所述预测数据,确定所述节点中的预留数据,包括:
根据所述预测数据以及所述SVM模型的历史误差,确定所述节点中的预留数据。
一方面,本发明实施例还提供一种节点数据预测装置,包括:
获取单元,用于获取参考时间段内节点的初始数据;
计算单元,用于针对所述参考时间段内的任一参考单位时间段,根据所述参考单位时间段内的初始数据,确定所述参考单位时间段的参考数据;
所述获取单元,还用于获取支持向量机SVM模型的模型参数;
预测单元,用于将所述模型参数以及所述参考时间段内的所有参考数据,输入所述SVM模型,计算出预测时间段的预测数据;
控制单元,用于根据所述预测数据,确定所述节点中的预留数据。
一种可选的实施例中,所述参考时间段内包括N个参考单位时间段;
针对第i个参考单位时间段,所述计算单元,具体用于:
将所述参考单位时间段内的初始数据相加,确定所述参考单位时间段的加权数据;
将第i个参考单位时间段的加权数据与所述第i-1个参考时间段的加权数据相比,并将比值取对数,得到所述第i个参考单位时间段的参考数据;其中1<i≤N。
一种可选的实施例中,还包括训练单元,用于:
获取历史时间段内节点的初始数据;
针对所述历史时间段内的任一训练单位时间段,根据所述训练单位时间段内的初始数据,确定所述训练单位时间段的训练数据;
将所述历史时间段内的训练数据作为参数,输入所述SVM模型中,确定所述SVM模型的模型参数。
一种可选的实施例中,还包括误差单元,用于:
针对所述历史时间段内的任一训练单位时间段,将所述训练单位时间段之前的训练数据,以及所述模型参数输入所述SVM模型,确定所述训练单位时间段的预测数据;将所述训练单位时间段的预测数据与所述训练单位时间段的训练数据相比较,确定所述训练单位时间段的误差数据;
根据所述历史时间段内所有训练单位时间段的误差数据,确定所述SVM模型的历史误差;
所述控制单元,还用于:
根据所述预测数据以及所述SVM模型的历史误差,确定所述节点中的预留数据。
一方面,本发明实施例还提供一种电子设备,包括:
至少一个处理器;以及,
与所述至少一个处理器通信连接的存储器;其中,
所述存储器存储有可被所述至少一个处理器执行的指令,所述指令被所述至少一个处理器执行,以使所述至少一个处理器能够执行如上所述的节点数据预测方法。
一方面,本发明实施例还提供一种非暂态计算机可读存储介质,所述非暂态计算机可读存储介质存储计算机指令,所述计算机指令用于使所述计算机执行如上所述的节点数据预测方法。
一方面,本发明实施例还提供了一种计算机程序产品,所述计算机程序产品包括存储在非暂态计算机可读存储介质上的计算机程序,所述计算机程序包括程序指令,当所述程序指令被计算机执行时,使所述计算机执行上述节点数据预测方法。
本发明实施例中,获取参考时间段内节点的初始数据,针对参考时间段内的任一参考单位时间段,根据参考单位时间段内的初始数据,确定参考时间段的参考数据,这样,参考时间段内的所有参考数据均与时间相关,组成时间序列数据。另一方面,又获取SVM模型的模型参数。将模型参数和参考时间段内的所有参考数据输入SVM模型中,计算出预测时间段的预测数据。再依据预测数据,确定节点中的预留数据。本发明实施例利用SVM模型对预测时间段进行预测,可以进行实时预留数据预测,且确定的参考数据为时间序列数据,受影响预留数据的所有相关因素合力影响,体现了所有因素的信息,因此将参考数据作为SVM模型的输入,准确率高。同时,只使用参考数据,计算量小,数据源要求简单,减轻了计算压力。
附图说明
为了更清楚地说明本发明实施例中的技术方案,下面将对实施例描述中所需要使用的附图作简要介绍,显而易见地,下面描述中的附图仅仅是本发明的一些实施例,对于本领域的普通技术人员来讲,在不付出创造性劳动性的前提下,还可以根据这些附图获得其他的附图。
图1为本发明实施例提供的一种可能的系统构架的结构示意图;
图2为本发明实施例提供的一种节点数据预测方法的流程示意图;
图3为本发明具体实施例提供的节点数据预测方法的流程示意图;
图4为本发明实施例提供的一种节点数据预测装置的结构示意图;
图5为本发明实施例提供的电子设备的结构示意图。
具体实施方式
为了使本发明的目的、技术方案和优点更加清楚,下面将结合附图对本发明作进一步地详细描述,显然,所描述的实施例仅仅是本发明一部份实施例,而不是全部的实施例。基于本发明中的实施例,本领域普通技术人员在没有做出创造性劳动前提下所获得的所有 其它实施例,都属于本发明保护的范围。
如图1所示,本发明实施例所适用的一种系统架构,包括头寸节点101、头寸管理设备102以及头寸预测设备103。头寸节点101、头寸管理设备102以及头寸预测设备103可以是计算机等网络设备,可以是一个独立的设备,也可以是多个服务器所形成的服务器集群。优选地,头寸节点101、头寸管理设备102以及头寸预测设备103可以采用云计算技术进行信息处理。
其中,头寸节点101中有业务资金的流入流出,头寸管理设备102从各个头寸节点101中收集数据,对各个头寸节点101进行头寸监测和管理。头寸预测设备103为头寸管理设备102的辅助设备,用于对各个头寸节点101进行头寸预测,从而给头寸管理设备102以参考,使得头寸管理设备102可以根据头寸预测设备103的实时预测,调整头寸节点101中的头寸数据。
为了便于理解,下面对本发明实施例中可能涉及的名词进行定义和解释。
SVM:支持向量机(Support Vector Machine),一种基于机器学习的分类回归算法,是一类按监督学习方式对数据进行二元分类的广义线性分类器,其决策边界是对学习样本求解的最大边距超平面。一般应用于人像识别、文本分类等模式识别场景。
Spark:一种开源的分布式平台下的计算框架。
Spark Mlib:一套基于Spark计算框架的机器学习开发包。
头寸预测:对各个头寸节点流入流出资金进行预测,给业务提供一个明确的量化数据参考,避免出现资金不足或头寸安排过多的情况。本发明实施例中节点的预留数据以头寸数据为例进行介绍。
时间序列:是指将同一统计指标的数值按其发生的时间先后顺序排列而成的数列。时间序列分析的主要目的是根据已有的历史数据对未来进行预测。经济数据中大多数以时间序列的形式给出。根据观察时间的不同,时间序列中的时间可以是年份、季度、月份或其他任何时间形式。
为了实施预测节点的头寸数据,并提高预测的准确性,本发明实施例提供了一种节点数据预测方法,如图2所示,本发明实施例提供的节点数据预测方法包括以下步骤:
步骤201、获取参考时间段内节点的初始数据。
其中,节点中的初始数据,可以为该节点中参考时间段内每个时间点的实时流入数据或者流出数据,可以利用正负号标记流入流出数据,如某一时间点的初始数据为正,则表明为流入数据;为负,则表明为流出数据。
这里的参考时间段为预测时间段之前的时间段,即根据参考时间段内的数据对预测时 间段内的数据进行预测。参考时间段与预测时间段之间的距离越近,参考数据与预测数据之间的关联性越高,则预测的准确性越高。因此,太久远的历史数据对预测的数据影响不大,本发明实施例中的参考时间段可以取最近1000小时内的数据。
举例来说,表1示出了头寸节点1的实时资金流入数据。
表1
头寸节点 时间 流入金额
头寸节点1 20190101 11:12:13 253 1234.12
头寸节点1 20190101 11:12:13 123 1223.34
头寸节点1 20190101 11:11:22 321 8976.67
头寸节点1 20190101 11:11:11 321 45645.00
…… …… ……
如表1所示,头寸节点1中在2019年1月1日的11时12分13秒253这个时间点,流入资金为1234.12元;在2019年1月1日11时12分13秒123这个时间点,流入资金为1223.34元;在2019年1月1日11时11分22秒321这个时间点,流入资金为8976.67元;在2019年1月1日11时11分11秒321这个时间点,流入资金为45645.00元。则头寸节点1的初始数据为1234.12、1223.34、8976.67、45645.00。
对于流出数据,可以另外进行统计,其与流入数据的计算方式类似,这里不多做赘述。
步骤202、针对所述参考时间段内的任一参考单位时间段,根据所述参考单位时间段内的初始数据,确定所述参考单位时间段的参考数据。
其中,初始数据为每个时间点的资金流动数据,参考数据为每个参考单位时间段内该节点的资金流动数据,参考单位时间段可以取为一分钟、一刻钟、一天等,一般根据需要,参考单位时间段取一个小时,即根据每个小时内节点中各个时间点的流入或者流出数据,确定该节点每个小时的资金流动情况。
步骤203、获取SVM模型的模型参数。
具体实施过程中,SVM模型可以从Spark Mlib包里调用,或者也可以利用开发语言python实现。SVM模型的模型参数可以为预先计算得出,需要进行预测时直接获取即可;或者在预测的过程中,模型参数也实时计算修正,以提高模型计算的准确性。
步骤204、将所述模型参数以及所述参考时间段内的所有参考数据,输入所述SVM模型,计算出预测时间段的预测数据。
本发明实施例中,获取参考时间段内节点的初始数据,针对参考时间段内的任一参考 单位时间段,根据参考单位时间段内的初始数据,确定参考时间段的参考数据,这样,参考时间段内的所有参考数据均与时间相关,组成时间序列数据。另一方面,又获取SVM模型的模型参数。将模型参数和参考时间段内的所有参考数据输入SVM模型中,计算出预测时间段的预测数据。再依据预测数据,确定节点中的头寸数据。本发明实施例利用SVM模型对预测时间段进行预测,可以进行实时头寸数据预测,且确定的参考数据为时间序列数据,受影响头寸数据的所有相关因素合力影响,体现了所有因素的信息,因此将参考数据作为SVM模型的输入,准确率高。同时,只使用参考数据,计算量小,数据源要求简单,减轻了计算压力。
进一步地,为了增加模型计算的准确性,可以对参考时间段内的数据对数化处理。所述参考时间段内包括N个参考单位时间段;
针对第i个参考单位时间段,所述根据所述参考单位时间段内的初始数据,确定所述参考单位时间段的参考数据,包括:
将所述参考单位时间段内的初始数据相加,确定所述参考单位时间段的加权数据;
将第i个参考单位时间段的加权数据与所述第i-1个参考时间段的加权数据相比,并将比值取对数,得到所述第i个参考单位时间段的参考数据;其中1<i≤N。
具体实施过程中,首先将参考单位时间段内每个时间点的初始数据相加,其中正负初始数据相互抵消。例如,表2示出了每个单位时间段内头寸节点1的资金流动数据。
表2
头寸节点 时间段 流入金额
头寸节点1 20190101 11点-12点 12323432.32
头寸节点1 20190101 10点-11点 34534534.67
头寸节点1 20190101 9点-10点 12312342.45
头寸节点1 20190101 8点-9点 32132423.00
…… …… ……
如表2所示,头寸节点1中在2019年1月1日的11时至12时,流入资金为12323432.32元;2019年1月1日的10时至11时,流入资金为34534534.67元;2019年1月1日的9时至10时,流入资金为12312342.45元;2019年1月1日的8时至9时,流入资金为32132423.00元。即头寸节点1的加权数据为12323432.32、34534534.67、12312342.45、32132423.00……。
之后,用后一个参考单位时间段的加权数据与前一个参考单位时间内段的加权数据相 对比,并将比值取对数。表3示出了参考时间段内对数化处理后得到的参考数据。
表3
头寸节点 时间段 对数
头寸节点1 20190101 11点-12点 ln(12323432.32/34534534.67)=-1.03
头寸节点1 20190101 10点-11点 ln(-34534534.67/12312342.45)=1.03
头寸节点1 20190101 9点-10点 ln(12312342.45/32132423.00)=-0.96
头寸节点1 20190101 8点-9点 ln(32132423.00/…)=xxxxx
…… …… ……
如表3所示,对数化处理后得到的参考数据,标示出每一单位时间段相对于上一单位时间段的变化率。将这种变化率输入SVM算法模型中,提高了预测的准确性。
进一步地,利用SVM模型进行计算之前,需训练出模型参数。所述获取SVM模型的模型参数之前,还包括:
获取历史时间段内节点的初始数据;
针对所述历史时间段内的任一训练单位时间段,根据所述训练单位时间段内的初始数据,确定所述训练单位时间段的训练数据;
将所述历史时间段内的训练数据作为参数,输入所述SVM模型中,确定所述SVM模型的模型参数。
具体实施过程中,根据历史时间段内的很多初始数据,将通过上文介绍的根据初始数据计算出的对数化处理后的变化率作为训练数据,训练数据即为训练单位时间段内该节点的对数化处理后的参数数据。将各个单位时间段内的训练数据作为SVM模型的输入参数,将各个单位时间段各自对应的时间序列数据作为SVM模型的输出参数,其中一个单位时间段对应的所述时间序列数据为该单位时间段之前的预设数量的单位时间段对应的训练数据按照时间由先到后的顺序排列生成的数据;显然在获取到大量的上述输入参数和输出参数的对应关系后,基于现有的SVM核函数的训练方法可以得到SVM模型的模型参数。
需要说明的是,训练过程对应的历史时间段与预测过程对应的参考时间段,可以为同一时间段,也可以为不同时间段,若历史时间段与参考时间段为不同时间段时,两个时间段可以有重叠也可以没有重叠。例如,历史时间段为当前时间点之前的1000个小时,参考时间段为当前时间点之前的999个小时;或者历史时间段为2018年1月至3月的每天上午9点至11点,参考时间段为2019年1月至3月的每天上午9点至11点。历史时间段与参考时间段的选取依据计算需要,本发明实施例中不做限制。
此外,训练数据的计算方式与上述参考数据的计算方式相似,这里不多做赘述。
本发明实施例中,SVM模型的模型参数可以为动态数据,或者为相对静态的数据。即,模型参数为动态数据时,每次进行头寸数据预测时,均实时计算模型参数,这样模型参数的准确性较高,但需要实时计算得出,服务器计算压力较大。模型参数为相对静态的数据时,可以预先计算出模型参数并存储,需要进行头寸数据预测时,直接从内存中获取,此时模型参数可以每隔一段较长时间,如一个月或者半年,进行一次更新。这种方式下,服务器无需实时计算模型参数,但其准确性较低。
值得注意地,显然在计算预设时间段的预测数据时,可以将预设时间段对应的时间序列数据(如上文介绍,时间序列数据为单位时间段之前的预设数量的单位时间段对应的训练数据按照时间由先到后的顺序排列生成的数据)作为输入参数输入通过上文训练出的SVM模型,通过上文训练好的SVM模型即可计算出在预设时间段后一时刻的对数化处理后的变化率,作为预测数据。
步骤205、根据所述预测数据,确定所述节点中的预留数据。
可以理解地,由于预测数据为上文提到地变化率,所以根据上文表3的计算方法的反向计算,在已知预设时间段的流入资金以及预测数据后就可以计算出预留数据了。现实情况中,银行节点会根据前几天节点中资金的流动情况以及经验,预计出当天节点中流出资金与流入资金的差值,从而在当天开始营业之前账户中预留一定的资金,该资金即为头寸。因此,本发明实施例中,银行节点可以依据预测数据确定预留数据,即头寸数据。
可选地,由于算法仍具有一定的误差,为了进一步提高准确性,本发明实施例将预测数据与历史时间中的真实值相对比,确定误差数据。所述将所述历史时间段内的训练数据作为参数,输入所述SVM模型中,确定所述SVM模型的模型参数之后,所述根据所述预测数据,确定所述节点中的预留数据之前,还包括:
针对所述历史时间段内的任一训练单位时间段,将所述训练单位时间段之前的训练单位时间段的训练数据,以及所述模型参数输入所述SVM模型,确定所述训练单位时间段的预测数据;将所述训练单位时间段的预测数据与所述训练单位时间段的训练数据相比较,确定所述训练单位时间段的误差数据;
根据所述历史时间段内所有训练单位时间段的误差数据,确定所述SVM模型的历史误差;
所述根据所述预测数据,确定所述节点中的预留数据,包括:
根据所述预测数据以及所述SVM模型的历史误差,确定所述节点中的预留数据。
具体实施过程中,对历史时间段内的每个训练单位时间段中的头寸数据进行预测,得 到历史时间段内的多个预测数据。具体计算每个训练单位时间段中的预测数据的方法与计算参考数据的方法相似。举例来说,历史时间段包括N个训练单位时间段,分别记为T1、T2、……TN。针对T1,将T1之前的训练数据,以及模型参数输入SVM模型,得到T1的预测数据,记为X1;将X1与T1的真实数据,即T1的训练数据,相减,得到T1的误差数据。针对T2至TN,均根据上述方式得到T2的误差数据至TN的误差数据。进一步,可以根据T1至TN的误差数据,算出误差数据的均值μ与方差σ,作为该SVM模型的历史误差。
最后,根据预测数据和历史误差,确定头寸数据。现实场景中,计算出的预测数据是一个预测值,将其提供给银行节点后,若银行节点直接将实际的头寸等于预测数据,则没有考虑到误差,且无法应付突发情况。因此,往往银行节点会在预测数据的基础上增加一些数据量,即实际的头寸数据会大于预测数据。因此,本发明实施例中向银行节点提供增加数据的参考值。可选的,可以根据误差数据的均值和/或方差得到误差范围,误差范围的个数可以根据需要选择,如选择三个误差范围,分别为μ、μ+σ和μ+2σ,在不同的需求条件下可以从这三个误差范围中选择一个加上预测数据,作为最终确定的节点的头寸数据。本发明实施例中的误差范围除了上述形式,还可以为μ和2μ,或者如σ、2σ、3σ和4σ,这里只为举例,不做限制。
为了更清楚地理解本发明,下面基于图1的架构,以具体实施例对上述流程进行详细描述,具体实施例的步骤如图3所示,包括:
步骤S301:头寸预测设备从头寸管理设备中,获取头寸节点1当前时间点之前999个小时内的初始数据。
步骤S302:针对任一个小时,头寸预测设备将该小时内的初始数据相加,得到该小时的加权数据,并与该小时的前一小时的加权数据相对比,并将比值取对数,得到该小时的参考数据。
步骤S303:头寸预测设备从头寸管理设备中,获取头寸节点1当前时间点之前1000个小时内的初始数据。
步骤S304:针对任一个小时,头寸预测设备将该小时内的初始数据相加,得到该小时的加权数据,并与该小时的前一小时的加权数据相对比,并将比值取对数,得到该小时的训练数据。
步骤S305:头寸预测设备将当前时间点之前1000个小时中,每一小时的训练数据作为参数,输入SVM算法模型中进行训练,得到SVM算法模型的模型参数。其中,核函数采用高斯径向基函数做为SVM核函数。
步骤S306:头寸预测设备将999个参考数据,输入训练好的SVM算法模型中进行计算,得到变化率数据,记为V。
步骤S307:头寸预测设备确定当前时间点之前一个小时内的参考数据,记为A,根据变化率数据V和参考数据A,计算出预测数据Y。其中预测数据Y根据以下公式计算得到:
Y=A·exp(V)
步骤S308:头寸预测设备获取误差范围。
步骤S309:头寸预测设备将误差范围与预测数据Y向头寸管理设备发送。
步骤S310:头寸管理设备根据误差范围与预测数据确定头寸节点1的头寸数据。
本发明实施例还提供了一种节点数据预测装置,如图4所示,包括:
获取单元401,用于获取参考时间段内节点的初始数据;
计算单元402,用于针对所述参考时间段内的任一参考单位时间段,根据所述参考单位时间段内的初始数据,确定所述参考单位时间段的参考数据;
所述获取单元401,还用于获取支持向量机SVM模型的模型参数;
预测单元403,用于将所述模型参数以及所述参考时间段内的所有参考数据,输入所述SVM模型,计算出预测时间段的预测数据;
控制单元404,用于根据所述预测数据,确定所述节点中的预留数据。
可选的,所述参考时间段内包括N个参考单位时间段;
针对第i个参考单位时间段,所述计算单元402,具体用于:
将所述参考单位时间段内的初始数据相加,确定所述参考单位时间段的加权数据;
将第i个参考单位时间段的加权数据与所述第i-1个参考时间段的加权数据相比,并将比值取对数,得到所述第i个参考单位时间段的参考数据;其中1<i≤N。
可选的,还包括训练单元405,用于:
获取历史时间段内节点的初始数据;
针对所述历史时间段内的任一训练单位时间段,根据所述训练单位时间段内的初始数据,确定所述训练单位时间段的训练数据;
将所述历史时间段内的训练数据作为参数,输入所述SVM模型中,确定所述SVM模型的模型参数。
可选的,还包括误差单元406,用于:
针对所述历史时间段内的任一训练单位时间段,将所述训练单位时间段之前的训练数据,以及所述模型参数输入所述SVM模型,确定所述训练单位时间段的预测数据;将所述训练单位时间段的预测数据与所述训练单位时间段的训练数据相比较,确定所述训练单 位时间段的误差数据;
根据所述历史时间段内所有训练单位时间段的误差数据,确定所述SVM模型的历史误差;
所述控制单元404,还用于:
根据所述预测数据以及所述SVM模型的历史误差,确定所述节点中的预留数据。
基于相同的技术构思,本发明还提供一种电子设备,如图5所示,包括:
包括处理器501、存储器502、收发机503、总线接口504,其中处理器501、存储器502与收发机503之间通过总线接口504连接;
所述处理器501,用于读取所述存储器502中的程序,执行下列方法:
获取参考时间段内节点的初始数据;
针对所述参考时间段内的任一参考单位时间段,根据所述参考单位时间段内的初始数据,确定所述参考单位时间段的参考数据;
获取支持向量机SVM模型的模型参数;
将所述模型参数以及所述参考时间段内的所有参考数据,输入所述SVM模型,计算出预测时间段的预测数据;
根据所述预测数据,确定所述节点中的预留数据。
可选的,所述参考时间段内包括N个参考单位时间段;
针对第i个参考单位时间段,所述处理器501,具体用于:
将所述参考单位时间段内的初始数据相加,确定所述参考单位时间段的加权数据;
将第i个参考单位时间段的加权数据与所述第i-1个参考时间段的加权数据相比,并将比值取对数,得到所述第i个参考单位时间段的参考数据;其中1<i≤N。
可选的,所述处理器501,具体用于:
获取历史时间段内节点的初始数据;
针对所述历史时间段内的任一训练单位时间段,根据所述训练单位时间段内的初始数据,确定所述训练单位时间段的训练数据;
将所述历史时间段内的训练数据作为参数,输入所述SVM模型中,确定所述SVM模型的模型参数。
可选的,所述处理器501,具体用于:
针对所述历史时间段内的任一训练单位时间段,将所述训练单位时间段之前的训练数据,以及所述模型参数输入所述SVM模型,确定所述训练单位时间段的预测数据;将所述训练单位时间段的预测数据与所述训练单位时间段的训练数据相比较,确定所述训练单 位时间段的误差数据;
根据所述历史时间段内所有训练单位时间段的误差数据,确定所述SVM模型的历史误差;
根据所述预测数据以及所述SVM模型的历史误差,确定所述节点中的预留数据。
基于相同的技术构思,本发明实施例还提供了一种非暂态计算机可读存储介质,所述非暂态计算机可读存储介质存储计算机指令,所述计算机指令用于使所述计算机执行如上所述的节点数据预测方法。
基于相同的技术构思,本发明实施例还提供了一种计算机程序产品,所述计算机程序产品包括存储在非暂态计算机可读存储介质上的计算机程序,所述计算机程序包括程序指令,当所述程序指令被计算机执行时,使所述计算机执行上述节点数据预测方法。
本发明是参照根据本发明实施例的方法、设备(系统)、和计算机程序产品的流程图和/或方框图来描述的。应理解可由计算机程序指令实现流程图和/或方框图中的每一流程和/或方框、以及流程图和/或方框图中的流程和/或方框的结合。可提供这些计算机程序指令到通用计算机、专用计算机、嵌入式处理机或其他可编程数据处理设备的处理器以产生一个机器,使得通过计算机或其他可编程数据处理设备的处理器执行的指令产生用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的装置。
这些计算机程序指令也可存储在能引导计算机或其他可编程数据处理设备以特定方式工作的计算机可读存储器中,使得存储在该计算机可读存储器中的指令产生包括指令装置的制造品,该指令装置实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能。
这些计算机程序指令也可装载到计算机或其他可编程数据处理设备上,使得在计算机或其他可编程设备上执行一系列操作步骤以产生计算机实现的处理,从而在计算机或其他可编程设备上执行的指令提供用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的步骤。
尽管已描述了本发明的优选实施例,但本领域内的技术人员一旦得知了基本创造性概念,则可对这些实施例作出另外的变更和修改。所以,所附权利要求意欲解释为包括优选实施例以及落入本发明范围的所有变更和修改。
显然,本领域的技术人员可以对本发明进行各种改动和变型而不脱离本发明的精神和范围。这样,倘若本发明的这些修改和变型属于本发明权利要求及其等同技术的范围之内,则本发明也意图包括这些改动和变型在内。

Claims (11)

  1. 一种节点数据预测方法,其特征在于,所述方法包括:
    获取参考时间段内节点的初始数据;
    针对所述参考时间段内的任一参考单位时间段,根据所述参考单位时间段内的初始数据,确定所述参考单位时间段的参考数据;
    获取支持向量机SVM模型的模型参数;
    将所述模型参数以及所述参考时间段内的所有参考数据,输入所述SVM模型,计算出预测时间段的预测数据;
    根据所述预测数据,确定所述节点中的预留数据。
  2. 如权利要求1所述的方法,其特征在于,所述参考时间段内包括N个参考单位时间段;
    针对第i个参考单位时间段,所述根据所述参考单位时间段内的初始数据,确定所述参考单位时间段的参考数据,包括:
    将所述参考单位时间段内的初始数据相加,确定所述参考单位时间段的加权数据;
    将第i个参考单位时间段的加权数据与所述第i-1个参考时间段的加权数据相比,并将比值取对数,得到所述第i个参考单位时间段的参考数据;其中1<i≤N。
  3. 如权利要求1所述的方法,其特征在于,所述获取SVM模型的模型参数之前,还包括:
    获取历史时间段内节点的初始数据;
    针对所述历史时间段内的任一训练单位时间段,根据所述训练单位时间段内的初始数据,确定所述训练单位时间段的训练数据;
    将所述历史时间段内的训练数据作为参数,输入所述SVM模型中,确定所述SVM模型的模型参数。
  4. 如权利要求3所述的方法,其特征在于,所述将所述历史时间段内的训练数据作为参数,输入所述SVM模型中,确定所述SVM模型的模型参数之后,所述根据所述预测数据,确定所述节点中的预留数据之前,还包括:
    针对所述历史时间段内的任一训练单位时间段,将所述训练单位时间段之前的训练数据,以及所述模型参数输入所述SVM模型,确定所述训练单位时间段的预测数据;将所述训练单位时间段的预测数据与所述训练单位时间段的训练数据相比较,确定所述训练单位时间段的误差数据;
    根据所述历史时间段内所有训练单位时间段的误差数据,确定所述SVM模型的历史误差;
    所述根据所述预测数据,确定所述节点中的预留数据,包括:
    根据所述预测数据以及所述SVM模型的历史误差,确定所述节点中的预留数据。
  5. 一种节点数据预测装置,其特征在于,包括:
    获取单元,用于获取参考时间段内节点的初始数据;
    计算单元,用于针对所述参考时间段内的任一参考单位时间段,根据所述参考单位时间段内的初始数据,确定所述参考单位时间段的参考数据;
    所述获取单元,还用于获取支持向量机SVM模型的模型参数;
    预测单元,用于将所述模型参数以及所述参考时间段内的所有参考数据,输入所述SVM模型,计算出预测时间段的预测数据;
    控制单元,用于根据所述预测数据,确定所述节点中的预留数据。
  6. 如权利要求5所述的装置,其特征在于,所述参考时间段内包括N个参考单位时间段;
    针对第i个参考单位时间段,所述计算单元,具体用于:
    将所述参考单位时间段内的初始数据相加,确定所述参考单位时间段的加权数据;
    将第i个参考单位时间段的加权数据与所述第i-1个参考时间段的加权数据相比,并将比值取对数,得到所述第i个参考单位时间段的参考数据;其中1<i≤N。
  7. 如权利要求5所述的装置,其特征在于,还包括训练单元,用于:
    获取历史时间段内节点的初始数据;
    针对所述历史时间段内的任一训练单位时间段,根据所述训练单位时间段内的初始数据,确定所述训练单位时间段的训练数据;
    将所述历史时间段内的训练数据作为参数,输入所述SVM模型中,确定所述SVM模型的模型参数。
  8. 如权利要求7所述的装置,其特征在于,还包括误差单元,用于:
    针对所述历史时间段内的任一训练单位时间段,将所述训练单位时间段之前的训练数据,以及所述模型参数输入所述SVM模型,确定所述训练单位时间段的预测数据;将所述训练单位时间段的预测数据与所述训练单位时间段的训练数据相比较,确定所述训练单位时间段的误差数据;
    根据所述历史时间段内所有训练单位时间段的误差数据,确定所述SVM模型的历史误差;
    所述控制单元,还用于:
    根据所述预测数据以及所述SVM模型的历史误差,确定所述节点中的预留数据。
  9. 一种电子设备,其特征在于,包括:
    至少一个处理器;以及,
    与所述至少一个处理器通信连接的存储器;其中,
    所述存储器存储有可被所述至少一个处理器执行的指令,所述指令被所述至少一个处理器执行,以使所述至少一个处理器能够执行权利要求1-4任一所述的方法。
  10. 一种非暂态计算机可读存储介质,其特征在于,所述非暂态计算机可读存储介质存储计算机指令,所述计算机指令用于使所述计算机执行权利要求1~4任一所述方法。
  11. 一种计算机程序产品,其特征在于,所述计算机程序产品包括存储在非暂态计算机可读存储介质上的计算机程序,所述计算机程序包括程序指令,当所述程序指令被计算机执行时,使所述计算机执行权利要求1~4任一所述方法。
PCT/CN2020/084322 2019-06-24 2020-04-10 一种节点数据预测方法及装置 Ceased WO2020258994A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910549462.1 2019-06-24
CN201910549462.1A CN110276492A (zh) 2019-06-24 2019-06-24 一种节点数据预测方法及装置

Publications (1)

Publication Number Publication Date
WO2020258994A1 true WO2020258994A1 (zh) 2020-12-30

Family

ID=67961679

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/084322 Ceased WO2020258994A1 (zh) 2019-06-24 2020-04-10 一种节点数据预测方法及装置

Country Status (2)

Country Link
CN (1) CN110276492A (zh)
WO (1) WO2020258994A1 (zh)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114154696A (zh) * 2021-11-19 2022-03-08 中国建设银行股份有限公司 资金流预测方法、系统、计算机设备及存储介质
CN114202231A (zh) * 2021-12-21 2022-03-18 瀚云科技有限公司 日用电数据缺失的补全方法、装置、电子设备及存储介质
CN116740946A (zh) * 2023-08-14 2023-09-12 山东华夏高科信息股份有限公司 一种公路车流量大数据分析处理方法

Families Citing this family (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110276492A (zh) * 2019-06-24 2019-09-24 深圳前海微众银行股份有限公司 一种节点数据预测方法及装置
CN111143169B (zh) * 2019-12-30 2024-02-27 杭州迪普科技股份有限公司 异常参数检测方法及装置、电子设备、存储介质
CN111752816A (zh) * 2020-06-30 2020-10-09 深圳前海微众银行股份有限公司 一种操作系统的分析方法及装置
US20220230236A1 (en) * 2021-01-20 2022-07-21 Bank Of America Corporation Artificial intelligence (ai) architecture with smart, automated triggers of incoming and outgoing actions and usage
CN113159790B (zh) * 2021-05-19 2025-05-06 中国银行股份有限公司 一种异常交易识别方法和装置

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20060026072A1 (en) * 2002-09-06 2006-02-02 Takao Suguro Safe stock amount calculation method, safe stock amount calculation device, order making moment calculation method , order making moment calculation device , and order making amount calculation method
CN103258388A (zh) * 2012-02-17 2013-08-21 日立(中国)研究开发有限公司 自动交易装置和服务器、以及用于预测现金需求量的方法
CN104424598A (zh) * 2013-09-06 2015-03-18 株式会社日立制作所 现金需求量预测装置与方法
CN106997497A (zh) * 2017-06-13 2017-08-01 中南大学 一种基于时间序列和节假日信息的银行备付金预测方法
CN110276492A (zh) * 2019-06-24 2019-09-24 深圳前海微众银行股份有限公司 一种节点数据预测方法及装置

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106952420A (zh) * 2016-01-07 2017-07-14 株式会社日立制作所 自动柜员机现金管理装置、系统及方法
CN106886846A (zh) * 2017-04-26 2017-06-23 中南大学 一种基于长短期记忆循环神经网络的银行网点备付金预测方法
CN109697522B (zh) * 2017-10-20 2021-07-09 北京京东尚科信息技术有限公司 一种数据预测的方法和装置

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20060026072A1 (en) * 2002-09-06 2006-02-02 Takao Suguro Safe stock amount calculation method, safe stock amount calculation device, order making moment calculation method , order making moment calculation device , and order making amount calculation method
CN103258388A (zh) * 2012-02-17 2013-08-21 日立(中国)研究开发有限公司 自动交易装置和服务器、以及用于预测现金需求量的方法
CN104424598A (zh) * 2013-09-06 2015-03-18 株式会社日立制作所 现金需求量预测装置与方法
CN106997497A (zh) * 2017-06-13 2017-08-01 中南大学 一种基于时间序列和节假日信息的银行备付金预测方法
CN110276492A (zh) * 2019-06-24 2019-09-24 深圳前海微众银行股份有限公司 一种节点数据预测方法及装置

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114154696A (zh) * 2021-11-19 2022-03-08 中国建设银行股份有限公司 资金流预测方法、系统、计算机设备及存储介质
CN114202231A (zh) * 2021-12-21 2022-03-18 瀚云科技有限公司 日用电数据缺失的补全方法、装置、电子设备及存储介质
CN116740946A (zh) * 2023-08-14 2023-09-12 山东华夏高科信息股份有限公司 一种公路车流量大数据分析处理方法
CN116740946B (zh) * 2023-08-14 2023-11-03 山东华夏高科信息股份有限公司 一种公路车流量大数据分析处理方法

Also Published As

Publication number Publication date
CN110276492A (zh) 2019-09-24

Similar Documents

Publication Publication Date Title
WO2020258994A1 (zh) 一种节点数据预测方法及装置
Khambam et al. The Cloud as A Financial Forecast: Leveraging AI For Predictive Analytics
AU2016328959A1 (en) Updating attribute data structures to indicate trends in attribute data provided to automated modeling systems
AU2016243106A1 (en) Optimizing neural networks for risk assessment
US20110153536A1 (en) Computer-Implemented Systems And Methods For Dynamic Model Switching Simulation Of Risk Factors
US11042677B1 (en) Systems and methods for time series simulation
WO2016084642A1 (ja) 与信審査用サーバと与信審査用システム及び与信審査用プログラム
CN111738504A (zh) 企业财务指标资金量预测方法和装置、设备及存储介质
JP6771513B2 (ja) 債務不履行確率を算出する装置、方法及びそのためのプログラム
CN112734516A (zh) 基于客户现金流预测的金融产品推荐方法及系统
CN117333025A (zh) 基于多维数据的企业风险评估方法、装置、设备和介质
Katranzhy et al. Modelling the dynamics of the adequacy of bank’s regulatory capital
JP7298286B2 (ja) モデル提供プログラム、モデル提供方法及びモデル提供装置
Tikhonov et al. The Relationship Between the Financial Performance of Banks and the Quality of Credit Scoring Models
CN112308293B (zh) 违约概率预测方法及装置
US8712890B2 (en) Systems and methods for optimizing capital structure of a financial institution
KR102767248B1 (ko) 자금세탁 의심거래 위험도 평가 방법 및 그 장치
Konovalova et al. Evaluation of credit policy implementation in commercial banks: Evidence from latvia
Faricha et al. Comparison study of transfer function and artificial neural network for cash flow analysis at Bank Rakyat Indonesia
KR20250033880A (ko) 조건부 확률 모델을 통한 추론 방법 및 이를 수행하는 컴퓨팅 장치
CN115659599A (zh) 一种数据处理方法及装置
Kubenka et al. Implementation of standards into predictors of financial stability
US20200302550A1 (en) Cost allocation estimation using direct cost vectors and machine learning
Sadoghi Measuring systemic risk: Robust ranking techniques approach
CN112308295B (zh) 违约概率预测方法及装置

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20831488

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20831488

Country of ref document: EP

Kind code of ref document: A1