An intelligent portfolio optimization pricing method for non-performing assets

Abstract

The present application relates to the technical field of financial risk control, and particularly relates to an intelligent portfolio optimization pricing method for non-performing asset packages, which comprises the following steps: extracting structured features of non-performing assets; performing interactive coding on the structured features through an interactive encoder to obtain context-aware embedding of assets in a target non-performing asset package and package-level aggregated representation; defining an objective function and constraint conditions of portfolio optimization, generating a candidate scheme set satisfying the constraint conditions and having a target function value higher than a preset screening threshold through a pointer network for multiple random sampling of the package-level aggregated representation; using a deep set encoder network to output a mixed Gaussian distribution of future recovery amount of a corresponding candidate portfolio in the candidate scheme set; calculating a reliability index and a comprehensive priority score of each candidate portfolio based on the mixed Gaussian distribution; and outputting a recommended scheme set based on the reliability index and the comprehensive priority score.

Description

Technical Field

[0001] This invention relates to the field of financial risk control technology, and in particular to an intelligent portfolio optimization pricing method for non-performing asset portfolios. Background Technology

[0002] Non-performing assets mainly refer to the claims held by banks and other financial institutions that cannot be recovered in full and on time, along with their associated collateral and guarantee rights. To quickly clear risks and release capital, financial institutions typically package dozens to hundreds of non-performing claims into "non-performing asset packages" for bulk transfer. The pricing and composition of these asset packages are the core challenges of these transactions.

[0003] Traditional valuation and portfolio construction of non-performing asset packages mainly rely on human experience: analysts review debt files one by one, estimate the expected recovery amount of a single debt based on factors such as collateral valuation, litigation progress, and guarantor's debt repayment ability, and then price it by simple summation and discount. This method has the following prominent technical problems: (1) It cannot model the complex non-linear risk relationship between assets. There are often multiple debts in an asset package that share the same guarantor, point to the same collateral, or belong to the same regional industry, resulting in systemic risk exposure. The traditional linear summation method cannot capture the risk resonance effect at all, leading to a serious underestimation of the tail risk of the portfolio. (2) The handling of missing information and uncertainty is too crude. Non-performing asset data is highly heterogeneous and seriously incomplete. Traditional methods usually fill missing values ​​with the mean or median, discarding the important risk signal of "missing information" and failing to explicitly retain and transmit uncertainty to the confidence interval of the final valuation. (3) The portfolio space search is inefficient. When the pool of candidate assets reaches hundreds, the number of possible combinations is exponential, which is impossible to exhaustively count manually. It is easy to miss hidden high-value combinations and it is also impossible to systematically find the Pareto optimal boundary between expected return and tail risk. (4) The pricing results lack confidence quantification. Traditional methods output a single expected value, which cannot inform decision-makers of the reliability of the estimate, nor can it answer tail risk inquiries, making investment decisions lack transparency and auditability in terms of risk quantification. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent portfolio optimization pricing method for distressed asset portfolios to solve the above-mentioned technical problems.

[0005] The technical solution proposed in this invention is: an intelligent portfolio optimization pricing method for non-performing asset portfolios, the method comprising:

[0006] Acquire multi-source heterogeneous data of the target non-performing asset portfolio, and extract the structured features of the non-performing assets from the multi-source heterogeneous data;

[0007] An interactive encoder is constructed to interactively encode structured features, thereby obtaining context-aware embeddings and package-level aggregated representations of assets within the target non-performing asset package; the interactive encoder is a multi-head self-attention network containing explicit relational biases.

[0008] Define the objective function and constraints for combinatorial optimization, and generate a set of candidate solutions that satisfy the constraints and whose objective function values ​​are higher than a preset screening threshold by multiple random samplings of the packet-level aggregate representation through a pointer network;

[0009] For candidate combinations in the candidate scheme set, a deep ensemble encoder network is used to output a Gaussian mixture distribution of the future recovery amount of the corresponding candidate combination;

[0010] Based on a Gaussian mixture distribution, the reliability index and comprehensive priority score of each candidate combination are calculated; a set of recommended solutions is output based on the reliability index and comprehensive priority score for manual review.

[0011] Preferred options also include:

[0012] A two-dimensional table of non-performing assets is constructed using structured features; the structured features include legal document fields, collateral file fields, and ledger data fields; the rows of the two-dimensional table represent assets, and the columns represent a field.

[0013] For numerical structured features, explicit missing encoding and masking are performed, and the mask indicating whether the data actually exists is treated as an independent feature.

[0014] One-hot encoding is performed on the categorical structured features;

[0015] The missing code, mask, and one-hot code are concatenated to form the input feature vector;

[0016] Using a conditional variational autoencoder, the input feature vector of each asset is encoded into a probability embedding consisting of a mean vector and a standard deviation vector, wherein the standard deviation vector is used to explicitly quantify the information uncertainty of the asset itself.

[0017] The decoder outputs the Gaussian distribution parameters and recovery probabilities of numerical structured features and the class probabilities of categorical structured features through the conditional variational autoencoder.

[0018] Preferably, the step of interactively encoding the structured features through an interactive encoder to obtain the context-aware embedding and package-level aggregated representation of each non-performing asset includes:

[0019] Based on probabilistic embedding, explicit relationship edges between two assets are extracted. These explicit relationship edges include shared guarantor relationships, shared collateral relationships, corporate affiliations, and regional-industry homogeneous relationships. A learnable edge embedding is assigned to each type of explicit relationship edge.

[0020] By using each attention head of the multi-head self-attention network, the edge embedding is linearly transformed and used as a bias term, which is added to the scaled dot product attention score to obtain the original score of each attention head, so that the attention weight depends on both semantic similarity and explicit relation.

[0021] Multiple encoder layers are stacked, each layer including a multi-head attention network, residual connections, layer normalization and a per-asset feedforward network, and dropout regularization is applied at the output of the self-attention network and the output of the feedforward network; the output of the last encoder layer is the context-aware embedding of each bad asset.

[0022] The encoder has no more than 3 layers and no more than 12 attention heads;

[0023] The package-level aggregation is obtained by linearly projecting the concatenation result of average pooling and learnable attention pooling.

[0024] Preferably, the definition of the objective function and constraints for combinatorial optimization includes:

[0025] The objective function is defined as the difference between the risk-adjusted expected internal rate of return and the tail risk penalty and complexity penalty;

[0026] Define constraints, including hard constraints: the total acquisition cost is less than a preset cost threshold; and soft constraints: the proportion of collateral types is less than a preset percentage threshold.

[0027] The expected internal rate of return is approximately expressed by a power relationship between the expected recovery amount, the total acquisition cost, and the average recovery period;

[0028] The tail risk penalty is quantified based on the difference between the conditional value at risk and the expected return; the complexity penalty includes penalties for the total number of assets exceeding a threshold and for the excessive proportion of small assets; the package-level aggregate representation is sorted according to the objective function value; wherein, the penalty coefficient weights in the tail risk penalty and the complexity penalty are calibrated through time-layered cross-validation to maximize the rank correlation coefficient between the objective function ranking and the historical true internal rate of return ranking.

[0029] Preferably, the step of using a pointer network as a policy network to generate a set of candidate solutions that satisfy the constraints and whose objective function values ​​are higher than a preset screening threshold by repeatedly randomly sampling the packet-level aggregate representation through the policy network includes:

[0030] The policy network uses the package-level aggregation representation of the current asset as the query and global interactive embedding as the key to calculate and obtain the asset. Relevance score to the current packet state ;in, The current package-level aggregation of assets represents, Embedded for global interaction; It is the bias vector; This is the projection vector, used to map the correlation to a scalar; For activation functions;

[0031] The global interactive embedding is obtained by inputting all assets in the distressed asset portfolio into the multi-head self-attention network;

[0032] The probability of an asset being selected from the target non-performing asset portfolio is calculated based on the correlation score and temperature coefficient; the formula for calculating the probability is as follows: ,in, For assets The probability of being selected A set of indexes for selectable assets; The temperature coefficient is used; the selectable assets are those that meet the hard constraints.

[0033] Calculate the objective function value of assets whose probability exceeds a preset selection probability threshold, and sort them in descending order according to the size of the objective function value. Assets with objective function values ​​higher than a preset screening threshold constitute a candidate portfolio of non-performing assets.

[0034] K different temperature coefficients are selected, and their probabilities are calculated for each. Based on the probabilities, the corresponding objective function values ​​are calculated, thereby obtaining K different candidate combinations, which constitute a candidate scheme set; the temperature coefficients range from 1 to 2.

[0035] The package status includes the currently selected assets, the remaining available budget, and a package-level aggregate representation of the selected assets; actions include selecting an asset from those that meet hard constraints to add to the asset package and a learnable stop action token.

[0036] Preferably, for each candidate combination in the candidate scheme set, a deep ensemble encoder network is used to output a Gaussian mixture distribution of the future recovery amount for the corresponding candidate combination, including:

[0037] Obtain context-aware embeddings and package-level aggregated representations of assets in candidate portfolios;

[0038] The context-aware embeddings of all assets in the candidate combination are first nonlinearly transformed and summed by a deep set encoder to obtain a permutation-invariant representation.

[0039] The obtained permutation-invariant representation is concatenated with the package-level scalar condition vector and then output by a global decoding network as a mixture of Gaussian distributions of the future recovery amount of the candidate combination; the package-level scalar condition vector includes the total cost of the candidate combination assets, the average recovery period, and the number of assets.

[0040] The deep ensemble encoder includes multiple multilayer perceptrons, each multilayer perceptron including three fully connected layers and a GELU activation function; the global decoding network includes four fully connected layers and a Sigmoid activation function.

[0041] Monte Carlo dropout is enabled after the fully connected layers of the deep ensemble encoder and the fully connected layers of the global decoder network, with a dropout rate of 0.05. The model uncertainty measure is obtained by performing independent forward propagation to obtain the posterior prediction distribution; the variance of the expected value of the posterior prediction distribution is used to calculate the model uncertainty measure.

[0042] Preferably, the calculation of the reliability index and comprehensive priority score for each candidate combination based on the Gaussian mixture distribution of the future recovery amounts of the candidate combinations includes:

[0043] The reliability index is obtained by normalizing and weighting the entropy of the posterior prediction distribution and the model uncertainty measure, and then taking the reciprocal.

[0044] The overall priority score is calculated by subtracting the penalty for the difference between the expected internal rate of return and the conditional value at risk (IVR) and the internal rate of return, and adding the reliability bonus.

[0045] Preferably, the set of recommended solutions output based on reliability index and comprehensive priority score includes:

[0046] The candidate combinations are sorted in descending order of their comprehensive priority scores to obtain the top... One recommended option;

[0047] Jaccard similarity of asset selection vectors for candidate combinations within the recommended scheme; the asset selection vectors ;in, For the first The number of assets in each candidate portfolio. Assets The selected value, Assets Being chosen Assets Not selected;

[0048] Before screening using diversity filtering algorithms One recommended scheme is to retain those with a similarity score below the similarity threshold in Jaccard. The recommended solution set consists of several solutions; at the same time, at least one representative solution that meets the preset risk attribute dimension is forcibly retained; the similarity threshold ranges from 0.5 to 0.9.

[0049] An electronic device for executing the aforementioned intelligent portfolio optimization pricing method for distressed asset portfolios.

[0050] A computer-readable storage medium storing a computer program that is executed by a processor to implement the aforementioned intelligent portfolio optimization pricing method for distressed asset portfolios.

[0051] The beneficial effects of this invention are:

[0052] 1. This invention transforms the uneven data quality from a “defect” into a quantifiable risk signal through the probabilistic embedding and explicit missing data encoding mechanism of the conditional variational autoencoder. This enables decision-makers to clearly understand which schemes have low valuation credibility due to weak basic data, and significantly reduces prediction calibration error in data missing scenarios.

[0053] 2. This invention encodes explicit relationships between assets into learnable relational edge embeddings, which are then injected into the attention score calculation of a multi-head self-attention network as a bias term. This makes the risk transmission between assets not only based on semantic similarity, but also directly dependent on explicit strong associations in business logic. This solves the problem that traditional pure data-driven methods cannot capture tail dependencies in sparse interaction scenarios, making the prediction of portfolio tail risk closer to the true loss distribution and avoiding overly optimistic estimates of portfolio diversification effects.

[0054] 3. This invention employs a mixture Gaussian distribution to output the combined recovery amount using a deep ensemble encoder. This integrates data uncertainty, model uncertainty, and strategy sampling stability into a unified reliability index, which, along with expected internal rate of return, conditional value at risk tail penalty, and soft constraint penalty, constitutes a comprehensive priority score. This upgrades investment decisions from a single expected value comparison to a three-dimensional transparent trade-off, with each figure traceable to the original data and model assumptions, meeting compliance audit requirements. Attached Figure Description

[0055] Figure 1 This is a flowchart of an intelligent portfolio optimization pricing method for non-performing asset portfolios according to the present invention. Detailed Implementation

[0056] The following description is intended to disclose the present invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious modifications will occur to those skilled in the art. The basic principles of the invention defined in the following description can be applied to other embodiments, modifications, improvements, equivalents, and other technical solutions that do not depart from the spirit and scope of the invention.

[0057] It is understood that the term "a" should be understood as "at least one" or "one or more," that is, in one embodiment, the number of an element can be one, while in another embodiment, the number of the element can be multiple, and the term "a" should not be understood as a limitation on the number.

[0058] refer to Figure 1 The technical solution provided by this invention is: an intelligent portfolio optimization pricing method for non-performing asset portfolios, the method comprising:

[0059] Step 1: Obtain multi-source heterogeneous data of the target non-performing asset portfolio, and extract the structured features of the non-performing assets from the multi-source heterogeneous data. This specifically includes the following steps:

[0060] Step 1.1: Obtain multi-source heterogeneous data of the target non-performing asset portfolio, including legal document data (loan contracts, judgments, rulings), collateral file data (property ownership certificates, land use certificates, etc.), and ledger data (Excel spreadsheets, business registration information, etc.).

[0061] Using Optical Character Recognition (OCR) and layout analysis models, fields such as creditor, debtor, guarantor, bond principal, interest calculation method, litigation status (sued, not sued, terminated), and mortgage certificate number are extracted from legal document data to form legal document fields. Similarly, fields such as address, area, use (residential, commercial, industrial), appraised value, and whether it is the first seizure are extracted from mortgage file data to form mortgage file fields. Fields such as industry classification, five-level classification, current balance, business operation status, and judicial auction transaction amount are extracted from ledger data to form ledger data fields.

[0062] Furthermore, the OCR engine outputs a confidence score for each recognition result. Fields with a confidence score below the threshold are marked as "low confidence" while retaining the original text fragment coordinates to prevent overfitting. For example, if the OCR engine recognizes the "500" in the principal amount of 5 million yuan as having a confidence score below the threshold of 0.95, it will mark the field "500" as "low confidence" and retain the image coordinates of that fragment in the original text for later cross-validation.

[0063] Step 1.2: Construct a two-dimensional table of non-performing assets using structured features, including legal document fields, collateral file fields, and ledger data fields; rows in the two-dimensional table represent assets, and columns represent a field.

[0064] Step 1.3: For numerical structured features, perform explicit missing data encoding and masking, using the mask indicating whether the data actually exists as an independent feature; for categorical structured features, perform one-hot encoding; concatenate the missing data encoding, mask, and one-hot encoding to form the input feature vector. Specifically:

[0065] For all numerical fields (such as principal, appraised value, and area), a dual-channel representation is used: the numerical channel is filled with the original value when it exists, and the global mean of the field when it is missing; the mask channel is a binary variable, where 1 indicates that the value exists and 0 indicates that the value is a filler value. For example, the appraised value of collateral for asset No. 37, which is 12 million yuan, is a true record, so its numerical channel is filled with log(1+12000000)≈16.30, and the mask channel is filled with 1. The appraised value of collateral for asset No. 128, which is 2 million yuan, is also filled with 1 in the mask after transformation. There is another asset in the candidate pool, No. 203, whose appraised value field is missing. The numerical channel is filled with the transformed value of the global mean, for example, -0.22, and the mask channel is filled with 0 to explicitly inform the model that this value is a filler value.

[0066] Categorical fields will treat "unknown / missing" as an independent category and perform one-hot encoding instead of using a mask. For long-tailed distribution variables such as principal amount and area, a logarithmic transformation will be performed followed by quantile truncation (truncation to the 1% and 99th quantiles).

[0067] For long texts such as collateral addresses, embeddings are extracted using a lightweight pre-trained language model, and a bottleneck layer is added on top of it (the dimension is reduced to 8 dimensions) to force the retention of only the most core semantics related to the valuation.

[0068] The obtained data is robustly scaled based on the interquartile range, and the median is used instead of the mean to remove the leverage effect of extreme outliers during the data cleaning stage.

[0069] Step 1.4: Using a conditional variational autoencoder, the input feature vector of each asset is encoded into a probability embedding consisting of a mean vector and a standard deviation vector. The standard deviation vector is used to explicitly quantify the information uncertainty of the asset itself. Specifically:

[0070] Construct a conditional variational autoencoder (CVAE). Let the input feature vector processed in step 1.3 be denoted as... The asset class criteria are denoted as (One-hot encoding of categorical fields). The encoder uses a three-layer fully connected network, with each layer followed by batch normalization and ReLU activation functions, outputting a mean vector. Sum of logarithmic variance Both dimensions are (In this embodiment, it is set to 64 dimensions), then the probability embedding ;

[0071] Latent variables are sampled from the encoded distribution using reparameterization techniques. ; The sampling coefficients are... It follows a standard normal distribution. It is the identity matrix;

[0072] The decoder employs a three-layer fully connected network symmetrical to the encoder, with the input... and conditions The output adopts a multi-output head structure: for numerical fields, it outputs the mean and log standard deviation of the Gaussian distribution; for categorical fields, it outputs the categorical probability normalized by Softmax; and for the mask, it outputs the probability of Sigmoid activation.

[0073] The training loss function of the decoder consists of the reconstruction loss and the KL divergence term:

[0074] Training loss function ; To rebuild the lost items, This is the KL divergence term;

[0075] in, For encoder distribution, For decoder distribution, This is a conditional prior distribution. for Follows encoder distribution, To perform KL divergence calculation;

[0076] The reconstruction loss is calculated separately for each output head, i.e., the numerical field is the Gaussian negative log-likelihood, the categorical field is the cross-entropy, and the mask is the binary cross-entropy. The reconstruction loss is only calculated for fields with a mask of 1 (true existence). The error of the filled value (mask=0) is masked to prevent the model from learning to fit the mean of arbitrarily filled values.

[0077] To prevent overfitting, for each new asset, its input features are reconstructed using a decoder. If the reconstruction error of multiple key fields (principal, appraised value, area) exceeds the threshold of 3 standard deviations on the training set at the same time, it is marked as an "out-of-distribution asset" and its embedding is not used directly but is handed over to manual verification.

[0078] Step 2: Construct an interactive encoder. The interactive encoder interactively encodes structured features to obtain context-aware embeddings and package-level aggregated representations of assets within the target distressed asset package. The interactive encoder is a multi-head self-attention network containing explicit relational biases. Specifically, it includes the following steps:

[0079] Based on probabilistic embedding, explicit relationship edges between two assets are extracted, specifically four types of explicit relationships:

[0080] Shared guarantor relationship: If the guarantor and mortgagor of two assets are the same, a shared guarantor edge is established; Shared collateral relationship (simultaneously recording the seizure priority bias, 0 for the first seizure, -1 for the second seizure, and -2 for the third seizure): If the claims of two assets are executed on the same collateral, a collateral edge is established.

[0081] Corporate relationships (equity relationships, upstream and downstream of the supply chain, or the same actual controller): If the debtor has direct equity relationships or upstream and downstream of the supply chain, then a corporate relationship edge is established.

[0082] The relationship between regions and industries is homogeneous. If assets in the same city and industry are homogeneous, then a regional industry homogeneous edge is established.

[0083] For example, Asset No. 37 and Asset No. 128 share the same guarantor and are both located in the Yangtze River Delta region. Therefore, a shared guarantor edge and a regional industry homogeneity edge are established between the two assets.

[0084] Assign a learnable edge embedding to each explicit relation edge. And the two assets , Relationship edge ; For assets The geometry of all types of relationships between assets, if there are no explicit relationships between assets. It is a zero vector.

[0085] For each attention head in a multi-head self-attention network, the edge embedding, after linear transformation, is used as a bias term and added to the scaled dot product attention score to obtain the original score for each attention head. This ensures that the attention weights depend simultaneously on semantic similarity and explicit relations. Specifically:

[0086] For input Linear projection is performed to obtain the query, key, and value matrix. There are H attention heads (H=8 in this embodiment), and each head has a dimension... .

[0087] ;in For the head The projection matrix. For querying the matrix, The key matrix, It is a value matrix.

[0088] The raw score for each attention head is obtained by adding the scaled dot product to the relational bias term, i.e.: the raw score of the attention head. Among them, the relation bias term , For the head The learnable parameters allow each attention head to determine its own dependence on explicit relationships.

[0089] Attention weights are normalized using Softmax: ;head The output is a weighted aggregation of value vectors: After being spliced ​​together from multiple outputs, the output is projected through an output matrix. Fusion to obtain head For assets Aggregated context vector ;

[0090] Multiple encoder layers are stacked, each including a multi-head attention network, residual connections, layer normalization, and a per-asset feedforward network, with dropout regularization applied at the outputs of the self-attention network and the feedforward network; the output of the last encoder layer is the context-aware embedding for each bad asset. ;

[0091] The package-level aggregation is obtained by linearly projecting the concatenation results of average pooling and learnable attention pooling. Specifically:

[0092] The context-aware embeddings of all assets within the package are aggregated through two paths: average pooling and learnable attentional pooling.

[0093] Package-level aggregation representation ;in Vector concatenation and attention pooling are used. Introducing Learnable Query Vectors After calculating the attention weights of each asset, a weighted sum is taken to capture the core assets that have the greatest impact on package-level risk and return. The concatenated result is then compressed back through a linear layer. dimension. This is the activation function for a linear layer. For example, a first-order linear function.

[0094] To prevent overfitting, the number of encoder layers is limited to no more than 3, and the number of attention heads is limited to no more than 12.

[0095] Step 3: Define the objective function and constraints for combinatorial optimization. Using a pointer network, repeatedly sample the packet-level aggregate representation to generate a set of candidate solutions that satisfy the constraints and whose objective function values ​​are higher than a preset screening threshold. Specifically:

[0096] Define the decision variables and objective function, including:

[0097] Define binary selection vector , This represents the total size of the candidate asset pool.

[0098] The objective function consists of three terms, including:

[0099] Expected internal rate of return The recovery factor is approximated by a power of the average recovery period, and the formula is as follows: ;

[0100] in, This represents the total expected recovery of the asset portfolio. For the assets predicted in step 4 Expected recovery amount; The total acquisition cost of the portfolio, For assets The expected acquisition cost; The weighted average expected payback period (in years). For assets The expected payback period.

[0101] Tail risk penalty Based on conditional value of risk The difference between the expected recovery amount and the actual recovery amount is calculated using the following formula:

[0102] ;

[0103] in Confidence level (e.g., 95%) conditional value at risk, i.e., the conditional expectation when the recovery is lower than the value at risk; This is the risk aversion coefficient (the value range in this embodiment is 0.5-2.0).

[0104] Complexity penalty Penalties will be imposed for exceeding the limit on the number of asset transactions and for having an excessively high proportion of small assets. The calculation formula is as follows:

[0105] ;

[0106] in To select the number of assets, The target upper limit, This is the threshold for recovering small amounts of assets. To prevent small positive numbers with a denominator of zero. , This is the penalty coefficient;

[0107] Then the objective function ;

[0108] All penalty coefficients were backtested using historical asset portfolio data to maximize the Spearman rank correlation coefficient between the ranking of the objective function value and the ranking of the historical true expected internal rate of return.

[0109] Define constraints, including hard constraints: the total acquisition cost is less than a preset cost threshold; and soft constraints: the proportion of collateral types is less than a preset percentage threshold.

[0110] The strategy network is constructed using a pointer network laser;

[0111] The policy network uses the package-level aggregation representation of the current asset as the query and global interactive embedding as the key to calculate and obtain the asset. Relevance score to the current packet state ;in, The current package-level aggregation of assets represents, Embedded for global interaction; It is the bias vector; This is the projection vector, used to map the correlation to a scalar; For activation functions;

[0112] The global interactive embedding is obtained by inputting all assets in the distressed asset portfolio into the multi-head self-attention network;

[0113] The probability of an asset being selected from the target non-performing asset portfolio is calculated based on the correlation score and temperature coefficient; the formula for calculating the probability is as follows: ,in, For assets The probability of being selected A set of indexes for selectable assets; The temperature coefficient is used; the selectable assets are those that meet the hard constraints.

[0114] Calculate the objective function value of assets whose probability exceeds a preset selection probability threshold, and sort them in descending order according to the size of the objective function value. Assets with objective function values ​​higher than a preset screening threshold constitute a candidate portfolio of non-performing assets.

[0115] Select K different temperature coefficients (to increase diversity), calculate the probability of each, and calculate the corresponding objective function value based on the probability to obtain K different candidate combinations, which constitute a candidate scheme set; the temperature coefficients range from 1 to 2.

[0116] Step 4: For candidate combinations in the candidate solution set, use a deep ensemble encoder network to output the Gaussian mixture distribution of the future recovery amount for the corresponding candidate combination. This specifically includes the following steps:

[0117] Obtain context-aware embeddings and package-level aggregated representations of assets in candidate portfolios;

[0118] By performing a nonlinear transformation and summing the context-aware embeddings of all assets within a candidate combination using a deep ensemble encoder, a permutation-invariant representation is obtained. ; ; Candidate combinations Assets within The standard deviation vector in the probability embedding; The average relevance score; Candidate combinations Assets within The context-aware embedding is obtained by reusing the encoder layer in step 2. A multilayer perceptron for a depth encoder;

[0119] The resulting permutation-invariant representation is concatenated with the packet-level scalar conditional vector and then passed through a global decoding network to output a mixture Gaussian distribution of the future recovery amount for the candidate combinations. , , . respectively represent the first The weights, standard deviations, and means of each component in the mixture.

[0120] Probability density function of future recovery amount . This refers to the set of assets corresponding to the candidate asset portfolio. This represents the future recovery amount of the candidate portfolio. The number of components is a hyperparameter, and in this embodiment, it is set to 4.

[0121] Mixed weights Mixed standard deviation Mixed mean . , For activation function, This represents a Gaussian distribution.

[0122] The package-level scalar condition vector includes the total cost of the candidate portfolio assets, the average payback period, and the number of assets.

[0123] The depth ensemble encoder includes multiple multilayer perceptrons. Each multilayer perceptron includes three fully connected layers and a GELU activation function; the global decoding network includes four fully connected layers and a Sigmoid activation function.

[0124] Monte Carlo dropout is enabled after the fully connected layers of the deep ensemble encoder and the fully connected layers of the global decoder network, with a dropout rate of 0.05. Sub-independent forward propagation to obtain the posterior predicted distribution ; The probability density function is represented by the first... This propagates forward.

[0125] For example, for a candidate portfolio containing 50 assets. The candidate plan has a total acquisition cost of 380 million yuan. The output is a Gaussian mixture distribution ( After activation, the components of the mixture are as follows:

[0126] Component 1: The mixed weight is 0.1, the mixed mean is 280 million yuan, and the mixed standard deviation is 60 million yuan;

[0127] Component 2: The mixed weight is 0.35, the mixed mean is 450 million yuan, and the mixed standard deviation is 80 million yuan;

[0128] Component 3: The mixed weight is 0.4, the mixed mean is 580 million yuan, and the mixed standard deviation is 70 million yuan;

[0129] Component 4: The mixed weight is 0.5, the mixed mean is 720 million yuan, and the mixed standard deviation is 100 million yuan;

[0130] but: 100 million yuan. 100 million yuan.

[0131] The model uncertainty measure is obtained by calculating the variance of the expected value of the posterior predicted distribution. Expected value of the posterior prediction distribution , It represents the average of the expected values ​​of the posterior prediction distribution.

[0132] Step 5: Based on a Gaussian mixture distribution, calculate the reliability index and comprehensive priority score for each candidate combination; output a recommended solution set based on the reliability index and comprehensive priority score for manual review. This specifically includes the following steps:

[0133] The reliability index is obtained by normalizing and weighting the entropy of the posterior prediction distribution and the model uncertainty measure, and then taking the reciprocal. ;in These are the normalized entropy of the posterior predicted distribution and the normalized model variance, which are measures of model uncertainty. The weight hyperparameter is set to a value of [value] in this embodiment. .

[0134] The comprehensive priority score is calculated by subtracting the penalty for the difference between the expected internal rate of return and the conditional value at risk (IVR) and the IRR, and then adding the reliability bonus. The calculation formula is as follows: .in, For reliability reward coefficient, To determine the overall priority score, To accurately predict the internal rate of return, , The average recovery period is... For acquisition costs, Candidate combinations The expected amount of future recoveries.

[0135] The candidate combinations are sorted in descending order of their comprehensive priority scores to obtain the top... One recommended option;

[0136] The Jaccard similarity of the asset selection vectors of candidate portfolios within the recommended scheme is calculated using the following formula:

[0137] ; For Jaccard similarity. The asset selection vector. ;in, For the first The number of assets in each candidate portfolio. Assets The selected value, Assets Being chosen Assets Not selected;

[0138] Before screening using a diversity filtering algorithm (greedy nonmaximum suppression) One recommended approach is to retain entries with a similarity to Jaccard below a similarity threshold of 0.7. The recommended solution set consists of several solutions. The similarity threshold ranges from 0.5 to 0.9.

[0139] At the same time, at least one representative scheme that meets the preset risk attribute dimensions must be retained, including: a short payback period scheme or a scheme with a residential component of more than 60%.

[0140] An electronic device for executing the aforementioned intelligent portfolio optimization pricing method for distressed asset portfolios.

[0141] The present invention also provides a computer-readable storage medium storing a computer program that is executed by a processor to implement the aforementioned intelligent portfolio optimization pricing method for distressed asset portfolios.

[0142] The processes described above with reference to the flowcharts in the embodiments disclosed in this invention can be implemented as computer software programs. The embodiments disclosed in this invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication component, and / or installed from a removable medium. When the computer program is executed by a central processing unit (CPU), it performs the functions defined in the methods of this application. It should be noted that the computer-readable medium described above in this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. Computer-readable storage media can be, for example, but not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wire segments, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on a computer-readable medium may be transmitted using any suitable medium, including but not limited to: wireless segments, wire segments, optical fibers, RF, etc., or any suitable combination thereof.

[0143] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0144] Those skilled in the art should understand that the embodiments of the present invention described above and shown in the accompanying drawings are merely examples and do not limit the present invention. The purpose of the present invention has been fully and effectively achieved. The functions and structural principles of the present invention have been shown and explained in the embodiments. Without departing from the principles described, the implementation of the present invention may have any changes or modifications.

Claims

1. A smart portfolio optimization pricing method for non-performing asset portfolios, characterized in that, The method includes: Acquire multi-source heterogeneous data of the target non-performing asset portfolio, and extract the structured features of the non-performing assets from the multi-source heterogeneous data; An interactive encoder is constructed to interactively encode structured features, thereby obtaining context-aware embeddings and package-level aggregated representations of assets within the target non-performing asset package; the interactive encoder is a multi-head self-attention network containing explicit relational biases. Define the objective function and constraints for combinatorial optimization, and generate a set of candidate solutions that satisfy the constraints and whose objective function values ​​are higher than a preset screening threshold by multiple random samplings of the packet-level aggregate representation through a pointer network; For candidate combinations in the candidate scheme set, a deep ensemble encoder network is used to output a Gaussian mixture distribution of the future recovery amount of the corresponding candidate combination; Based on a Gaussian mixture distribution, the reliability index and comprehensive priority score of each candidate combination are calculated; a set of recommended solutions is output based on the reliability index and comprehensive priority score for manual review.

2. The intelligent portfolio optimization pricing method for non-performing asset portfolios according to claim 1, characterized in that, Also includes: A two-dimensional table of non-performing assets is constructed using structured features; the structured features include legal document fields, collateral file fields, and ledger data fields; the rows of the two-dimensional table represent assets, and the columns represent a field. For numerical structured features, explicit missing encoding and masking are performed, and the mask indicating whether the data actually exists is treated as an independent feature. One-hot encoding is performed on the categorical structured features; The missing code, mask, and one-hot code are concatenated to form the input feature vector; Using a conditional variational autoencoder, the input feature vector of each asset is encoded into a probability embedding consisting of a mean vector and a standard deviation vector, wherein the standard deviation vector is used to explicitly quantify the information uncertainty of the asset itself. The decoder outputs the Gaussian distribution parameters and recovery probabilities of numerical structured features and the class probabilities of categorical structured features through the conditional variational autoencoder.

3. The intelligent portfolio optimization pricing method for non-performing asset portfolios according to claim 2, characterized in that, The step of interactively encoding structured features through an interactive encoder to obtain the context-aware embedding and package-level aggregated representation of each non-performing asset includes: Based on probabilistic embedding, explicit relationship edges between two assets are extracted. These explicit relationship edges include shared guarantor relationships, shared collateral relationships, corporate affiliations, and regional-industry homogeneous relationships. A learnable edge embedding is assigned to each type of explicit relationship edge. By using each attention head of the multi-head self-attention network, the edge embedding is linearly transformed and used as a bias term, which is added to the scaled dot product attention score to obtain the original score of each attention head, so that the attention weight depends on both semantic similarity and explicit relation. Multiple encoder layers are stacked, each layer including a multi-head attention network, residual connections, layer normalization and a per-asset feedforward network, and dropout regularization is applied at the output of the self-attention network and the output of the feedforward network; the output of the last encoder layer is the context-aware embedding of each bad asset. The encoder has no more than 3 layers and no more than 12 attention heads; The package-level aggregation is obtained by linearly projecting the concatenation result of average pooling and learnable attention pooling.

4. The intelligent portfolio optimization pricing method for non-performing asset portfolios according to claim 3, characterized in that, The defined objective function and constraints for combinatorial optimization include: The objective function is defined as the difference between the risk-adjusted expected internal rate of return and the tail risk penalty and complexity penalty; Define constraints, including hard constraints: the total acquisition cost is less than a preset cost threshold; and soft constraints: the proportion of collateral types is less than a preset percentage threshold. The expected internal rate of return is approximately expressed by a power relationship between the expected recovery amount, the total acquisition cost, and the average recovery period; The tail risk penalty is quantified based on the difference between the conditional value at risk and the expected return; the complexity penalty includes penalties for the total number of assets exceeding a threshold and for the excessive proportion of small assets; the package-level aggregate representation is sorted according to the objective function value; wherein, the penalty coefficient weights in the tail risk penalty and the complexity penalty are calibrated through time-layered cross-validation to maximize the rank correlation coefficient between the objective function ranking and the historical true internal rate of return ranking.

5. The intelligent portfolio optimization pricing method for non-performing asset portfolios according to claim 4, characterized in that, The method of using a pointer network as a policy network to generate a set of candidate solutions that satisfy the constraints and whose objective function values ​​are higher than a preset screening threshold through multiple random samplings of the packet-level aggregate representation includes: The policy network uses the package-level aggregation representation of the current asset as the query and global interactive embedding as the key to calculate and obtain the asset. Relevance score to the current packet state ;in, The current package-level aggregation of assets represents, Embedded for global interaction; It is the bias vector; This is the projection vector, used to map the correlation to a scalar; For activation functions; The global interactive embedding is obtained by inputting all assets in the distressed asset portfolio into the multi-head self-attention network; The probability of an asset being selected from the target non-performing asset portfolio is calculated based on the correlation score and temperature coefficient; the formula for calculating the probability is as follows: ,in, For assets The probability of being selected A set of indexes for selectable assets; The temperature coefficient is used; the selectable assets are those that meet the hard constraints. Calculate the objective function value of assets whose probability exceeds a preset selection probability threshold, and sort them in descending order according to the size of the objective function value. Assets with objective function values ​​higher than a preset screening threshold constitute a candidate portfolio of non-performing assets. K different temperature coefficients are selected, and their probabilities are calculated for each. Based on the probabilities, the corresponding objective function values ​​are calculated, thereby obtaining K different candidate combinations, which constitute a candidate scheme set; the temperature coefficients range from 1 to 2. The package status includes the currently selected assets, the remaining available budget, and a package-level aggregate representation of the selected assets; actions include selecting an asset from those that meet hard constraints to add to the asset package and a learnable stop action token.

6. The intelligent portfolio optimization pricing method for non-performing asset portfolios according to claim 5, characterized in that, The process of using a deep ensemble encoder network to output a Gaussian mixture distribution of the future recovery amount for each candidate combination in the candidate scheme set includes: Obtain context-aware embeddings and package-level aggregated representations of assets in candidate portfolios; The context-aware embeddings of all assets in the candidate combination are first nonlinearly transformed and summed by a deep set encoder to obtain a permutation-invariant representation. The obtained permutation-invariant representation is concatenated with the package-level scalar condition vector and then output by a global decoding network as a mixture of Gaussian distributions of the future recovery amount of the candidate combination; the package-level scalar condition vector includes the total cost of the candidate combination assets, the average recovery period, and the number of assets. The deep ensemble encoder includes multiple multilayer perceptrons, each multilayer perceptron including three fully connected layers and a GELU activation function; the global decoding network includes four fully connected layers and a Sigmoid activation function. Monte Carlo dropout is enabled after the fully connected layers of the deep ensemble encoder and the fully connected layers of the global decoder network, with a dropout rate of 0.

05. The model uncertainty measure is obtained by performing independent forward propagation to obtain the posterior prediction distribution; the variance of the expected value of the posterior prediction distribution is used to calculate the model uncertainty measure.

7. The intelligent portfolio optimization pricing method for non-performing asset portfolios according to claim 6, characterized in that, The reliability index and comprehensive priority score for each candidate combination are calculated based on a Gaussian mixture distribution of the future recovery amounts of the candidate combinations, including: The reliability index is obtained by normalizing and weighting the entropy of the posterior prediction distribution and the model uncertainty measure, and then taking the reciprocal. The overall priority score is calculated by subtracting the penalty for the difference between the expected internal rate of return and the conditional value at risk (IVR) and the internal rate of return, and adding the reliability bonus.

8. The intelligent portfolio optimization pricing method for non-performing asset portfolios according to claim 7, characterized in that, The set of recommended solutions output based on reliability index and comprehensive priority score includes: The candidate combinations are sorted in descending order of their comprehensive priority scores to obtain the top... One recommended option; Jaccard similarity of asset selection vectors for candidate combinations within the recommended scheme; the asset selection vectors ;in, For the first The number of assets in each candidate portfolio. Assets The selected value, Assets Being chosen Assets Not selected; Before screening using diversity filtering algorithms One recommended scheme is to retain those with a similarity score below the similarity threshold in Jaccard. The recommended solution set consists of several solutions; at the same time, at least one representative solution that meets the preset risk attribute dimension is forcibly retained; the similarity threshold ranges from 0.5 to 0.

9.

9. An electronic device, characterized in that, The electronic device is used to execute the intelligent portfolio optimization pricing method for distressed asset portfolios as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that is executed by a processor to implement the intelligent portfolio optimization pricing method for distressed asset portfolios as described in any one of claims 1-8.

Images