CN120746696A - Intelligent evaluation pricing method and system for individual bad assets based on graphic neural network - Google Patents

Abstract

The invention relates to the technical field of intersection of financial science and technology and artificial intelligence, and particularly discloses an intelligent evaluation pricing method for individual bad assets based on a graphic neural network, which comprises the steps of acquiring original data in real time, extracting multi-modal feature data based on the original data, and fusing the multi-modal feature data; establishing a debtor association network map based on the debtor information, and performing risk assessment based on the debtor association network map and the fused multi-mode feature data; the method comprises the steps of predicting the recovery rate of the assets, quantifying uncertainty, optimizing the asset package combination based on a Markov decision process model, pricing based on a risk assessment result, an asset recovery rate prediction result and an optimization strategy, and by means of a deep fusion technology of multi-mode data and strong nonlinear fitting capacity of a deep learning model, the system can more fully mine value information contained in the data and help investors to avoid potential major investment risks by constructing and analyzing a debtor associated map.

Description

Intelligent evaluation pricing method and system for individual bad assets based on graphic neural network

Technical Field

The invention relates to the technical field of intersection of financial science and technology and artificial intelligence, in particular to an intelligent evaluation pricing method for individual bad assets based on a graphic neural network.

Background

As economic cycles fluctuate and credit scales expand, the size of bad assets on commercial banking personal loans continues to increase, posing a potential threat to the stability of financial systems. The disposal of bad assets has become one of the important links and core businesses for financial institution risk management. Traditional bad asset pack evaluation pricing mainly relies on manual experience judgment and sampling audit, and has the following obvious technical defects:

first, the evaluation efficiency is low, and it is difficult to meet the demand of fast market trade. The traditional method needs a large amount of manual examination of loan files, carefully analyzes complex information such as repayment capacity, mortgage value, guarantee condition and the like of borrowers, and has long time consumption and high labor cost. For property bags containing thousands or even tens of thousands of loans, a complete evaluation cycle often takes weeks or even months.

Second, pricing accuracy is inadequate and subjective factors are more relevant. Manual assessment is difficult to adequately consider numerous complex risk factors and their interactions, subject to subjective bias and experience limitations. There may be a large variance in valuation of the same asset pack by different valuators or institutions, which directly affects the scientificity of the transaction decision and the effectiveness of asset recovery.

Third, risk identification capability is limited, especially for lack of identification of implicit and associated risks. The conventional method is difficult to deeply mine and find the implicit association relationship between debtors, such as common guarantee, family association, enterprise association, and the like, so that systematic risks and conduction risks cannot be effectively identified and quantified. For example, multiple debtors within the same guaranty circle may be at risk for associated default, but such deep associations are often ignored or imperceptible in traditional evaluations.

Fourth, dynamic adaptation is poor and model updates lag market changes. External factors such as market environment, policy regulations, regional economics, etc. are constantly changing, and these changes directly affect the recovery prospects of the undesirable assets. Once the traditional static evaluation model is established, the traditional static evaluation model is difficult to adjust in time according to the latest market dynamics, so that the pricing result deviates from the actual market situation.

Disclosure of Invention

The invention aims to provide a method for intelligently evaluating and pricing individual bad assets based on a graphic neural network, so as to solve the problems in the background art.

In order to achieve the purpose, the invention provides the following technical scheme that the intelligent evaluation pricing method for the individual bad assets based on the graphic neural network comprises the following steps:

Acquiring original data in real time, wherein the original data comprises bank internal data, bank external data and market data;

Extracting multi-mode feature data based on the original data, and fusing the multi-mode feature data;

establishing a debtor association network map based on the debtor information, and performing risk assessment based on the debtor association network map and the fused multi-mode feature data;

Predicting asset recovery rate and quantifying uncertainty based on Bayesian deep learning method and risk conduction analysis result;

optimizing the asset pack combination based on the Markov decision process model to generate an optimization strategy;

pricing is performed based on the risk assessment results, asset recovery prediction results, and optimization strategies to generate pricing advice.

As a further aspect of the present invention, the step of extracting the multi-modal feature data based on the raw data specifically includes:

Cleaning and standardizing the original data;

constructing basic features and derivative features and calculating;

and carrying out extraction analysis on structural features, text features and time sequence features of the cleaned and standardized data.

As a further aspect of the present invention, the step of fusing the multi-modal feature data specifically includes:

Calculating a correlation matrix among the structural features, the text features and the time sequence features;

Assigning fusion weights to the structured features, text features, and timing features based on an attention mechanism;

and fusing the multi-modal feature data based on the fusion weight to generate a multi-modal feature representation vector.

As a further aspect of the present invention, the step of establishing a debtor-associated network map based on the debtor information, and performing risk assessment based on the debtor-associated network map and the integrated multimodal feature data specifically includes:

establishing a debtor association network map based on the debtor information;

constructing a dynamic heterogeneous graph network model based on the liability person associated network graph;

determining a risk conduction path and a risk conduction rule;

and performing risk assessment based on the dynamic heterograph network model.

As a further scheme of the invention, the asset pack combination is optimized based on the Markov decision process model, and the step of generating the optimization strategy specifically comprises the following steps:

defining an objective function consisting of maximizing the expected combined benefit and minimizing the combined risk;

Constraint conditions are determined, including budget constraints, concentration constraints, and flowability constraints.

And (3) constructing a Markov decision process model, optimizing the asset pack combination based on the Markov decision process model, and generating an optimization strategy.

As a further aspect of the present invention, the step of generating pricing advice specifically includes:

Constructing a weighted scoring system based on the risk assessment result, the asset recovery rate prediction result and the optimization strategy;

Adjusting the risk based on the uncertainty;

introducing a market comparison method to perform preliminary calibration;

a pricing model is built based on an interpretable machine learning method, and pricing suggestions are generated.

As a further aspect of the present invention, the pricing model is expressed as:

;

Wherein B is the present value of the expected recoverable amount of a single or packaged asset based on model prediction, D is the discount rate determined by comprehensively considering the time value of the funds, the specific risk premium of the asset and the liquidity premium factor, and M is the adjustment coefficient of the basic value according to the current market supply and demand condition, the average premium or discount level of the comparable transaction and the macro economic environment factor. U is an adjustment factor set based on the quantized prediction uncertainty.

As a further aspect of the present invention, updating the pricing proposal based on the real-time raw data is also included.

As a further aspect of the present invention, the step of updating the pricing proposal based on the real-time raw data specifically includes:

acquiring original data, and performing data drift detection and concept drift detection based on the original data;

generating performance degradation early warning based on the detection result;

and performing iterative optimization updating by adopting incremental learning and transfer learning based on performance degradation early warning.

The invention also discloses a system for intelligently evaluating and pricing the lending bad assets based on the graph neural network, which comprises the following steps:

the multi-source data acquisition module is used for acquiring original data in real time, wherein the original data comprises bank internal data, bank external data and market data;

The multi-mode feature fusion module is used for extracting multi-mode feature data based on the original data and fusing the multi-mode feature data;

The debtor association graph construction and risk assessment module is used for building a debtor association network graph based on the debtor information and carrying out risk assessment based on the debtor association network graph and the fused multi-mode feature data;

The time sequence prediction and uncertainty quantification module is used for predicting the asset recovery rate and quantifying the uncertainty based on a Bayesian deep learning method and a risk conduction analysis result;

The asset pack combination optimization module is used for optimizing the asset pack combination based on the Markov decision process model to generate an optimization strategy;

The intelligent pricing decision module is used for pricing based on the risk assessment result, the asset recovery rate prediction result and the optimization strategy to generate pricing advice;

Compared with the prior art, the invention has the beneficial effects that asset pack evaluation work which usually needs weeks or even longer in the traditional mode can be greatly shortened to be completed within hours or days by advanced automatic data processing flow and efficient deep learning model reasoning, so that the overall work efficiency of bad asset disposal is greatly improved;

By means of the deep fusion technology of the multi-mode data and the strong nonlinear fitting capability of the deep learning model, the system can more fully mine the value information contained in the data, and compared with the traditional manual evaluation method, the pricing error can be obviously reduced, and the pricing error is more close to the real intrinsic value of the asset;

By constructing and analyzing the liability person association graph and applying the graph neural network technology, the system can find and quantify complex and implicit association relations (such as a guarantee chain, an enterprise assignment and the like) among the liability persons, thereby effectively identifying systematic risks and potential risk conduction paths and helping investors avoid potential major investment risks;

The application of the innovative Bayesian deep learning method ensures that the system not only can give the estimated value of the expected point of asset recovery, but also can provide the confidence interval and complete risk probability distribution of the predicted result, thereby providing more comprehensive and deeper information support and risk consideration dimension for investment decision-makers when facing uncertainty;

the introduced reinforcement learning method can fully consider the correlation among various assets in the asset package and the risk benefit characteristics, and dynamically optimize the composition and weight distribution of the asset package, so that the overall investment portfolio benefit after risk adjustment is effectively improved under given risk preference;

the built-in self-adaptive updating mechanism ensures that each model in the system can timely respond to the change of the market environment, the appearance of new data and the feedback of the actual treatment effect, and the accuracy of model prediction and the market fitting degree are maintained through continuous learning and iteration;

The interpretable machine learning method and the generated detailed evaluation analysis report adopted by the system provide a solid and data-driven scientific basis for investment decision, pricing strategy and risk management of a management layer, can explain the cause of key decision to a certain extent, and meet the supervision requirement.

Drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following description will briefly introduce the drawings that are needed in the embodiments or the description of the prior art, and it is obvious that the drawings in the following description are only some embodiments of the present invention.

FIG. 1 is a flow chart of a method for intelligently evaluating and pricing individual lended bad assets based on a graph neural network, which is provided by an embodiment of the invention.

Detailed Description

In order to make the technical problems, technical schemes and beneficial effects to be solved more clear, the invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention.

Fig. 1 is a flow chart of a method for intelligently evaluating and pricing individual bad assets based on a graph neural network, and in an embodiment of the invention, the method includes steps S100 to S600:

Acquiring original data in real time, wherein the original data comprises bank internal data, bank external data and market data;

Extracting multi-mode feature data based on the original data, and fusing the multi-mode feature data;

establishing a debtor association network map based on the debtor information, and performing risk assessment based on the debtor association network map and the fused multi-mode feature data;

Predicting asset recovery rate and quantifying uncertainty based on Bayesian deep learning method and risk conduction analysis result;

optimizing the asset pack combination based on the Markov decision process model to generate an optimization strategy;

pricing is performed based on the risk assessment results, asset recovery prediction results, and optimization strategies to generate pricing advice.

In this embodiment, the internal data of the bank mainly refers to the data in the system of the commercial bank itself, such as detailed loan contract terms, historical repayment records, collection process records (text, voice-to-text), guarantee contracts and guarantee person information.

The external data of the bank refers to supplementary data acquired from outside the bank, such as personal credit reports, lawsuits and execution information, business registration and management information, public social media utterances (required compliance acquisition), and the like.

Market data reflects market environment and data of comparable trade conditions, such as historical trading prices for similar bad asset packs, macro economic indicators (GDP speed increase, rate of loss), related industry scenic index, latest regulatory policy and regulation changes, etc.

As a preferred embodiment of the present invention, the step of extracting the multi-modal feature data based on the raw data specifically includes:

Cleaning and standardizing the original data;

constructing basic features and derivative features and calculating;

and carrying out extraction analysis on structural features, text features and time sequence features of the cleaned and standardized data.

In this embodiment, the data missing values are processed by using a suitable multiple interpolation method (Multiple Imputation) for different types of missing data, such as interpolation based on predictive models of other relevant features or interpolation based on the mean/median/mode distribution of similar samples.

And (3) detecting and processing abnormal values, namely identifying the abnormal values in the data by using statistical methods such as an Isolation Forest algorithm, a box diagram, a Z-score and the like, and judging whether to correct, delete or perform special marking processing according to business logic.

Data normalization/normalization the numerical features were normalized (Z-score normalized to 0 mean and 1 standard deviation) and the class features were single-heat coded.

The basic characteristics are constructed by directly extracting key information from the original data, such as the total loan amount, the current overdue days, the guarantee mode (credit, mortgage and assurance), the geographical position of borrowers and the like.

The method comprises the steps of constructing derivative features, namely generating new features with more information based on basic features through calculation or logic combination, such as historical average overdue rate change trend, borrower repayment capability index (such as debt income ratio), mortgage estimated devaluation rate, promotion frequency and the like, constructing interaction features, namely exploring combination and interaction items among different basic features or derivative features to capture nonlinear relations, wherein interaction of 'age' and 'income level' can influence repayment willingness.

And (3) extracting and processing structural features:

Numerical features are generally directly input as a model after normalization or used for derivative feature calculation, category features are generally converted into numerical dense vectors through a single thermal coding or embedding layer (Embedding Layer) so as to be input into a deep learning model, and the spliced structural features are subjected to nonlinear transformation through a multi-layer perceptron (MLP) to extract higher-order abstract feature representations.

Text feature extraction and analysis:

Key information fragments such as the frequency of collection, communication attitude (such as cooperation and negative), whether repayment commitment is made, the reason of difficulty is mentioned and the like are extracted from the text, the important legal states such as the stage of litigation (such as first trial, second trial and execution), the judgment result, the execution progress condition, whether the asset is sealed and frozen and the like are automatically identified, text data are encoded by using a pre-trained BERT model or a variant (such as FinBERT) optimized for the financial field, and deep semantic feature vectors are extracted.

Timing feature extraction and pattern recognition:

analyzing time sequence data formed by historical repayment amount, overdue days and the like of the borrower to mine repayment regularity, seasonal fluctuation and the like of the borrower, evaluating change trends of different repayment measures on repayment behaviors of the borrower at different time points, such as whether repayment probability is improved in a short period after repayment, and extracting complex time sequence modes and dependency relations by using a long and short period memory network (LSTM), a gate control circulation unit (GRU), a Transformer and other sequence models.

As a preferred embodiment of the present invention, the step of fusing the multi-modal feature data specifically includes:

Calculating a correlation matrix among the structural features, the text features and the time sequence features;

Assigning fusion weights to the structured features, text features, and timing features based on an attention mechanism;

and fusing the multi-modal feature data based on the fusion weight to generate a multi-modal feature representation vector.

In this embodiment, firstly, a correlation matrix or similarity score between features of different modes (such as a structured feature, a text feature vector, and a time sequence feature vector) is calculated, and secondly, different fusion weights are adaptively allocated to the features of each mode according to the contribution degree of the features of the different modes to the final prediction task through an attention mechanism (such as self-attention and cross-attention).

Finally, each mode characteristic after weighting or some combination thereof (such as dimension reduction through a full connection layer after splicing) is generated into a unified multi-mode characteristic expression vector with more comprehensive information.

In particular, for structured data(WhereinIn order to obtain the number of samples,To structure feature dimensions), its deep feature representation is extracted by a multi-layer perceptron (MLP):

;

Wherein the method comprises the steps of AndRepresenting the weight matrix and the bias vector respectively,Representing an activation function.

For text data(E.g., collect records, legal documents, etc.), deep semantic features are extracted using a pre-trained BERT model, and their domain relevance is enhanced by adaptive fine tuning for financial domain text, while sequence information is retained in combination with position coding (Positional Encoding):

;

For time series data (WhereinFor the time step size of the time step,For a time-series feature dimension, such as a historical repayment sequence), a modified transducer encoder (Transformer Encoder) architecture is employed in conjunction with a one-dimensional time-series convolution (TemporalConv 1D) to capture local time-series dependencies:

;

and designing a Cross-Modal Attention mechanism (Cross-Modal Attention) to perform interaction and fusion among different Modal features, and calculating Attention weights:

;

Wherein the method comprises the steps of ,,Representing different modal characteristicsQuery (Query) and Key (Key) obtained through linear transformation.

The multi-modal characteristics are adaptively fused through a gating mechanism, so that the contribution degree of each modal characteristic is dynamically adjusted:

;

;

Wherein the method comprises the steps of Is a gate-control signal, and the control signal is a gate-control signal,Is a parameter that can be learned and is,The element-level product is represented by a number of elements,Representing a stitching operation.

As a preferred embodiment of the present invention, the step of establishing a debtor-associated network map based on the debtor information, and performing risk assessment based on the debtor-associated network map and the fused multimodal feature data specifically includes:

establishing a debtor association network map based on the debtor information;

constructing a dynamic heterogeneous graph network model based on the liability person associated network graph;

determining a risk conduction path and a risk conduction rule;

and performing risk assessment based on the dynamic heterograph network model.

In this embodiment, explicit association discovery refers to directly extracting a well-defined association from data, such as a well-defined common borrower relationship in a loan contract, a guarantee relationship in a guarantee contract, and a relative relationship (e.g., spouse, parent-child) recorded in borrower information.

Implicit association mining refers to the discovery of potential, not easily directly observable associations through data mining techniques, such as multiple borrowers working units being identical, living or communicating addresses being in close proximity, there being non-public past funds records (requiring compliance to acquire data), and the discovery of new entities and their potential associations between them from unstructured text (e.g., revenue records, news stories) using entity recognition (NER) and Relationship Extraction (RE) techniques in natural language processing.

Node types are defined, namely, the nodes are divided into main debtor nodes, guarantor nodes (individuals/enterprises), associated enterprise nodes, even key contact nodes and the like according to different roles.

Defining edge types, namely dividing the edges into a guarantee edge, a relative edge, a common borrowing edge, an enterprise subordinate/investment edge, a business incoming and outgoing edge, a geographic adjacent edge and the like according to different association properties, and endowing different types of edges with different initial weights.

Dynamic calculation of edge weights the edge weights can be dynamically calculated and adjusted based on the associated strength (e.g., amount of vouch for), risk correlation (e.g., historical common breach probability), efficiency of information transfer, and the like.

The risk propagation rules and decay functions are defined by setting how the risk of a node is conducted to neighboring nodes through different types of edges when it breaks down or the risk level increases, and how the strength of the conduction decays with distance or relationship type.

End-to-end learning is performed on the constructed graph through a graph neural network (GNNs, such as GRAPHSAGE, etc.), automatically learning and capturing complex risk conduction patterns, rather than relying solely on preset rules.

Systematic risk and community risk assessment:

key centrality indexes of the network topology structure are calculated, such as centrality (DEGREE CENTRALITY, the connection number of the nodes), betweenness centrality (Betweenness Centrality, the number of times the nodes serve as shortest path bridges), feature vector centrality (Eigenvector Centrality, the degree of connection of the nodes to important nodes) and the like.

Identifying system importance nodes (SYSTEMICALLY IMPORTANT NODES, SINs) or "large but not overwhelming" liability entities, and closely connected risk communities (Community Detection).

Vulnerability (Vulnerability) and robustness (Robustness) of the entire associated network are evaluated, such as the extent and speed of risk spread after certain critical nodes "fail".

Specifically, constructing an isomerism mapWhereinRepresenting a collection of nodes (e.g. debtors, guarantors etc.),Represents a set of edges (representing the association between nodes),Representing a set of relationship types (e.g., guaranty relationships, relatives, etc.);

design of relationship-aware messaging mechanisms (Relation-AWARE MESSAGE PASSING) for different types of relationships NodeFrom its type ofIs a neighbor of (a)Receiving information and aggregating:

;

Wherein the method comprises the steps of Is a nodeIs characterized in that,Is a relationship typeA characteristic transformation matrix is used to transform the data,Is the attention factor, typically calculated by:

;

is a relationship type Is used for the vector of the attention parameter of (c),Representing a stitching operation.

Information transmitted from different types of relations is aggregated by adopting a Multi-Head Attention mechanism (Multi-Head Attention), and a characteristic representation after node updating is obtained:

;

Wherein the method comprises the steps ofIs the number of attention tips and,Is the firstSub-attention-head pair relationship typeWeights for transformation of the information of (a).

Introducing a time decay factorTo process timing information in the dynamic graph such that the impact of historical information on the current node state weakens over time:

;

Wherein the method comprises the steps of Is a nodeLast time stampIs used for the control of the state of (a),Is the information currently obtained by messaging updates.

Obtaining risk characterization of an entire asset pack or a particular sub-image level through a pooling operationFor example using pooling based on an attention mechanism:

;

Wherein the method comprises the steps of Is a nodeThe resulting representation of the characteristics of the device,Is a nodeImportance of (2) is weighted by a small multi-layer perceptronAnd (5) calculating to obtain the product.

The asset recovery prediction and uncertainty quantization stage first builds an accurate prediction model:

Well-defined predicted target variables may be the expected Recovery Rate (Recovery Rate) of the individual asset, the specific expected Recovery amount, the Time required to expect Recovery to complete (Time to Recovery), or the cost that may occur during Recovery, etc.

The appropriate input features are selected using the multimodal fusion feature generated in stage S2 and the node level or sub-level risk characterization learned from the graph network in stage S3 as input to the model.

Advanced predictive model architectures such as the Bayesian Neural Network (BNN) described above, or gradient-lifted trees (e.g., XGBoost, lightGBM) are designed and employed for integration with deep learning models (Ensemble Learning).

Model training is carried out:

the loss function combines regression losses such as Mean Square Error (MSE), mean Absolute Error (MAE) and the like, and for BNN, KL divergence term is also needed to be contained as regularization, so that model fitting degree and complexity are balanced, and an efficient AdamW self-adaptive learning rate optimizer is selected.

Model overfitting is prevented using Dropout, L2 regularization (weight decay), early stop (Early Stopping), etc.

Occasional uncertainties, due to inherent noise, measurement errors, or inherent randomness of the data itself, cannot be eliminated even with infinite data. Quantization can be achieved by the model directly outputting the variance of the predicted value.

Cognitive uncertainty results from uncertainty in the model parameters themselves, typically due to insufficient amounts of training data or inadequate model structure. Quantification is performed by bayesian methods (e.g. weight distribution in BNN).

The total uncertainty is obtained by adding the variance of the data uncertainty and the model uncertainty.

Generating comprehensive prediction distribution information:

providing a point estimate, i.e., predicting the expected or median of the distribution, such as the desired recovery.

Providing an interval estimate, i.e. a confidence interval (Confidence Interval, CI) or a prediction interval (Prediction Interval, PI) giving a prediction result, e.g. 95% CI, indicates that there is a 95% probability that the true value falls within this interval.

And providing a complete probability distribution map, namely a complete Probability Density Function (PDF) and a Cumulative Distribution Function (CDF) of the visual prediction result, and providing the most comprehensive risk information for a decision maker.

Specifically, a Bayesian neural network is constructed, and BNN is different from the traditional neural network in terms of network weightGiving a priori distributionTypically, a gaussian distribution is assumed:

;

Wherein the method comprises the steps of Is a single weight in the network and,Is the mean value of the values,Is the a priori variance.

Approximation of true posterior distribution using variational inference (Variational Inference, VI)(Given dataDistribution of post weights). Introducing a parameterized variation distribution(Typically gaussian distribution is also chosen) to approximate the posterior:

;

Wherein the method comprises the steps of Is the variation parameter, i.e., the mean and variance of each weight.

Optimizing variational parameters by maximizing the lower bound of evidence (Evidence Lower Bound, ELBO):

;

The first term is expected log likelihood, the fitting degree of the model to the data is measured, and the second term is KL divergence between the variational distribution and the prior distribution, which is used as a regularization term to prevent overfitting.

Gradient calculations and sampling are performed using re-parameterization techniques to optimize by standard back-propagation algorithms. For Gaussian distribution, weightsCan be expressed as:

;

Wherein the method comprises the steps of AndIs the mean and standard deviation vector of the variation distribution,Is noise sampled from a standard normal distribution,Representing the element-level product.

The prediction distribution is obtained by monte carlo sampling. During the prediction phase, learned posterior distributionMedium multiple sampling weightsForward propagation is performed for a plurality of times to obtain a series of prediction results, so that the prediction distribution is obtained by approximate integration:

;

Wherein the method comprises the steps of Is the number of samples to be taken,Is an input to which the user is exposed,Is the predicted output.

Calculating the uncertainty of the prediction, and decomposing the prediction into accidental uncertainty and cognitive uncertainty:

Occasional uncertainty (data uncertainty) reflecting noise and inherent randomness of the data itself, estimated by modeling the variance of the direct predicted output, e.g. (If the model outputs variance).

Cognitive uncertainty (model uncertainty) reflecting the uncertainty of the model to the parameters due to insufficient data or model structure limitations, which can be estimated by the variance of the predicted values obtained by multiple samples, e.gWhereinIs the firstThe predicted mean value of the sub-samples,Is the average of all the sampled prediction means.

As a preferred embodiment of the present invention, the step of generating an optimization strategy specifically includes:

defining an objective function consisting of maximizing the expected combined benefit and minimizing the combined risk;

Constraint conditions are determined, including budget constraints, concentration constraints, and flowability constraints.

And (3) constructing a Markov decision process model, optimizing the asset pack combination based on the Markov decision process model, and generating an optimization strategy.

In the present embodiment, the combined optimization objective function is explicitly defined:

Maximizing the expected combined return, i.e., the weighted average expected recovery amount or recovery of the selected subset of assets, minimizing the combined risk, i.e., typically measured by the overall risk indicator of the combination, such as the variance, standard deviation, risk value (VaR), or conditional risk value (CVaR) of the portfolio.

Balance benefit versus risk, e.g., maximize risk adjusted benefit index such as the Sharpe Ratio (Sharpe Ratio), the Sorpe Ratio (Sortino Ratio), etc.

Budget constraints mean that the total investment cost (e.g., purchase price) of the selected asset cannot exceed a preset total budget upper limit. Concentration constraints refer to setting an upper limit on the asset ratio for a single region, a single industry, a single debtor, or a single risk level to spread the risk.

The flowability constraint refers to limiting the configuration proportion of the assets with different period limits by considering the expected recovery period of the assets, so that the flowability of the whole combination is ensured to meet the requirement.

Other business constraints such as minimum asset count requirements, inclusion or exclusion of certain types of assets, etc.

Dynamic optimization using reinforcement learning:

the selection and configuration problems of asset packs are accurately modeled as a Markov Decision Process (MDP), defining states, actions, rewards, and state transition probabilities.

An Agent is trained by a deep reinforcement learning algorithm PPO to learn continuously in interactions with the simulated environment to find an optimal strategy (Policy) to achieve an optimization objective under given constraints.

The reinforcement learning model can be designed to dynamically adjust its portfolio strategy based on market feedback and new asset information, with some adaptation.

The recommended subset of assets is output, explicitly listing which bad assets the model suggests to purchase or reserve under the current conditions.

The weight allocation of the selected assets is given by allocating an appropriate investment ratio or amount to each selected asset.

Expected revenue targets, various risk targets (e.g., vaR, CVaR, volatility) and other key performance targets (KPIs) of the optimized combination are provided.

Specifically, a Markov decision process (Markov Decision Process, MDP) is defined to formalize asset pack selection and configuration issues, state spaceInformation describing the current environment, such as the current portfolio of selected assets in the asset pack, market overall status indicators (e.g., macro economic index, industry Jing Qidu), risk indicators for the current portfolio (e.g., vaR, CVaR), etc., action spaceActions that agents (agents) may take, such as selecting which assets to join in a portfolio in a candidate asset list, how much investment weight or proportion of funds to allocate to the selected assets, rewarding functionsEvaluation of the stateTake action downwardsThe immediate return obtained later is typically designed as a risk-adjusted return, such as a change in the summer ratio, the sovereign ratio, or directly as a net return minus a risk penalty term.

A special Actor-Critic network architecture is designed, which is a policy optimization framework commonly used in DRL:

Actor network (policy network): Input the current state The output taking various possible actions in this stateThe probability distribution (specific value of motion is output for continuous motion space) of which the parameters are。

Critic network (value network): And/or Value for evaluating current stateValue) or value of state-action pairValue), the parameters of which are respectivelyAnd. The Critic network helps the Actor network to update policies better.

A risk-sensitive reward function is introduced to ensure that the optimization process fully considers risk factors:

;

Wherein the method comprises the steps of Is the moment of timeIs added to the original benefit of (1),Is the distribution of the losses and,Is at a confidence levelThe risk value of the conditions below is,Is an index for measuring the diversity of investment portfolios (such as the reciprocal of the herford index),AndIs the corresponding penalty or prize coefficient.

The policy network is updated using a near-end policy optimization (Proximal Policy Optimization, PPO) algorithm. PPO is an advanced strategy gradient method, which limits the magnitude of each strategy update by introducing a tailored objective function, thus ensuring the stability of the learning process:

;

Wherein, the Is the probability ratio of the new strategy to the old strategy,Is a merit function (ADVANTAGE FUNCTION),Is a clipping parameter.

The learning efficiency and the sample utilization rate are improved through experience playback (Experience Replay) and priority sampling (Prioritized Experience Replay, PER), wherein an experience playback mechanism stores a sequence (state, action, rewards and next state) generated by interaction of an agent and the environment in a playback buffer zone, and small batches of samples are randomly extracted from the experience playback mechanism for learning during training, so that the data correlation is broken. The preferential sampling is based on the importance of the sample (usually by time differential errorMeasured by the size of (a) of) the number of samples to be sampled is assigned to different sampling probabilitiesThe important samples with poor learning effect are more easily selected:

;

Wherein the method comprises the steps of Is a small normal number to ensure a non-zero probability,The degree of prioritization is controlled.

As a preferred embodiment of the present invention, the pricing is based on risk assessment results, asset recovery prediction results, and optimization strategies, the step of generating pricing advice specifically includes:

Constructing a weighted scoring system based on the risk assessment result, the asset recovery rate prediction result and the optimization strategy;

Adjusting the risk based on the uncertainty;

introducing a market comparison method to perform preliminary calibration;

a pricing model is built based on an interpretable machine learning method, and pricing suggestions are generated.

In this embodiment, the individual recovery prediction of S4, the associated risk assessment of S3, and other key features are combined, a composite risk score or quality score is calculated for each asset or whole asset pack, and when calculating the base value, the value of the model preliminary assessment is corrected by using the uncertainty information quantified by S4 (such as the lower limit of the confidence interval or the expected value adjusted by the risk aversion coefficient), and referring to the historical trading prices or current quotes of similar bad asset packs on the market.

Constructing a multi-factor pricing model, one possible pricing formula can be expressed as:

;

Wherein B is the present value of the expected recoverable amount of a single or packaged asset based on model prediction, D is the discount rate determined by comprehensively considering the time value of funds, the specific risk premium of the asset, the liquidity premium and other factors, M is the adjustment coefficient of the basic value according to the current market supply and demand condition, the average premium or discount level of the comparable transaction, the macroscopic economic environment and other factors, U is an adjustment factor set based on the prediction uncertainty (such as the size of the cognitive uncertainty) quantified by S4, and the higher the uncertainty, the smaller the factor is possible to represent more conservative pricing.

The automated generation of detailed pricing analysis reports gives a clear recommended pricing interval, not just a single suggested price, but a reasonable price range (e.g., converted based on a 90% confidence interval of predicted recovery).

Detailing the key assumptions and core computational process upon which pricing is based, increases transparency.

And carrying out identification and quantitative analysis of main risk points, carrying out sensitivity analysis (SENSITIVITY ANALYSIS) on the influence of key parameter changes such as recovery rate, discount rate and the like on final pricing, and carrying out transverse comparison analysis on the pricing result of the asset pack and a market comparable transaction case to provide market positioning reference.

Explicit investment advice is given, e.g., whether the system recommends purchasing (or selling) the asset pack, and suggested trading strategies.

Providing negotiating policy references, providing buyers or sellers with possible negotiating spaces and key negotiating points, emphasizing major potential risk factors and their possible financial impact, and providing corresponding risk countermeasures or slow release suggestions based on analysis of seller offers and self-pricing results.

As a preferred embodiment of the present invention, the step of updating the pricing proposal based on the real-time raw data specifically comprises:

acquiring original data, and performing data drift detection and concept drift detection based on the original data;

generating performance degradation early warning based on the detection result;

and performing iterative optimization updating by adopting incremental learning and transfer learning based on performance degradation early warning.

In this embodiment, a comprehensive model performance monitoring system is established, accuracy of key prediction indexes (such as recovery rate and recovery time) is continuously monitored, common evaluation indexes include Mean Absolute Error (MAE), root Mean Square Error (RMSE), mean Absolute Percentage Error (MAPE) and the like, predicted pricing of a model is compared with actual achievement price, the size and direction of deviation are analyzed, market fitness of the model is evaluated, and actual recovery amount and recovery progress of an asset pack are continuously tracked and compared with predictions of the model.

And detecting data drift (DATA DRIFT), and monitoring whether the statistical distribution of the input features of the model changes significantly with time, such as borrowing person figures, macroscopic economic indexes and the like.

Concept Drift (Concept Drift) detection monitors whether a potential relationship between the input feature and the target variable has changed, e.g., the same borrower feature may correspond to different default probabilities at different times.

And the performance degradation early warning mechanism automatically triggers early warning when the performance index (such as accuracy rate and AUC) of the model continuously drops and is lower than a preset threshold value.

By adopting the incremental learning and online updating technology, new samples are effectively integrated, and the design model can effectively learn new modes and information while maintaining historical knowledge when new data samples (such as newly generated bad loans and finally recovered data of treated assets) come.

The online updating capability of the parameters can be used for small-amplitude and high-frequency adjustment of the model parameters based on the latest market feedback and small-batch new data for partial models (such as models based on gradient descent optimization).

Applying transfer learning (TRANSFER LEARNING), when faced with new scenarios or new product types where data is sparse, knowledge learned from related fields or historical data (e.g., feature extractors, pre-trained model parameters) can be transferred to new assessment tasks to speed up learning and improve performance.

Normalized model version management and iteration:

And the strict version control system records detailed information of each model update, including data version, code version, parameter configuration, performance and the like, so as to ensure the traceability and reproducibility of the model.

A/B test or Shadow Mode deployment is implemented, namely before the new model is formally put on line, the new model is compared and verified with the old model in parallel through the A/B test, or the performance of the new model is observed under the Shadow Mode but is not directly used for decision making, so that the stability and superiority of the new model are ensured.

And a quick rollback mechanism is established, so that once a new online model has serious performance problems or unpredictable errors, the model can be quickly switched back to the former stable and reliable old version model, and the service continuity is ensured.

Example application a commercial bank is to market a portfolio containing 5000 individual consumer bad loans with a total of approximately 10 billion primes, the borrowers involved being widely distributed in 20 different provinces and regions throughout the country. The intelligent assessment pricing system of the present invention is now used to comprehensively assess and price the asset pack.

The loan-related basic data is derived from the internal systems such as a core credit system, a risk management system and the like of the commercial bank and mainly comprises loan detailed information such as a unique number of each loan, an original loan amount, a loan date, a contract expiration date, a loan interest rate, repayment modes (such as equal cost, etc.), loan product types and the like, borrower basic information such as names, unique identification numbers, birth dates (used for calculating ages), sexes, occupation types, annual income levels (or reporting income), marital status, highest academic, detailed residence addresses and the like of borrowers, overdue information such as the date of first overdue occurrence, current continuous overdue days, historical accumulated overdue times, current undercost principal, current debit interest total amount, default interest amount and the like, and guaranty and mortgage information such as a guarantee mode (such as pure credit, third party guarantee, house mortgage, vehicle mortgage and the like) of the like of the loan, and guaranty basic information such as value information such as the type, value information and the position information of the like of the mortgage if the guaranty is included.

Meanwhile, valuable supplementary information for evaluation is acquired from a compliant external data source, wherein personal credit information is acquired by an authorized credit inquiry channel, liabilities of borrowers at other financial institutions, historical credit records (such as overdue credit cards, other default loan conditions and the like), inquiry records and the like are acquired, judicial complaint data is acquired by a public court judge document network, an execution information public network and other channels, whether the borrowers have complaint information, are recorded as executed persons, are listed as credit losing executed persons ("elder Lai") and the like, and a collection process record is recorded by arranging historical collection record texts of a collection mechanism inside a bank or outside the bank, and comprises the contact times, main contact modes, attitudes of the borrowers in a communication process, whether repayment promises, mentioned difficult reasons and the like.

Next, the collected raw data is subjected to strict preprocessing such as missing value processing, filling by a multiple interpolation method of borrower groups based on similar characteristics (such as occupation, age and region) for example, finding that borrower income information in 23 loan records is abnormal (such as older than 100 years or less than 18 years) by checking identity card number information, data normalization and transformation such as logarithmic transformation of amount type data (such as loan amount and income) to reduce deviation thereof, and normalization processing of [0,1] interval for comparative example, comparative example type data (such as historical overdue rate) is performed.

Finally, feature engineering is carried out, derivative variables with business significance are constructed, namely, a liability income ratio (Debt-to-Income Ratio, DTI) =total monthly payouts/total monthly incomes are calculated, a overdue severity index=current continuous overdue days/historical longest continuous overdue days are calculated, and a regional economic development index is constructed, wherein the regional economic development index is calculated by weighting based on the GDP growth rate of borrower living places or workplaces, average dominant incomes, town registration loss rates and other macroscopic economic indexes.

The method comprises the steps of extracting structured features, namely inputting feature dimensions, wherein the original data comprises 85 basic structured features and 42 derivative features generated in a feature engineering stage, adding 127 features in total, carrying out nonlinear transformation and dimension reduction on the structured features through a 3-layer fully connected neural network (MLP, the number of neurons is 128-64-32 respectively and a ReLU activation function is used), and outputting to obtain a 32-dimensional structured feature vector serving as a representative feature of the mode.

Text feature extraction, namely, preprocessing a collected furnacing record text, performing word segmentation (such as using Jieba words), removing stop words (such as 'words', 'words' and other words without practical meaning), constructing a dictionary (assuming that the final vocabulary quantity is 8567), using a BERT model (such as FinBERT) pre-trained in the financial field, performing field adaptive fine tuning on about 10 ten thousand historical furnacing record samples accumulated in the bank, extracting key information from the output of the BERT model or performing classification tasks, such as judging the repayment willingness level (such as strong/medium/weak) of a borrower, evaluating the difficulty (such as easy/medium/hard) of a contact borrower, identifying whether special conditions (such as no-business, serious diseases, no-coupling and the like) exist, outputting text feature vectors with 768 dimensions (or higher) of the original BERT output, and compressing the text feature vectors to 32 dimensions through a full-mode layer so as to be fused with other features.

The method comprises the steps of time sequence feature extraction, namely constructing a borrower repayment behavior sequence of the last 24 months, wherein each time point in the sequence comprises information such as the repayment amount, the actual repayment amount, whether overdue occurs, the overdue days and the like of the month, using a two-way long-short term memory network (Bi-LSTM network), setting the hidden layer dimension of the two-way long-short term memory network to be 128-dimensional for capturing long-term dependence and complex patterns in the repayment sequence, possibly identifying certain seasonal repayment modes by a model, for example, finding that the repayment rate of part of borrowers at the annual bottoms (such as after prize release) is about 15% of the promotion rule, and outputting a 32-dimensional time sequence feature vector after Bi-LSTM processing.

The cross-modal feature fusion comprises the steps of calculating a correlation matrix between every two of three modal (structuring, text and time sequence) feature vectors, for example, finding that 0.67 correlation coefficient exists between the text modal extracted feature and the structuring modal extracted feature, indicating that the text modal extracted feature and the structuring modal extracted feature have certain overlapping but are complementary, calculating the optimal weight of each modal feature in fusion through a designed cross-modal attention mechanism (gating fusion or attention weighting as described above), assuming that the obtained weights are structuring (0.45), text (0.35) and time sequence (0.20), carrying out weighted summation on the three 32-dimensional feature vectors according to the calculated attention weights, or finally carrying out fusion through a full-connection layer after splicing, and finally generating a 96-dimensional (or unified target dimension) fusion feature vector which characterizes each debtor more comprehensively.

The method comprises the steps of carefully identifying various association relations, namely discovering 892 pairs of guarantee relations in a property bag and among association parties by analyzing loan contracts and guarantee contracts, wherein 156 pairs are complex cross-guarantee or chain-guarantee, preliminarily identifying 523 potential family association units by analyzing contact information (such as emergency contact phones and relations), common residence addresses, similar surnames and other clues filled by borrowers, discovering that 218 borrowers possibly have colleague relations or enterprise groups belonging to the same actual controller by combining business data based on work unit information declared by the borrowers, and defining geographical proximity relations among borrowers with space distances within a range of 500 meters based on residence addresses or longitude and latitude information of the work unit addresses provided by the borrowers, thereby discovering 1876 pairs of geographically adjacent debtors.

The method comprises the steps of constructing an alien composition (Heterogeneous Graph) including 5000 main debtor nodes and 1235 independent vouchers (after de-duplication) identified through vouchers, wherein the number and the type of edges are summed to generate about 3509 edges representing different types of association, including 892 vouchers edges, 523 relatives edges, 218 colleagues/enterprises association edges and 1876 geographic proximity edges, analyzing connectivity of the composition, finding that the largest connected subgraph (Maximal Connected Component) in the composition contains 3821 nodes, and indicating that the potential risk conduction range among the asset package domestic debt operators is very wide.

The risk conduction path analysis is carried out by identifying 12 central nodes (Hubs) with higher risk conduction capability in the network through a graph algorithm (such as PageRank variety and K-core decomposition), wherein the average centrality of the nodes is larger than 15, successfully identifying 3 more obvious guarantee rings (Guarantee Circles), wherein the largest one guarantee ring relates to 37 liabilities, the total unreliability of all liabilities in the guarantee rings (potential risk opening) is up to 2100 ten thousand yuan, and the model discovery proves that 43 loans of staff of a large textile mill and the associated guarantee persons of the large textile mill are subjected to centralized default due to poor operation of the local textile mill, and the model discovery verifies the validity of the model identification associated risks.

The method comprises the steps of constructing a Graph Neural Network (GNN) model, inputting the model, wherein the initial characteristic of each node is 96-dimensional multi-mode fusion characteristic vector generated in the step 2, the characteristic of an edge can adopt one-time thermal coding of a relation type, the GNN framework comprises a GRAPHSAGE model (a inductive GNN capable of processing newly added nodes) of 3 layers, each node can aggregate the information of the adjacent nodes to update the representation of the node in each layer, and attention mechanism application introduces attention mechanisms (Graph Attention Network, GAT) when the adjacent information is aggregated, so that the edges of different types or the neighbors with different importance are given different weights when the information is transferred. For example, experiments have found that the weight of the vouching relationship is typically highest (e.g., the average weight found by modeling is 0.89).

Model training process data set partitioning, randomly partitioning nodes in the graph into training sets (e.g., 3500 nodes, 70%), verification sets (e.g., 750 nodes, 15%) and test sets (e.g., 750 nodes, 15%). The goal is to predict future probability of breach or recovery performance of the node, train the iteration, model trains 50 epochs (cycles) on the training set, evaluate performance (such as AUC, F1-score) on the verification set at the end of each epoch, adjust the super parameters or early stop according to the verification set performance, finally AUC (Area Under ROC Curve) obtained on the test set reaches 0.847, indicating that the model has good risk discrimination capability.

The model outputs a risk score between 0 and 100 for each borrower (node in the figure) in the asset package, wherein the higher the score is, the worse the associated risk and the individual risk are comprehensively evaluated, the greater the future recovery difficulty is, the risk scores are layered according to the risk scores, namely, a high risk group (score 80-100 min) is composed of 723 people in total, a medium risk group (score 60-80 min) is composed of 2145 people in total, 42.9% of the total, and a low risk group (score 0-60 min) is composed of 2132 people in total, and 42.6% of the total.

Constructing a Bayesian Neural Network (BNN) to predict the recovery rate:

The network architecture, the input layer receives 128-dimensional features (which may be a concatenation or some combination of node representation of GNN output and original multi-modal features), followed by two hidden layers (the number of neurons is 256 and 128, respectively, using tanh or ReLU activation functions), the output layer is designed to output 2 values, mean of predicted recovery and logarithmic variance of predicted recovery, to capture occasional uncertainty, the weight a priori distribution is set for all weight parameters in the network, e.g., 0 mean, 0.1 standard deviation (i.e., 0.01 variance), i.e. Training method training using variance inference, the loss function consists of the expected log likelihood term and KL divergence term (for regularization, weight set to 0.001).

The overall expected recovery of the property bag is 18.7% by weight-averaging the predicted recovery of all loans in the bag (per principal weight). Meanwhile, the 95% Confidence Interval (CI) of the recovery is [16.2% -21.5% ] by Monte Carlo sampling of BNN (for example, 100 predictions with different weights) and the recovery is predicted by the risk level, namely, the high risk group (from the division of the step 4) is predicted by the average expected recovery of 5.3%, the 95% CI is [3.1% -7.8% ], the medium risk group is 15.6% by the average expected recovery of 95% CI is [13.2% -18.3% ], and the low risk group is 31.2% by the average expected recovery of 95% CI is [28.5% -34.1% ].

Occasional uncertainty, i.e., noise and randomness inherent to the data, is estimated to account for 68% of the total predicted uncertainty. The model mainly derives from the factors such as unpredictability of future repayment capability, fluctuation of mortgage value and the like of borrowers, and the cognition uncertainty is that the model accounts for 32% of the total uncertainty due to uncertainty of own parameters (usually derived from insufficient training data or imperfect model structure). In this case, the model's predictive confidence in these populations may be poor, mainly due to the relatively small sample size of certain specific areas or types of borrowers.

State space, the environment describing the current decision time. Including the portfolio of assets currently selected for holding, the remaining available budget, and some key market risk indicators (e.g., simulated macro economic index changes), action space: actions that agents (agents) can take. Simplified here to select which subset of assets to purchase from the asset pack. To simplify the problem, the original 5000 assets may be pre-clustered or screened into 100 candidate sub-asset packs (or portfolios). Actions are selected as one or more of the candidate combinations, and a reward function is used to evaluate the quality of an action. Designed as. Coefficients 0.5 and 0.1 are super parameters.

Model training process RL algorithm selection, a near-end policy optimization (PPO) algorithm is used. The Actor network and the Critic network are designed to comprise 3 full connection layers, and training iteration is that 10000 episode training is performed. Each episode simulates an investment decision period, for example, 6 months in holding period, during which changes in income and risk are observed, and an experience playback (Experience Replay) is provided with an experience playback buffer zone of 10000 for storing (state, action, rewards, next state) transition samples, randomly sampling from which to train, breaking data correlation.

The recommended property is that after training by the RL agency, about 3200 loans in the property package are recommended to be purchased under given budget and risk preference, the recommended property package accounts for 64% of the total loan amount, the recommended performance index of the recommended combination is that the expected annual gain rate is 22.3% -the risk value VaR (95% confidence level, 1 year): -8.7% (i.e. the 5% probability annual loss exceeds 8.7%) -the Charpy ratio (assuming no risk is 3%): (22.3% -3%)/Portfolio_Std_Dev-is calculated to obtain the combined standard deviation) -the region dispersity is 0.087 by calculating the principal distribution of the selected property in different provinces, indicating that the region distribution is highly dispersed and the risk is lower.

Calculating the base value (Present Value of Expected Recoveries) of the asset pack, namely, the original principal sum of the asset pack is 10 hundred million yuan, the expected recoverable total amount is 18.7 percent of the expected recoverable 1.87 hundred million yuan according to the overall expected recovery of the step 5, the average recovery period is 24 months (2 years) of the expected average recovery period by comprehensively considering the disposal speed and the recovery time distribution, and the applicable discount rate is calculated based on the current market Risk-free rate (such as national debt yield) plus the Risk Premium (Risk Premium) and the liquidity Premium for the bad asset, assuming that the comprehensive discount rate is 12 percent (annual), and the recovery value is calculated by the present value:。

Data is collected by collecting transaction cases of 20 similar packages of personal consumer bad assets occurring in the last 6 months of the market, and by analyzing comparable transactions by calculating the average principal discount rate (price/principal) of those comparable transactions. Assuming an average discount rate of 82.5% (i.e., a cost of about 17.5%) and adjusting the characteristics of the present package of assets by giving a +2.3% adjustment to account for certain characteristics of the present package of assets (e.g., borrower territory distribution is better, average individual amounts are smaller and thus more diffuse) that may be better than the average market level, the adjusted market reference discount rate: (where +2.3% of the original text is understood to be the reduction of the discount rate, i.e. the increase of the value, it should be Or is alsoIt should be noted that the term "positive adjustment" is understood herein to mean a decrease in discount, i.e., an increase in the price ratio. If the intention is that the discount rate itself is added by 2.3 percentage points, then it is. The rate of discount is here reversed by "80.2%", more like (100-80.2)% = 19.8% of the price ratio principal. Assuming that the average price of the comparable transaction is 17.5% of principal, the adjusted principal is 19.8%, the adjustment coefficient is. The original text "discount rate after adjustment: 80.2%" implies a price of about 19.8% of principal. Here, the final result is explained. ) Assuming that the market reference's own asset package price is approximately principal after adjustment. The market method valuation is。

The results of the basic value method (DCF variant) and the market comparison method are combined, and uncertainty is considered, and the quotation interval recommended by the system is 1.42 hundred million yuan to 1.56 hundred million yuan. This interval may be based on the confidence interval of the underlying value calculation and adjusted with reference to the market result, the system suggests a median bid of 1.49 billion yuan (consistent with the underlying value calculation), the rate of discount of this bid relative to principal:。

Recovery sensitivity the estimated value of the asset pack increases approximately ± per 1 percent improvement (e.g., from 18.7% to 19.7%) in model predicted overall recovery )/(Here, 790 tens of thousands may be slightly simplified or different assumptions), discount rate sensitivity, the estimated value of the asset pack increases approximately (reckoned on a 1.49 billion basis) for every 1 percent reduction in discount rate used (e.g., from 12% to 11%),Namely 270 ten thousand yuan, the original 124 ten thousand calculation modes need to be checked, and the calculation modes may be derivatives of the discount rate), and the sensitivity of the recovery period, namely, the average recovery period is shortened by 1 month, the total recovery amount is assumed to be unchanged, and the evaluation value is increased approximately due to the improvement of the time value (the specific calculation is complex, the cash flow distribution is depended, and 52 ten thousand are taken as an example here).

The system automatically generates a detailed evaluation report of about 50 pages, and the main content structure and the core key points are as follows:

Summary of the basic case of the package of assets including 5000 individual consumer loans, 10 billions of the total original principal, core evaluation conclusion, 1.49 billions of primordial notes on the recommended purchase price (or fair value), 1.42-1.56 billions of reasonable quotation intervals, expected return on investment analysis, expected internal yield (IRR) of 22.3% if purchased at the recommended price, expected average investment recovery period of 24 months, and major risk point prompt, namely, briefly listing major risks faced by the package of assets, such as specific security circle risks, risks excessively concentrated in part borrower areas, and the like.

The detailed analysis section is used for carrying out statistical description and visual display on key attributes such as age structure, occupation distribution, income level, regional distribution and the like of borrowers, loan overdue feature analysis, loan overdue time distribution (such as M1, M2, M3+ overdue duty ratio), overdue amount distribution feature and the like, association network and risk conduction analysis, namely attaching a key debtor association network diagram (visualization) to indicate main risk conduction paths and high risk communities, and carrying out recovery prospect prediction details to detail the expected recovery rate, recovery amount and corresponding confidence intervals according to risk layering (high, medium and low risk groups), and analyzing the sources of uncertainty.

The risk prompt and early warning are used for prompting that if the macroscopic economic situation is obviously descending in the future (such as GDP speed-up is lower than the expected 2 percentage points), the overall recovery rate is possibly reduced by 5-8%, the important attention is paid to a risk object, namely, 3 large-scale guaranty rings which need important attention are clearly pointed out, the potential linkage violation risks are prompted, and the regional concentration risk is that the principal rate of loan of a single province in a property package reaches 23%, a certain regional concentration risk exists, and the specific economy or policy change of the region needs to be focused.

Investment and disposal advice, which is to explicitly advise buyers to bid or negotiate at a price of not more than 1.52 yuan, disposal policy initiatives, which advises that after successful acquisition, quick and flexible disposal means (such as liability reorganization and small amount quick collection) are preferably adopted for low-risk group assets, and that the assets can be expected to recover about 30% of the expected amount in 6 months, and risk monitoring mechanism advice, which advises that special monitoring mechanism is established for the identified risk of the security circle and closely tracks the operating condition of the core enterprise in the circle.

Assuming that the intelligent assessment system of the present invention has been on-line for 6 months at an asset management company, models are now reviewed, updated and optimized based on actual asset disposition data and market feedback accumulated during that period.

Model historical performance evaluation the overall expected recovery of a batch of asset packs predicted at the initial stage of system on-line was 18.7%, the actual recovery of the batch of asset packs at stages (recovered amount/total principal) was 17.2% by statistics over 6 months of actual treatment (note: this is only stage recovery, not completely treated, may be different from the final recovery, but could be used as an early tracking indicator).

The predicted deviation under simple comparison: . However, if time factors are considered, such as an expected recovery of 18.7% for 2 years, for example 40% of the expected total recovery is recovered for half a year, and in practice only 35% is completed, the deviation will be greater. "prediction bias: +8.7%" herein may mean that the predicted value is 8.7% higher than the actual value (under some reference). ( Assume that the original "predicted deviation: +8.7%" means that the predicted value is 18.7% 8.7% higher than some comparable reference point actually observed, or 8.7% higher. It is understood here that the absolute deviation, i.e. the prediction, should be 18.7% -8.7% = 10%. This requires a clearer definition. We continue as original text values. )

The deviation causes are deeply analyzed, namely regional burst factors, and through data drilling analysis, the fact that the economic condition of a certain area occupying a larger area in the asset package is unexpected deteriorated due to sudden events (such as important industrial policy adjustment, natural disasters and the like), so that the actual recovery rate of the assets in the area is only 60% of the original predicted value.

Global external impact-for example, high infectious respiratory virus transmission has a significant impact on normal apheresis (such as upper door apheresis, judicial procedures), resulting in an average delay of overall recovery schedule of about 2 months compared to the original program.

Introducing new influence factors, and adding the virus transmission influence factors into a model feature library, wherein the factors can be quantified, and for example, the factors are constructed through data of the influence degree (such as dining and travel are greatly influenced) of the virus transmission of the industries.

Adding 500 newly generated partially or completely treated asset samples (including the actual recovery data) during the period of 6 months into a training set, performing incremental learning or retraining on the existing model to learn the latest market mode, adjusting the weight of key features, and properly improving the weight of a regional economic advance index (such as PMI and regional financial income acceleration) in the model according to the latest data analysis, for example, improving the contribution weight of the regional economic advance index to a prediction result from the original 0.15 to 0.22.

And (3) verifying the effect of the updated model, namely evaluating the model on a brand-new test set which does not participate in the updating training, wherein the prediction error (such as MAPE or RMSE) of the model on the recovery rate is obviously reduced to 4.2% from 8.7% before, and particularly for the region which is seriously affected by the virus transmission, the prediction accuracy of the updated model is improved by about 35%, and the recovery condition of the current special period can be reflected.

A large Asset Management Company (AMC) plans to acquire a batch of bad assets from 5 different commercial banks, involving a total of 12 independent asset packs, requiring fast and efficient batch evaluation and combinatorial optimization across packs.

The batch processing and efficiency optimization comprises the steps of starting a distributed computing architecture, distributing the evaluation tasks of 12 asset packs to a plurality of computing nodes for parallel processing, greatly shortening the overall evaluation time, sharing and multiplexing a feature extraction module, designing a common feature extraction logic (such as BERT coding of text and LSTM processing of time sequence) into a service module for sharing and calling of all asset pack evaluation tasks, avoiding repeated calculation and resource waste, constructing a global association network of crossing asset packs, and when a liability person association map is constructed, not only considering the association inside a single asset pack, but also trying to identify whether common liabilities, guarantors or association parties exist among different asset packs, and constructing a larger-range association network covering all 12 asset packs.

The method comprises the steps of analyzing batch processing results, namely, total processing time, carrying out detailed evaluation and preliminary pricing of all 12 asset packages through parallelization and module multiplexing, wherein the total time is only 4.5 hours, if a traditional serial processing mode is adopted, the requirement of about 18 hours is expected, cross-package association is found, the association relation of debtors crossing different original asset packages is successfully identified 127 in a built global association network, a large cross-package risk community is identified, and further analysis finds that 2 large guarantee networks related to a plurality of original asset packages exist, and once risks of the networks explode, the multiple asset combinations purchased by AMC can be affected, so that special attention is needed.

Optimizing decisions across portfolios, optimizing targets, optimizing combinations, applying reinforcement learning or other portfolio optimization algorithms (such as mean-variance models, risk flat models, etc.) described in step 6 under set total investment budget limits (e.g., no more than 50 billion yuan), comprehensively considering expected profitability, risk levels (such as CVaR), and correlations among them, of each portfolio, and outputting recommended combinations, the model eventually recommending 7 portfolios for purchase, the overall expected annual profitability of the recommended combination being expected to reach 24.6%, which is significantly higher than the average expected profitability (assuming 19.8%) of simply purchasing all 12 portfolios or randomly selected.

The embodiment of the invention also provides a system for intelligently evaluating and pricing the lending bad assets based on the graphic neural network, which comprises the following steps:

the multi-source data acquisition module is used for acquiring original data in real time, wherein the original data comprises bank internal data, bank external data and market data;

The multi-mode feature fusion module is used for extracting multi-mode feature data based on the original data and fusing the multi-mode feature data;

The debtor association graph construction and risk assessment module is used for building a debtor association network graph based on the debtor information and carrying out risk assessment based on the debtor association network graph and the fused multi-mode feature data;

The time sequence prediction and uncertainty quantification module is used for predicting the asset recovery rate and quantifying the uncertainty based on a Bayesian deep learning method and a risk conduction analysis result;

The asset pack combination optimization module is used for optimizing the asset pack combination based on the Markov decision process model to generate an optimization strategy;

and the intelligent pricing decision module is used for pricing based on the risk assessment result, the asset recovery rate prediction result and the optimization strategy to generate pricing advice.

The multi-mode feature fusion module adopts an innovative cross-mode attention mechanism, can deeply and effectively fuse features of different modes such as structured data (such as form data), text data (such as a collection record and a legal document), time sequence data (such as a repayment sequence) and the like, and constructs a unified high-dimensional feature representation with rich information, so that the representation of each debtor is more comprehensively depicted;

The debtor association map construction and risk assessment module is used for constructing a dynamic debtor association network map by utilizing a map database technology based on multi-dimensional information such as a guarantee relationship, a relative relationship, a common borrowing relationship, enterprise association, similar geographic positions and the like among borrowers, so as to explicitly represent and analyze complex associations among the debtors and identify potential risk conduction paths and risk communities;

The special heterogeneous graph neural network architecture is designed, information transmission and feature aggregation between nodes are carried out on the constructed liability person association graph, so that risk characterization of each liability person node in the association network is learned, and systematic risks and conduction effects of the risks are effectively identified and quantified;

The time sequence prediction and uncertainty quantization module adopts a Bayesian deep learning method to carry out probabilistic prediction on the future recovery rate, recovery amount or recovery time of the bad asset, not only provides a predicted point estimation value, but also can provide uncertainty quantization of a prediction result, and comprises a confidence interval for giving the prediction value and complete posterior risk distribution;

The asset pack combination optimization module fully considers the relativity and the risk and income characteristics among different assets in the asset pack based on the modern investment portfolio theory, adopts a reinforcement learning optimization algorithm, dynamically optimizes the screening and combination strategy of the asset pack, and aims at maximizing the expected income after risk adjustment;

The intelligent pricing decision module synthesizes the output results (such as individual risk score, prediction recovery rate, associated risk assessment, combined optimization suggestion and the like) of the modules, generates final pricing suggestion by adopting an interpretable machine learning method (such as SHAP, LIME assisted decision tree or linear model), and can provide detailed pricing basis, key influence factor analysis and potential risk prompt;

The model self-adaptive updating module is further arranged, and can continuously and automatically optimize and iterate each model in the system according to the latest market feedback data (such as actual recycling conditions and newly generated bad asset data) and external environment changes by deploying the online learning and transfer learning technology so as to maintain the timeliness and the prediction accuracy of the model.

The foregoing description of the preferred embodiments of the invention is not intended to be limiting, but rather is intended to cover all modifications, equivalents, and alternatives falling within the spirit and principles of the invention.

Claims (10)

1. A method for intelligently evaluating and pricing individual bad assets based on a graph neural network, which is characterized by comprising the following steps:

Acquiring original data in real time, wherein the original data comprises bank internal data, bank external data and market data;

Extracting multi-mode feature data based on the original data, and fusing the multi-mode feature data;

establishing a debtor association network map based on the debtor information, and performing risk assessment based on the debtor association network map and the fused multi-mode feature data;

Predicting asset recovery rate and quantifying uncertainty based on Bayesian deep learning method and risk conduction analysis result;

optimizing the asset pack combination based on the Markov decision process model to generate an optimization strategy;

pricing is performed based on the risk assessment results, asset recovery prediction results, and optimization strategies to generate pricing advice.

2. The method for intelligently assessing and pricing individual lending undesirable assets based on a graph neural network as recited in claim 1, wherein the step of extracting multi-modal feature data based on raw data specifically comprises:

Cleaning and standardizing the original data;

constructing basic features and derivative features and calculating;

and carrying out extraction analysis on structural features, text features and time sequence features of the cleaned and standardized data.

3. The method for intelligently assessing and pricing individual lending undesirable assets based on a graph neural network according to claim 2, wherein the step of fusing the multimodal feature data specifically comprises:

Calculating a correlation matrix among the structural features, the text features and the time sequence features;

Assigning fusion weights to the structured features, text features, and timing features based on an attention mechanism;

and fusing the multi-modal feature data based on the fusion weight to generate a multi-modal feature representation vector.

4. The intelligent evaluation and pricing method for individual bad assets based on graphic neural network according to claim 1, wherein the step of establishing a debtor association network map based on the debtor information, and performing risk evaluation based on the debtor association network map and the fused multimodal feature data specifically comprises:

establishing a debtor association network map based on the debtor information;

constructing a dynamic heterogeneous graph network model based on the liability person associated network graph;

determining a risk conduction path and a risk conduction rule;

and performing risk assessment based on the dynamic heterograph network model.

5. The intelligent evaluation and pricing method for individual bad assets based on a graphic neural network according to claim 1, wherein the step of generating an optimization strategy specifically comprises:

defining an objective function consisting of maximizing the expected combined benefit and minimizing the combined risk;

Constraint conditions are determined, including budget constraints, concentration constraints, and flowability constraints.

And (3) constructing a Markov decision process model, optimizing the asset pack combination based on the Markov decision process model, and generating an optimization strategy.

6. The intelligent evaluation and pricing method for individual bad assets based on a graphic neural network according to claim 1, wherein the step of generating pricing advice specifically comprises:

Constructing a weighted scoring system based on the risk assessment result, the asset recovery rate prediction result and the optimization strategy;

Adjusting the risk based on the uncertainty;

introducing a market comparison method to perform preliminary calibration;

a pricing model is built based on an interpretable machine learning method, and pricing suggestions are generated.

7. The intelligent evaluation pricing method for individual lending undesirable assets based on a graph neural network of claim 6, wherein the pricing model is expressed as:

Wherein B is the present value of the expected recoverable amount of a single or packaged asset based on model prediction, D is the discount rate determined by comprehensively considering the time value of the funds, the specific risk premium of the asset and the liquidity premium factor, and M is the adjustment coefficient of the basic value according to the current market supply and demand condition, the average premium or discount level of the comparable transaction and the macro economic environment factor. U is an adjustment factor set based on the quantized prediction uncertainty.

8. The intelligent assessment pricing method for individual lending undesirable assets based on the graphic neural network of claim 1, further comprising updating the pricing advice based on real-time raw data.

9. The intelligent evaluation pricing method for individual lending undesirable assets based on the graphic neural network of claim 8, wherein the updating the pricing advice based on the real-time raw data step comprises:

acquiring original data, and performing data drift detection and concept drift detection based on the original data;

generating performance degradation early warning based on the detection result;

and performing iterative optimization updating by adopting incremental learning and transfer learning based on performance degradation early warning.

10. A graphic neural network-based individual malfunctioned asset intelligent assessment pricing system for implementing the graphic neural network-based individual malfunctioned asset intelligent assessment pricing method of any of claims 1-9, the system comprising:

the multi-source data acquisition module is used for acquiring original data in real time, wherein the original data comprises bank internal data, bank external data and market data;

The multi-mode feature fusion module is used for extracting multi-mode feature data based on the original data and fusing the multi-mode feature data;

The debtor association graph construction and risk assessment module is used for building a debtor association network graph based on the debtor information and carrying out risk assessment based on the debtor association network graph and the fused multi-mode feature data;

The time sequence prediction and uncertainty quantification module is used for predicting the asset recovery rate and quantifying the uncertainty based on a Bayesian deep learning method and a risk conduction analysis result;

The asset pack combination optimization module is used for optimizing the asset pack combination based on the Markov decision process model to generate an optimization strategy;

and the intelligent pricing decision module is used for pricing based on the risk assessment result, the asset recovery rate prediction result and the optimization strategy to generate pricing advice.