CN106886846A - A kind of bank outlets' excess reserve Forecasting Methodology that Recognition with Recurrent Neural Network is remembered based on shot and long term - Google Patents
A kind of bank outlets' excess reserve Forecasting Methodology that Recognition with Recurrent Neural Network is remembered based on shot and long term Download PDFInfo
- Publication number
- CN106886846A CN106886846A CN201710281286.9A CN201710281286A CN106886846A CN 106886846 A CN106886846 A CN 106886846A CN 201710281286 A CN201710281286 A CN 201710281286A CN 106886846 A CN106886846 A CN 106886846A
- Authority
- CN
- China
- Prior art keywords
- day
- bank outlets
- excess reserve
- shot
- neural network
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
- G06N3/061—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using biological neurons, e.g. biological neurons connected to an integrated circuit
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION 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
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION 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
- G06Q40/00—Finance; Insurance; Tax strategies; Processing of corporate or income taxes
- G06Q40/02—Banking, e.g. interest calculation or account maintenance
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Business, Economics & Management (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Strategic Management (AREA)
- Molecular Biology (AREA)
- Economics (AREA)
- Software Systems (AREA)
- Marketing (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Mathematical Physics (AREA)
- Data Mining & Analysis (AREA)
- Computational Linguistics (AREA)
- Artificial Intelligence (AREA)
- Accounting & Taxation (AREA)
- Finance (AREA)
- Development Economics (AREA)
- Neurology (AREA)
- General Health & Medical Sciences (AREA)
- Human Resources & Organizations (AREA)
- General Business, Economics & Management (AREA)
- Technology Law (AREA)
- Game Theory and Decision Science (AREA)
- Microelectronics & Electronic Packaging (AREA)
- Entrepreneurship & Innovation (AREA)
- Operations Research (AREA)
- Quality & Reliability (AREA)
- Tourism & Hospitality (AREA)
- Financial Or Insurance-Related Operations Such As Payment And Settlement (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a kind of bank outlets' excess reserve Forecasting Methodology that Recognition with Recurrent Neural Network is remembered based on shot and long term, including data prediction, model training and prediction three phases.Data preprocessing phase counts bank outlets' cash transaction day total deposit, day withdrawal total value, and same day date property, construction feature vector in units of day；According to the day net amount that daily cash transaction record is calculated.In the model training stage, according to history feature vector sum day net amount data training LSTM models.The characteristic vector of some days, interval where the day net amount on the day of input LSTM model predictions before forecast period, statistics bank outlets forecast date, takes excess reserve demand of the random value in the interval as the same day.The present invention takes full advantage of the advantage of historical data and shot and long term memory Recognition with Recurrent Neural Network on data time series analysis, effectively increases the accuracy rate of bank outlets' excess reserve prediction.
Description
Technical field
The present invention relates to a kind of bank outlets' excess reserve Forecasting Methodology that Recognition with Recurrent Neural Network is remembered based on shot and long term.
Background technology
In recent years, with the rise and development of Third-party payment mechanism, business bank's monetary transaction receives certain punching
Hit, Intermediary Business in Commercial Bank stands in the breach.Intermediary Business in Commercial Bank mainly include payment and settlement, guarantee, promise, transaction,
Inquiry etc., is most important part as the payment and settlement business of traditional media.Business bank's Corporate Performance system principal measure
Mobility, security, profitability three indexs, wherein, excess reserve ratio is the key factor for influenceing three norms.
For bank outlets, excess reserve represents the monetary transaction liveness of its radiation scope, and this liveness receives the date
The influence of attribute.Generally, bank outlets conclude the business more frequently on weekdays, and especially service for corporate customers is in great demand, and false in section
During day, bank outlets generally only provide self-service withdrawal service, and block trade is less.Therefore, prediction bank outlets excess reserve
, it is necessary to consider date property first during demand.
On the one hand, cash transaction situation of bank outlets' excess reserve prediction based on bank outlets in a period of time in the past, main
Excess reserve demand is predicted according to the monetary transaction of bank outlets day net amount.Within a period of time, bank outlets' business
Day net amount is one group of time series, therefore, bank outlets' excess reserve prediction can be converted into time series forecasting problem.Both at home and abroad
Researcher has carried out correlative study on time series direction.The method of conventional construction time series models, such as autoregression
Model, linear dynamic model, it is linear correlation that this class method is typically simulated time sequence, is not particularly suited for bank's provision
The prediction of this nonlinear organization data of gold.
On the other hand, developing rapidly with deep learning, nerual network technique is also being constantly updated.Recurrent neural network
(RNNs, Recurrent Neural Networks) is due to self-learning property, and its expansion model is exactly Time Dependent mould
Type, therefore it is particularly suitable for time series forecasting.Word is just predicted by using RNNs in natural language processing (NLP).Reason
For upper, RNNs can remember permanent information, carry out predicted time sequence, but in fact, RNNs to long-range Dependency Specification at
Reason scarce capacity.
It can be seen that, either in practical application or scientific research, time series has important researching value.However,
Existing achievement in research is not met by the new demand that application scenarios are proposed to time series forecasting for becoming increasingly complex, current
Excess reserve prediction uses statistical method mostly, it is difficult to efficiently, accurately tackle excess reserve forecast demand.Therefore, design a kind of
Can efficiently and precisely predict the method for bank outlets' excess reserve demand has great Research Significance.
The content of the invention
Technical problem solved by the invention is, in view of the shortcomings of the prior art, proposing a kind of based on shot and long term memory circulation
Bank outlets' excess reserve forecasting system of neutral net (Long Short Term Memory, LSTM), bank outlets day are net
Volume as excess reserve demand normative reference, and with reference to forecast date date property for bank outlets set up LSTM prediction mould
Type, effectively excavates the excess reserve demand of bank outlets' scheduled date, energy from the passing cash transaction record of bank outlets
Effectively improve bank outlets' excess reserve precision of prediction.
In order to solve the above technical problems, the technical solution adopted in the present invention is：
A kind of bank outlets' excess reserve Forecasting Methodology that Recognition with Recurrent Neural Network is remembered based on shot and long term, it is characterised in that bag
Include data preprocessing phase, model training stage and forecast period；
The data preprocessing phase, collects bank outlets' record of the cash transaction of some days as training set, builds and hands over
Easy database, database includes recording the characteristic vector built in units of day according to daily cash transaction, and according to
The day net amount that daily cash transaction record is calculated；
The model training stage, the shot and long term memory Recognition with Recurrent Neural Network mould with several hidden layers is built, i.e.,
LSTM models, wherein each hidden layer include several neurons, using the characteristic vector in transaction data base and day net amount pair
LSTM models are trained, and obtain optimal L STM Model Weight parameters；
The forecast period, determines bank outlets and forecast date to be predicted, collects some days before the bank outlets
Cash transaction is recorded, and characteristic vector is then converted into respectively and is input in LSTM forecast models, and output predicts the outcome, i.e. excess reserve
Predicted value.
Further, in data preprocessing phase, in units of day, count daily total deposit, total withdrawal volume and
Date property, builds characteristic vector, and form is as follows：[month, date, total deposit, volume of always withdrawing the money, date property], wherein, the date
Attribute is divided into working day, Saturday, Sunday and festivals or holidays, is respectively labeled as 0,1,2,3, and festivals or holidays attribute is better than other dates
Attribute；
Day net amount is calculated according to cash transaction record；The day net amount distribution of bank outlets, is put in one section of cycle of statistics
Letter is interval；And it is total interval using confidential interval as prediction, it is divided into N_CLASS subinterval；According to daily day net amount
Residing subinterval is marked as an one-hot vector for N_CLASS dimensions, each dimension correspondence in one-hot vectors
One subinterval, the vector element of subinterval correspondence dimension is designated as 1 residing for the day net amount of this day, the vector element mark of other dimensions
It is 0.
Further, the N_CLASS values are 5.
Market day net amount distribution map (Fig. 2) is trained according to bank outlets as can be seen that bank outlets day net amount substantially meets just
The characteristics of state is distributed.Thus, it is supposed that data Normal Distribution to be predicted, to determine the border of its span.Next,
Just different intervals can will be divided into border, so as to obtain mark of the different samples by interval classification.Based on normal distribution
It is assumed that calculating the average (MEAN) and standard deviation (STD) of the day net amount of training set sample.Using 99% confidential interval as all samples
This day net amount distribution [MEAN-2.576*STD, MEAN+2.576*STD].This distribution is divided into N_CLASS
Individual subinterval, these subintervals are closed the right side and are opened as the criteria for classifying with a left side.After being divided into N_CLASS subinterval, each subinterval is pressed
According to 0 to N_CLASS-1 numberings, and determine it is interval where the day net amount of all samples, not N_CLASS interval ranges data then
The 0th interval is divided to nearby or the one N_CLASS-1 interval；The interval classification of the day net amount of sample is changed into one-
Hot vector forms.Therefore, this bank outlets excess reserve prediction seeks to predict that the excess reserve demand of scheduled date will fall in tool
In the middle of which subinterval of body.
, be input to the characteristic vector of not same date in LSTM forecast models successively, by input gate by the forecast period
Layer, the three kinds of matrixings forgotten gate layer, export gate layer, the one-hot vectors of output prediction day.
Further, the forecast period, the one-hot vectors according to prediction day are obtained residing for the day net amount of prediction day
Subinterval, and the random value in the subinterval is taken as excess reserve predicted value.
Further, the model training stage, iteration is trained to LSTM models several times, uses random small real number
Initialize LSTM models weight parameter (including be input to hidden layer, hidden layer to exporting, the connection weight of hidden layer to hidden layer
Weight U, V and W), Regularization is carried out using random inactivation, determine to export as activation primitive using tanh nonlinear functions, and
Cross entropy is chosen as loss function, prediction is obtained into the vectorial one-hot vectors corresponding with true day net amount of one-hot is carried out
Contrast, counting loss, and use stochastic gradient descent (Stochastic Gradient Descent, SGD) and backpropagation
(BackPropagation, BP) method is combined the optimal weights parameter found and cause loss reduction.
Further, the model training stage, hidden layer number is 5, and each hidden layer includes 50 neurons, iteration
100000 times LSTM models are trained.
Each step to the model training stage is described in detail below.
Step one：Weights initialisation；In LSTM model process, weight parameter U, V and W are needed by constantly updating acquisition most
The figure of merit.If each neuron in network calculates identical output, in back-propagation process, identical ladder will be obtained
Degree, causes parameter U, V and W consistent updates, neuron to be always maintained at symmetrically, influenceing training effect.Therefore, it is suitable in order to obtain
LSTM model learning parameters, the present invention have selected the initialization of small random number, i.e. weight parameter U, V and W initial values close to 0,
And it is not equal to 0.Initialized with minimum numerical value so that neuron original state is random and unequal, then calculates
Different renewals, and in gradient descent procedures, the different piece of whole network is converted itself into, break symmetry.It is small with
The computing formula of machine number weights initialisation is as follows：
W=0.01*np.random.randn (n) * sqrt (2.0/n)
Wherein, np.random is the method for generating random number in Python in Numpy storehouses, and randn functions are based on standard
Difference isGaussian Profile generate random number, wherein n is the neuronal quantity of input layer.According to this computing formula, each
The weight vectors of neuron can all be initialized to a random vector, and these random vectors obey a multivariate Gaussian
Distribution, therefore, in the input space, the sensing of all of neuron is all random.
Step 2：Regularization；For the learning ability of control neural network, over-fitting is prevented, in propagated forward mistake
Cheng Zhongxu carries out regularization.The random inactivation (Dropout) of present invention selection is used as regularization method.In the training process, to complete
Whole neutral net is sampled out a part, and the parameter of sub-network is updated based on input data, allows neuron to be swashed with the probability of p
It is living；Default setting p of the present invention is 0.7.
Step 3：Activation primitive, LSTM models need to use a nonlinear function f to determine the output of neuron.It is conventional
Activation primitive have tanh and ReLU.The present invention is initialized using tanh nonlinear functions, and be compressed to for real number value by tanh
Between [- 1,1], its output is zero center.Computing formula is as follows：
Tanh (x)=2 δ (2x) -1
Wherein, x representative functions input variable, δ is Sigmoid functions, and its computing formula is as follows：
Step 4：SGD and backpropagation.During model training, shot and long term memory Recognition with Recurrent Neural Network is run to last
Penalty values can be gone out according to the error calculation predicted the outcome between actual value using loss function after one layer, backpropagation (BP) is calculated
This penalty values opposite direction can be passed to each neuron by method, and then each neuron can be repaiied using stochastic gradient descent (SGD)
Positive LSTM weight parameters U, V and W.During model training, LSTM weight parameters U, V and W are constantly updated, meet it minimum
Change loss condition.
SGD is a kind of method that iterative minimizes loss.In order to find a local minimum for function, SGD meetings
Search is iterated to the regulation step distance point of the opposite direction of current point correspondence gradient on function.In SGD, each iteration is not
All samples must be traveled through, it will be using a sample come undated parameter, and parameter will update along negative gradient direction.
Back-propagation algorithm is updated using the method for batch updating to the weights of neutral net and biasing, its treatment
Step is as shown in table 1.
Loss function represents that this method uses cross entropy as loss function counting loss, to weigh model training with L
Effect, its computing formula of cross entropy is as follows：
Wherein, N is training samples number, onIt is the day net amount predicted value of n-th training sample of LSTM models output, yn
It is n-th day net amount actual value of training sample.If onWith ynBetween gap it is bigger, then lose L (y, o) it is bigger.
Beneficial effect：
Be converted into for excess reserve sequence prediction problem by the characteristics of in view of excess reserve sequence there is higher-dimension, height to make an uproar, the present invention
Time series forecasting problem.Furthermore, it is contemplated that shot and long term memory Recognition with Recurrent Neural Network model, i.e. LSTM models are in time series
There is good performance in classification problem, and rational interval being had more than one exact value of prediction of prediction one is of practical significance,
Excess reserve forecasting problem is converted into the present invention classification prediction of time series.The corresponding day net amount data of sample are carried out first
Conversion, it is sorted out according to span, class label on mark；Then using the training of real bank site cash transaction data
LSTM models, were finally input to LSTM models by the characteristic vector of some days before forecast date, and prediction bank outlets excess reserve is needed
Ask, compared to current existing ARIMA Forecasting Methodologies, the present invention take full advantage of different date properties and time series according to
Lai Xing, the prediction algorithm of proposition can more fully excavate dependence between bank outlets' excess reserve, so as to solve bank's net
The constantly complicated problem of point excess reserve prediction, improves excess reserve predictablity rate and precision, it is to avoid prediction occur too high or too low
Situation.Show in the model prediction result of 453 bank outlets, compared to more existing ARIMA Forecasting Methodologies, the present invention is carried
The method for going out, in prediction and evaluation standard MAD (mean absolute error), RMSE (root-mean-square error), MAPE (average absolute percentages
Error) contrast in, hence it is evident that better than ARIMA Forecasting Methodologies.
Brief description of the drawings
Fig. 1 is bank outlets' excess reserve Forecasting Methodology flow that recirculating network is remembered based on shot and long term；
Fig. 2 is the day net amount distribution of certain bank outlets' training set；
Fig. 3 is bank outlets' excess reserve forecast model training process that Recognition with Recurrent Neural Network is remembered based on shot and long term；
Fig. 4 is bank outlets' excess reserve forecast interval and day net amount contrast based on shot and long term memory Recognition with Recurrent Neural Network
Figure；
Fig. 5 is on June 9,1 day to 2015 January in 2015, bank outlets' excess reserve predicted value and true day net amount and
ARIMA predicted value comparison diagrams；
Fig. 6 is on April 30,1 day to 2015 April in 2015, bank outlets' excess reserve predicted value and true day net amount and
ARIMA predicted value comparison diagrams；
Fig. 7 is on June 9,1 day to 2015 January in 2015, and bank outlets' excess reserve predicts the outcome and ARIMA predicted values
MAD Comparison of standards figures；
Fig. 8 is on June 9,1 day to 2015 January in 2015, and bank outlets' excess reserve predicts the outcome and ARIMA predicted values
RMSE Comparison of standards figures；
Fig. 9 is on June 9,1 day to 2015 January in 2015, and bank outlets' excess reserve predicts the outcome and ARIMA predicted values
MAPE Comparison of standards figures.
Specific embodiment
Present system framework including data prediction, model training and bank outlets' excess reserve as shown in figure 1, predict three
Part.In data preprocessing phase, the cash transaction record of bank outlets is collected, and transaction record is counted, be somebody's turn to do
The day net amount sequence of bank outlets.The present invention will build the feature of [month, date, total deposit, total to withdraw the money, date property] to
Amount.For according to site day net amount distribution, calculate its average value and standard deviation, the confidential interval using 99% as forecast interval,
And forecast interval is divided into N_CLASS classes, it is one-hot vectors by every recording mark.
In the model training stage, the default parameters configuration for using is as shown in table 1.And at the beginning of by small random number carrying out weight
Beginningization, tanh nonlinear functions activate, will predict the outcome and one-hot vector contrasts, calculate cross entropy, parameter is reversely being passed
Quickly restrained during broadcasting.
Table 1LSTM forecast model default configuration parameters values
| Parameter name | Default configuration |
| Implicit layer number | 5 |
| LSTM model training sample number of days | 10 |
| Neuronal quantity in hidden layer | 50 |
| Activation primitive | Tanh |
| Loss function | Cross entropy |
| Learning rate | 0.001 |
| LSTM model batch processing sizes | 200 |
| Initialization weight method | Small random number |
| Iterations | 100000 |
Back-propagation algorithm is：
1. pair all of hidden layer (2≤l≤L), if Δ W(l)=0, Δ b(l)=0, i.e. Δ W(l)With Δ b(l)It is respectively complete
Null matrix and full null vector；
2.Fori=1:m
(1) back-propagation algorithm is used, the gradient matrix of each layer neuron weights and biasing is calculatedWith
(2) calculate
(3) calculate
3. weights and biasing are updated：
(1) calculate
(2) calculate
Wherein, m is the number of training sample.
In the model prediction stage, the characteristic vector of the bank cash transaction day net amount of preceding ten days is chosen, be sequentially inputted to
In LSTM models, and random value is taken as excess reserve requirement forecasting value in the forecast interval of forecast model output.
This method is tested using real bank site cash transaction data.In Forecasting Methodology, due to only considered
Date property and excess reserve as time series influence, be that bank outlets set up forecast model, its configuration parameter is adjustable
Section.LSTM forecast models default configuration parameters value is as shown in table 2 in the present embodiment.
As shown in figure 4, under conditions of certain bank outlets predicts January 1 to 9 days June in 2015 in 2015, based on LSTM
Forecast model given interval in the ratio comprising true day net amount be 85.0679%.
As shown in figure 5, by taking random value in forecast interval, as final excess reserve requirement forecasting, compared to ARIMA
Forecasting Methodology, LSTM model prediction results contrasts press close to true day net amount.
As shown in fig. 6, LSTM model predictions result more stably simulates the trend of true day net amount, ARIMA prediction knots
Fruit is more steady, it is impossible to reflect the trend of true day net amount well.
Additionally, according to existing time series evaluation criterion, this method compared for LSTM Forecasting Methodologies and ARIMA
(Autoregressive Integrated Moving Average Model, ARMA model) Forecasting Methodology exists
Value in MAD (mean absolute error), RMSE (root-mean-square error), three indexs of MAPE (mean absolute percentage error).
As shown in fig. 7, MAD value of the MAD values of LSTM Forecasting Methodologies significantly lower than ARIMA methods, in addition, with classification area
Between quantity growth, the MAD values of LSTM Forecasting Methodologies will be gradually reduced.
As shown in figure 8, RMSE value of the RMSE value of LSTM Forecasting Methodologies substantially with ARIMA Forecasting Methodologies is equal, but with
The growth of class interval quantity, the RMSE value of LSTM Forecasting Methodologies is declined slightly.
As shown in figure 9, MAD value of the MAPE values of LSTM Forecasting Methodologies significantly lower than ARIMA methods, additionally, with classification
The growth of interval quantity, the MAPE values of LSTM Forecasting Methodologies will be gradually reduced, and substantially level off to 0.
Claims (7)
1. it is a kind of based on shot and long term remember Recognition with Recurrent Neural Network bank outlets' excess reserve Forecasting Methodology, it is characterised in that including
Data preprocessing phase, model training stage and forecast period；
The data preprocessing phase, collects bank outlets' record of the cash transaction of some days as training set, builds number of deals
According to storehouse, database includes recording the characteristic vector built in units of day according to daily cash transaction, and according to daily
The cash transaction day net amount that is calculated of record；
The model training stage, build shot and long term memory Recognition with Recurrent Neural Network model, i.e. LSTM with several hidden layers
Model, wherein each hidden layer include several neurons, using the characteristic vector in transaction data base and day net amount to LSTM
Model is trained, and obtains optimal L STM Model Weight parameters；
The forecast period, determines bank outlets and forecast date to be predicted, collects the cash of some days before the bank outlets
Transaction record, is then converted into characteristic vector and is input in LSTM forecast models respectively, and output predicts the outcome, i.e. excess reserve prediction
Value.
2. it is according to claim 1 based on shot and long term remember Recognition with Recurrent Neural Network bank outlets' excess reserve Forecasting Methodology,
Characterized in that, in data preprocessing phase, in units of day, the daily total deposit of statistics, total withdrawal volume and date category
Property, characteristic vector is built, form is as follows：[month, date, total deposit, volume of always withdrawing the money, date property], wherein, date property point
It is working day, Saturday, Sunday and festivals or holidays, is respectively labeled as 0,1,2,3, and festivals or holidays attribute is better than other date properties；
Day net amount is calculated according to cash transaction record；The day net amount distribution of bank outlets, obtains confidence area in one section of cycle of statistics
Between；And it is total interval using confidential interval as prediction, it is divided into N_CLASS subinterval；According to residing for daily day net amount
Subinterval is marked as an one-hot vector for N_CLASS dimensions, each dimension correspondence one in one-hot vectors
Subinterval, the vector element of subinterval correspondence dimension is designated as 1 residing for the day net amount of this day, and the vector element of other dimensions is designated as 0.
3. it is according to claim 2 based on shot and long term remember Recognition with Recurrent Neural Network bank outlets' excess reserve Forecasting Methodology,
Characterized in that, the N_CLASS values are 5.
4. it is according to claim 3 based on shot and long term remember Recognition with Recurrent Neural Network bank outlets' excess reserve Forecasting Methodology,
Characterized in that, be input to the characteristic vector of not same date in LSTM forecast models successively by the forecast period, by input
Gate layer, the three kinds of matrixings forgotten gate layer, export gate layer, the one-hot vectors of output prediction day.
5. it is according to claim 4 based on shot and long term remember Recognition with Recurrent Neural Network bank outlets' excess reserve Forecasting Methodology,
Characterized in that, the forecast period, obtains predicting the sub-district residing for the day net amount of day according to the one-hot vectors for predicting day
Between, and the random value in the subinterval is taken as excess reserve predicted value.
6. it is according to claim 3 based on shot and long term remember Recognition with Recurrent Neural Network bank outlets' excess reserve Forecasting Methodology,
Characterized in that, the model training stage, iteration is trained to LSTM models several times, uses random small real number to initialize
The weight parameter of LSTM models, Regularization is carried out using random inactivation, true as activation primitive using tanh nonlinear functions
Fixed output, and cross entropy is chosen as loss function, prediction is obtained into the vectorial one-s corresponding with true day net amount of one-hot
Hot vectors are contrasted, counting loss, and use stochastic gradient descent to be combined searching with back-propagation method so that losing most
Small optimal weights parameter.
7. it is according to claim 6 based on shot and long term remember Recognition with Recurrent Neural Network bank outlets' excess reserve Forecasting Methodology,
Characterized in that, the model training stage, hidden layer number is 5, and each hidden layer includes 50 neurons, iteration 100000
It is secondary that LSTM models are trained.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710281286.9A CN106886846A (en) | 2017-04-26 | 2017-04-26 | A kind of bank outlets' excess reserve Forecasting Methodology that Recognition with Recurrent Neural Network is remembered based on shot and long term |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710281286.9A CN106886846A (en) | 2017-04-26 | 2017-04-26 | A kind of bank outlets' excess reserve Forecasting Methodology that Recognition with Recurrent Neural Network is remembered based on shot and long term |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN106886846A true CN106886846A (en) | 2017-06-23 |
Family
ID=59183052
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710281286.9A Pending CN106886846A (en) | 2017-04-26 | 2017-04-26 | A kind of bank outlets' excess reserve Forecasting Methodology that Recognition with Recurrent Neural Network is remembered based on shot and long term |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106886846A (en) |
Cited By (41)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107491508A (en) * | 2017-08-01 | 2017-12-19 | 浙江大学 | A kind of data base querying time forecasting methods based on Recognition with Recurrent Neural Network |
| CN107506414A (en) * | 2017-08-11 | 2017-12-22 | 武汉大学 | A kind of code based on shot and long term memory network recommends method |
| CN107609974A (en) * | 2017-09-26 | 2018-01-19 | 北京福布罗科技有限公司 | Method for testing pressure and device |
| CN107786369A (en) * | 2017-09-26 | 2018-03-09 | 广东电网有限责任公司电力调度控制中心 | Based on the perception of IRT step analyses and LSTM powerline network security postures and Forecasting Methodology |
| CN107909427A (en) * | 2017-10-25 | 2018-04-13 | 浙江大学 | A kind of Recognition with Recurrent Neural Network method for lifting recommended models timing driving ability |
| CN108090686A (en) * | 2017-12-29 | 2018-05-29 | 北京大学 | A kind of medical events risk-assessment method and system |
| CN108615096A (en) * | 2018-05-10 | 2018-10-02 | 平安科技（深圳）有限公司 | Server, the processing method of Financial Time Series and storage medium |
| CN109034493A (en) * | 2018-08-22 | 2018-12-18 | 阿里巴巴集团控股有限公司 | A kind of account balance prediction technique and device |
| CN109165727A (en) * | 2018-09-04 | 2019-01-08 | 成都品果科技有限公司 | A kind of data predication method based on Recognition with Recurrent Neural Network |
| CN109272164A (en) * | 2018-09-29 | 2019-01-25 | 清华大学深圳研究生院 | Learning behavior dynamic prediction method, device, equipment and storage medium |
| CN109447354A (en) * | 2018-10-31 | 2019-03-08 | 中国银行股份有限公司 | A kind of intelligent bank note distribution method and device based on GA-BP neural network |
| WO2019056499A1 (en) * | 2017-09-20 | 2019-03-28 | 平安科技（深圳）有限公司 | Prediction model training method, data monitoring method, apparatuses, device and medium |
| CN109685290A (en) * | 2019-02-11 | 2019-04-26 | 南方电网科学研究院有限责任公司 | Deep learning-based power consumption prediction method, device and equipment |
| CN109829818A (en) * | 2019-02-03 | 2019-05-31 | 中国银行股份有限公司 | Cash demand amount prediction technique, device, electronic equipment and readable storage medium storing program for executing |
| CN109934385A (en) * | 2019-01-29 | 2019-06-25 | 跨越速运集团有限公司 | Goods amount prediction technique and system based on length Memory Neural Networks |
| CN110019401A (en) * | 2017-12-25 | 2019-07-16 | 顺丰科技有限公司 | Part amount prediction technique, device, equipment and its storage medium |
| WO2019165944A1 (en) * | 2018-02-28 | 2019-09-06 | 中国银联股份有限公司 | Transition probability network based merchant recommendation method and system thereof |
| CN110245782A (en) * | 2019-05-20 | 2019-09-17 | 中国银行股份有限公司 | Prediction model training method, device and equipment |
| CN110255303A (en) * | 2019-06-13 | 2019-09-20 | 天津大学 | A kind of elevator waiting floor optimization method |
| CN110276492A (en) * | 2019-06-24 | 2019-09-24 | 深圳前海微众银行股份有限公司 | A kind of node data prediction technique and device |
| CN110363357A (en) * | 2019-07-17 | 2019-10-22 | 中国银行股份有限公司 | A kind of intelligence bank note distribution method and system |
| WO2019201001A1 (en) * | 2018-04-20 | 2019-10-24 | 阿里巴巴集团控股有限公司 | Fund forecasting method and device, and electronic device |
| CN110490392A (en) * | 2019-09-20 | 2019-11-22 | 中国银行股份有限公司 | Site bank note distribution method and device |
| CN110659787A (en) * | 2018-06-29 | 2020-01-07 | 北京九章云极科技有限公司 | Bank data analysis method and system |
| CN111090571A (en) * | 2019-12-18 | 2020-05-01 | 中国建设银行股份有限公司 | Information system maintenance method, device and computer storage medium |
| CN111130839A (en) * | 2019-11-04 | 2020-05-08 | 清华大学 | Flow demand matrix prediction method and system |
| CN111145541A (en) * | 2019-12-18 | 2020-05-12 | 深圳先进技术研究院 | Traffic flow data prediction method, storage medium, and computer device |
| CN111490995A (en) * | 2020-06-12 | 2020-08-04 | 支付宝(杭州)信息技术有限公司 | Model training method and device for protecting privacy, data processing method and server |
| CN111639792A (en) * | 2020-05-12 | 2020-09-08 | 深圳市星耀蓝图科技有限公司 | Method for intelligently adding bank ATM (automatic teller machine) money based on artificial intelligence |
| CN111738507A (en) * | 2020-06-19 | 2020-10-02 | 中国工商银行股份有限公司 | Bank clearing position fund payment amount prediction method, device, equipment and medium |
| CN112098714A (en) * | 2020-08-12 | 2020-12-18 | 国网江苏省电力有限公司南京供电分公司 | ResNet-LSTM-based electricity stealing detection method and system |
| CN112580505A (en) * | 2020-12-18 | 2021-03-30 | 平安银行股份有限公司 | Method and device for identifying opening and closing states of network points, electronic equipment and storage medium |
| WO2021082811A1 (en) * | 2019-10-29 | 2021-05-06 | 山东科技大学 | Foreign exchange time series prediction method |
| CN112802278A (en) * | 2020-12-31 | 2021-05-14 | 中国银行股份有限公司 | Method and device for adjusting teller machine, electronic device, and storage medium |
| CN113095934A (en) * | 2021-04-30 | 2021-07-09 | 中国银行股份有限公司 | Method and device for managing reserve payment of cash machine |
| CN113420902A (en) * | 2020-12-09 | 2021-09-21 | 上海东普信息科技有限公司 | Component prediction model training method, component prediction method and related equipment |
| CN113487110A (en) * | 2021-07-28 | 2021-10-08 | 中国银行股份有限公司 | Spare payment management method and device |
| CN113988461A (en) * | 2021-11-11 | 2022-01-28 | 中国工商银行股份有限公司 | Position prediction method, position prediction device, storage medium and electronic equipment |
| CN114781741A (en) * | 2022-05-06 | 2022-07-22 | 中国工商银行股份有限公司 | Option buying and selling opportunity determination method, apparatus, device, medium and program product |
| CN115130879A (en) * | 2022-07-05 | 2022-09-30 | 中国银行股份有限公司 | Method and device for controlling cash by bank |
| CN115440317A (en) * | 2022-07-20 | 2022-12-06 | 中南大学 | Prediction method and device for leached uranium concentration, electronic equipment and storage medium |
-
2017
- 2017-04-26 CN CN201710281286.9A patent/CN106886846A/en active Pending
Cited By (53)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107491508B (en) * | 2017-08-01 | 2020-05-26 | 浙江大学 | Database query time prediction method based on recurrent neural network |
| CN107491508A (en) * | 2017-08-01 | 2017-12-19 | 浙江大学 | A kind of data base querying time forecasting methods based on Recognition with Recurrent Neural Network |
| CN107506414A (en) * | 2017-08-11 | 2017-12-22 | 武汉大学 | A kind of code based on shot and long term memory network recommends method |
| CN107506414B (en) * | 2017-08-11 | 2020-01-07 | 武汉大学 | Code recommendation method based on long-term and short-term memory network |
| WO2019056499A1 (en) * | 2017-09-20 | 2019-03-28 | 平安科技（深圳）有限公司 | Prediction model training method, data monitoring method, apparatuses, device and medium |
| CN107609974A (en) * | 2017-09-26 | 2018-01-19 | 北京福布罗科技有限公司 | Method for testing pressure and device |
| CN107786369A (en) * | 2017-09-26 | 2018-03-09 | 广东电网有限责任公司电力调度控制中心 | Based on the perception of IRT step analyses and LSTM powerline network security postures and Forecasting Methodology |
| CN107909427A (en) * | 2017-10-25 | 2018-04-13 | 浙江大学 | A kind of Recognition with Recurrent Neural Network method for lifting recommended models timing driving ability |
| CN110019401A (en) * | 2017-12-25 | 2019-07-16 | 顺丰科技有限公司 | Part amount prediction technique, device, equipment and its storage medium |
| CN110019401B (en) * | 2017-12-25 | 2024-04-05 | 顺丰科技有限公司 | Method, device, equipment and storage medium for predicting part quantity |
| CN108090686A (en) * | 2017-12-29 | 2018-05-29 | 北京大学 | A kind of medical events risk-assessment method and system |
| CN108090686B (en) * | 2017-12-29 | 2022-01-25 | 北京大学 | Medical event risk assessment analysis method and system |
| WO2019165944A1 (en) * | 2018-02-28 | 2019-09-06 | 中国银联股份有限公司 | Transition probability network based merchant recommendation method and system thereof |
| WO2019201001A1 (en) * | 2018-04-20 | 2019-10-24 | 阿里巴巴集团控股有限公司 | Fund forecasting method and device, and electronic device |
| CN108615096A (en) * | 2018-05-10 | 2018-10-02 | 平安科技（深圳）有限公司 | Server, the processing method of Financial Time Series and storage medium |
| CN110659787A (en) * | 2018-06-29 | 2020-01-07 | 北京九章云极科技有限公司 | Bank data analysis method and system |
| CN109034493A (en) * | 2018-08-22 | 2018-12-18 | 阿里巴巴集团控股有限公司 | A kind of account balance prediction technique and device |
| CN109165727A (en) * | 2018-09-04 | 2019-01-08 | 成都品果科技有限公司 | A kind of data predication method based on Recognition with Recurrent Neural Network |
| CN109272164A (en) * | 2018-09-29 | 2019-01-25 | 清华大学深圳研究生院 | Learning behavior dynamic prediction method, device, equipment and storage medium |
| CN109447354A (en) * | 2018-10-31 | 2019-03-08 | 中国银行股份有限公司 | A kind of intelligent bank note distribution method and device based on GA-BP neural network |
| CN109934385A (en) * | 2019-01-29 | 2019-06-25 | 跨越速运集团有限公司 | Goods amount prediction technique and system based on length Memory Neural Networks |
| CN109829818A (en) * | 2019-02-03 | 2019-05-31 | 中国银行股份有限公司 | Cash demand amount prediction technique, device, electronic equipment and readable storage medium storing program for executing |
| CN109685290A (en) * | 2019-02-11 | 2019-04-26 | 南方电网科学研究院有限责任公司 | Deep learning-based power consumption prediction method, device and equipment |
| CN109685290B (en) * | 2019-02-11 | 2023-06-16 | 南方电网科学研究院有限责任公司 | Power consumption prediction method, device and equipment based on deep learning |
| CN110245782A (en) * | 2019-05-20 | 2019-09-17 | 中国银行股份有限公司 | Prediction model training method, device and equipment |
| CN110255303A (en) * | 2019-06-13 | 2019-09-20 | 天津大学 | A kind of elevator waiting floor optimization method |
| CN110276492A (en) * | 2019-06-24 | 2019-09-24 | 深圳前海微众银行股份有限公司 | A kind of node data prediction technique and device |
| CN110363357A (en) * | 2019-07-17 | 2019-10-22 | 中国银行股份有限公司 | A kind of intelligence bank note distribution method and system |
| CN110490392A (en) * | 2019-09-20 | 2019-11-22 | 中国银行股份有限公司 | Site bank note distribution method and device |
| WO2021082811A1 (en) * | 2019-10-29 | 2021-05-06 | 山东科技大学 | Foreign exchange time series prediction method |
| CN111130839A (en) * | 2019-11-04 | 2020-05-08 | 清华大学 | Flow demand matrix prediction method and system |
| CN111130839B (en) * | 2019-11-04 | 2021-07-16 | 清华大学 | Flow demand matrix prediction method and system |
| CN111090571A (en) * | 2019-12-18 | 2020-05-01 | 中国建设银行股份有限公司 | Information system maintenance method, device and computer storage medium |
| CN111145541B (en) * | 2019-12-18 | 2021-10-22 | 深圳先进技术研究院 | Traffic flow data prediction method, storage medium, and computer device |
| CN111090571B (en) * | 2019-12-18 | 2024-01-23 | 中国建设银行股份有限公司 | Maintenance method, device and computer storage medium for information system |
| CN111145541A (en) * | 2019-12-18 | 2020-05-12 | 深圳先进技术研究院 | Traffic flow data prediction method, storage medium, and computer device |
| CN111639792A (en) * | 2020-05-12 | 2020-09-08 | 深圳市星耀蓝图科技有限公司 | Method for intelligently adding bank ATM (automatic teller machine) money based on artificial intelligence |
| CN111639792B (en) * | 2020-05-12 | 2023-04-18 | 深圳市星耀蓝图科技有限公司 | Method for intelligently adding bank ATM (automatic teller machine) money based on artificial intelligence |
| CN111490995A (en) * | 2020-06-12 | 2020-08-04 | 支付宝(杭州)信息技术有限公司 | Model training method and device for protecting privacy, data processing method and server |
| CN111738507A (en) * | 2020-06-19 | 2020-10-02 | 中国工商银行股份有限公司 | Bank clearing position fund payment amount prediction method, device, equipment and medium |
| CN112098714A (en) * | 2020-08-12 | 2020-12-18 | 国网江苏省电力有限公司南京供电分公司 | ResNet-LSTM-based electricity stealing detection method and system |
| CN112098714B (en) * | 2020-08-12 | 2023-04-18 | 国网江苏省电力有限公司南京供电分公司 | Electricity stealing detection method and system based on ResNet-LSTM |
| CN113420902A (en) * | 2020-12-09 | 2021-09-21 | 上海东普信息科技有限公司 | Component prediction model training method, component prediction method and related equipment |
| CN112580505A (en) * | 2020-12-18 | 2021-03-30 | 平安银行股份有限公司 | Method and device for identifying opening and closing states of network points, electronic equipment and storage medium |
| CN112580505B (en) * | 2020-12-18 | 2023-09-19 | 平安银行股份有限公司 | Method and device for identifying network point switch door state, electronic equipment and storage medium |
| CN112802278A (en) * | 2020-12-31 | 2021-05-14 | 中国银行股份有限公司 | Method and device for adjusting teller machine, electronic device, and storage medium |
| CN113095934A (en) * | 2021-04-30 | 2021-07-09 | 中国银行股份有限公司 | Method and device for managing reserve payment of cash machine |
| CN113487110A (en) * | 2021-07-28 | 2021-10-08 | 中国银行股份有限公司 | Spare payment management method and device |
| CN113988461A (en) * | 2021-11-11 | 2022-01-28 | 中国工商银行股份有限公司 | Position prediction method, position prediction device, storage medium and electronic equipment |
| CN114781741A (en) * | 2022-05-06 | 2022-07-22 | 中国工商银行股份有限公司 | Option buying and selling opportunity determination method, apparatus, device, medium and program product |
| CN115130879A (en) * | 2022-07-05 | 2022-09-30 | 中国银行股份有限公司 | Method and device for controlling cash by bank |
| CN115440317B (en) * | 2022-07-20 | 2023-05-26 | 中南大学 | Method and device for predicting uranium leaching concentration, electronic equipment and storage medium |
| CN115440317A (en) * | 2022-07-20 | 2022-12-06 | 中南大学 | Prediction method and device for leached uranium concentration, electronic equipment and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106886846A (en) | A kind of bank outlets' excess reserve Forecasting Methodology that Recognition with Recurrent Neural Network is remembered based on shot and long term | |
| CN108022001A (en) | Short term probability density Forecasting Methodology based on PCA and quantile estimate forest | |
| CN109785003A (en) | A kind of Pharmaceutical retail industry medicine sales forecasting system and method | |
| CN109118013A (en) | A kind of management data prediction technique, readable storage medium storing program for executing and forecasting system neural network based | |
| CN108830430A (en) | Convolutional neural networks multiple spot regressive prediction model for traffic flow forecasting | |
| CN107622329A (en) | The Methods of electric load forecasting of Memory Neural Networks in short-term is grown based on Multiple Time Scales | |
| CN111178611B (en) | Method for predicting daily electric quantity | |
| CN107832897A (en) | A kind of Stock Price Forecasting method based on deep learning | |
| Enke et al. | A hybrid neuro-fuzzy model to forecast inflation | |
| CN112232561B (en) | Power load probability prediction method based on constrained parallel LSTM fractional regression | |
| CN111626785A (en) | CNN-LSTM network fund price prediction method based on attention combination | |
| CN110348608A (en) | A kind of prediction technique for improving LSTM based on fuzzy clustering algorithm | |
| CN104751363B (en) | Stock Forecasting of Middle And Long Period Trends method and system based on Bayes classifier | |
| CN110503104A (en) | A kind of remaining parking stall quantitative forecasting technique in short-term based on convolutional neural networks | |
| CN107463993A (en) | Medium-and Long-Term Runoff Forecasting method based on mutual information core principle component analysis Elman networks | |
| CN113554466A (en) | Short-term power consumption prediction model construction method, prediction method and device | |
| Bhandari et al. | An artificial intelligence ATM forecasting system for hybrid neural networks | |
| Struga et al. | Bitcoin Price Prediction with Neural Networks. | |
| CN111882114B (en) | Short-time traffic flow prediction model construction method and prediction method | |
| CN104751254A (en) | Line loss rate prediction method based on non-isometric weighted grey model and fuzzy clustering sorting | |
| Liu et al. | Stock price trend prediction model based on deep residual network and stock price graph | |
| CN107704944B (en) | Construction method of stock market fluctuation interval prediction model based on information theory learning | |
| Liu et al. | A new LASSO-BiLSTM-based ensemble learning approach for exchange rate forecasting | |
| Asad et al. | A long-short-term-memory based model for predicting atm replenishment amount | |
| Turki et al. | Study Estimating hourly traffic flow using Artificial Neural Network: A M25 motorway case |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| TA01 | Transfer of patent application right | ||
| TA01 | Transfer of patent application right |
Effective date of registration: 20181227 Address after: 410000 No. 678 Qingshan Road, Yuelu District, Changsha City, Hunan Province Applicant after: Hunan Creator Information Technologies Co.,Ltd. Address before: 410083 932 south foot Road, Yuelu District, Changsha, Hunan Applicant before: Central South University Applicant before: Hunan Hyderabad Data Technology Co. Ltd. |
|
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20170623 |