Regulatory initial margin (IM) is being implemented across the financial industry in accordance with BCBS-IOSCO requirements. The regulations target uncleared over-the-counter derivative trading and, among other issues, aim to mitigate counter party credit risk by defining a comprehensive set of rules for initial margin between trading parties. Computing the funding costs of IM, or Margin Valuation Adjustment (MVA), is a major challenge for xVA systems as it requires the future projection of dynamic IM positions. This is particularly challenging for callable products, such as Bermudan swaptions which are complex to price and require path wise exercise tracking in exposure simulations. Brute force simulation of future IM is not feasible due to the excessive computational demands of model calibration and numerical pricing methods. Approximate MVA methods, such as regression techniques, are difficult to design due to the high-dimensionality of the problem. In this paper, we propose a method based on Deep Neural Networks to approximate the Bermudan swaption pricing function and sensitivities. We exploit neural network's high-dimensionality and universal approximation properties to train networks based on prices and sensitivities generated from existing numerical pricing models. The trained neural networks are then used for extremely fast IM simulation where computationally intense numerical methods are replaced by optimized and hardware accelerated neural network inference. We demonstrate that the neural network models deliver exceptional performance, capable of pricing Bermudan Swaption MVAs over 100,000 times faster than traditional approaches while maintaining a high degree of accuracy.
{"title":"Deep MVA: Deep Learning for Margin Valuation Adjustment of Callable Products","authors":"Anup Aryan, Allan Cowan","doi":"10.2139/ssrn.3634059","DOIUrl":"https://doi.org/10.2139/ssrn.3634059","url":null,"abstract":"Regulatory initial margin (IM) is being implemented across the financial industry in accordance with BCBS-IOSCO requirements. The regulations target uncleared over-the-counter derivative trading and, among other issues, aim to mitigate counter party credit risk by defining a comprehensive set of rules for initial margin between trading parties. Computing the funding costs of IM, or Margin Valuation Adjustment (MVA), is a major challenge for xVA systems as it requires the future projection of dynamic IM positions. This is particularly challenging for callable products, such as Bermudan swaptions which are complex to price and require path wise exercise tracking in exposure simulations. Brute force simulation of future IM is not feasible due to the excessive computational demands of model calibration and numerical pricing methods. Approximate MVA methods, such as regression techniques, are difficult to design due to the high-dimensionality of the problem. In this paper, we propose a method based on Deep Neural Networks to approximate the Bermudan swaption pricing function and sensitivities. We exploit neural network's high-dimensionality and universal approximation properties to train networks based on prices and sensitivities generated from existing numerical pricing models. The trained neural networks are then used for extremely fast IM simulation where computationally intense numerical methods are replaced by optimized and hardware accelerated neural network inference. We demonstrate that the neural network models deliver exceptional performance, capable of pricing Bermudan Swaption MVAs over 100,000 times faster than traditional approaches while maintaining a high degree of accuracy.","PeriodicalId":406435,"journal":{"name":"CompSciRN: Other Machine Learning (Topic)","volume":"58 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2020-06-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"116438524","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Joint prediction and optimization problems are common in many business applications ranging from customer relationship management and marketing to revenue and retail operations management. These problems involve a first-stage learning model, where outcomes are predicted from features, and a second-stage decision process, which selects the optimal decisions based on these outcomes. In practice, these two stages are conducted separately, but is sub-optimal. In this work, we propose a novel model that solves both parts as a whole, but is computationally tractable under many circumstances. Specifically, we introduce the notion of a regularizer that measures the value of a predictive model in terms of the cost incurred in the decision process. We term this decision-driven regularization, and it is centred on the premise that the bias-variance trade-off in the learning problem is not transformed linearly by the subsequent decision problem. Additionally, this accounts for the ambiguity in the definition of the cost function, which we identify. We prove key properties of our model, namely, that it is consistent, robust to wrong estimation, and has bounded bias. We also examine special cases under which we draw links to existing models in the literature, propose hybrid models and are able to describe their effectiveness using our framework as a theoretical basis. In our numerical experiments, we illustrate the behaviour of our model, and its performance against other models in the literature.
{"title":"Decision-Driven Regularization: Harmonizing the Predictive and Prescriptive","authors":"G. Loke, Qinshen Tang, Yangge Xiao","doi":"10.2139/ssrn.3623006","DOIUrl":"https://doi.org/10.2139/ssrn.3623006","url":null,"abstract":"Joint prediction and optimization problems are common in many business applications ranging from customer relationship management and marketing to revenue and retail operations management. These problems involve a first-stage learning model, where outcomes are predicted from features, and a second-stage decision process, which selects the optimal decisions based on these outcomes. In practice, these two stages are conducted separately, but is sub-optimal. In this work, we propose a novel model that solves both parts as a whole, but is computationally tractable under many circumstances. Specifically, we introduce the notion of a regularizer that measures the value of a predictive model in terms of the cost incurred in the decision process. We term this decision-driven regularization, and it is centred on the premise that the bias-variance trade-off in the learning problem is not transformed linearly by the subsequent decision problem. Additionally, this accounts for the ambiguity in the definition of the cost function, which we identify. We prove key properties of our model, namely, that it is consistent, robust to wrong estimation, and has bounded bias. We also examine special cases under which we draw links to existing models in the literature, propose hybrid models and are able to describe their effectiveness using our framework as a theoretical basis. In our numerical experiments, we illustrate the behaviour of our model, and its performance against other models in the literature.","PeriodicalId":406435,"journal":{"name":"CompSciRN: Other Machine Learning (Topic)","volume":"19 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2020-06-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"122099030","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
DataGene is developed to identify data set similarity between real and synthetic datasets as well as train, test, and validation datasets. For many modelling and software development tasks there is a need for datasets to have share similar characteristics. This has traditionally been achieved with visualizations, DataGene seeks to replace these visual methods with a range of novel quantitative methods. Please see the GitHub repository to inspect and install the Python code.
{"title":"DataGene: A Framework for Dataset Similarity","authors":"Derek Snow","doi":"10.2139/ssrn.3619626","DOIUrl":"https://doi.org/10.2139/ssrn.3619626","url":null,"abstract":"DataGene is developed to identify data set similarity between real and synthetic datasets as well as train, test, and validation datasets. For many modelling and software development tasks there is a need for datasets to have share similar characteristics. This has traditionally been achieved with visualizations, DataGene seeks to replace these visual methods with a range of novel quantitative methods. Please see the GitHub repository to inspect and install the Python code.","PeriodicalId":406435,"journal":{"name":"CompSciRN: Other Machine Learning (Topic)","volume":"89 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2020-06-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"116811292","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
B. Rani, M. Kumari, Kumari Sobha ,, Pinki Kumari, Jyotipragyan Majhi, Subham Chakraborty
Big data is such a field where we collect and store the related data from agriculture. In this paper we researched about advanced technology of agriculture with big data, Smart Farming/Organic Farming which is lacking in India. There are many farmers in India who are deprived of advanced technology.In which we want to provide true knowledge to the farmers through soil, irrigation, environment, pesticides and genetic engineering so that the economy of the farmer can be improved so that more and more people are connected with agriculture.
{"title":"Application of Big Data in Smart Agriculture","authors":"B. Rani, M. Kumari, Kumari Sobha ,, Pinki Kumari, Jyotipragyan Majhi, Subham Chakraborty","doi":"10.2139/ssrn.3611514","DOIUrl":"https://doi.org/10.2139/ssrn.3611514","url":null,"abstract":"Big data is such a field where we collect and store the related data from agriculture. In this paper we researched about advanced technology of agriculture with big data, Smart Farming/Organic Farming which is lacking in India. There are many farmers in India who are deprived of advanced technology.In which we want to provide true knowledge to the farmers through soil, irrigation, environment, pesticides and genetic engineering so that the economy of the farmer can be improved so that more and more people are connected with agriculture.","PeriodicalId":406435,"journal":{"name":"CompSciRN: Other Machine Learning (Topic)","volume":"50 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2020-05-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"115518868","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
In this paper, I prove my promise that Machine Learning is one of the most important parts to provide tools and methods to go deeper and nurture the data properly. The most amazing part is to analyze the large chunks of data in a very precise way, and high-value predictions that can guide better decisions and smart actions in real-time without human intervention. I give an overview over different proposed structures of Data Science and mention the impact of Machine learning such as algorithms, model evaluation and selection, pipeline. I also indicate all misconceptions when neglecting Machine learning reasoning.
{"title":"Data Science: The Impact of Machine Learning","authors":"S. Islam","doi":"10.2139/ssrn.3674357","DOIUrl":"https://doi.org/10.2139/ssrn.3674357","url":null,"abstract":"In this paper, I prove my promise that Machine Learning is one of the most important parts to provide tools and methods to go deeper and nurture the data properly. The most amazing part is to analyze the large chunks of data in a very precise way, and high-value predictions that can guide better decisions and smart actions in real-time without human intervention. I give an overview over different proposed structures of Data Science and mention the impact of Machine learning such as algorithms, model evaluation and selection, pipeline. I also indicate all misconceptions when neglecting Machine learning reasoning.","PeriodicalId":406435,"journal":{"name":"CompSciRN: Other Machine Learning (Topic)","volume":"27 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2020-05-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"131891432","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Quantitative factor portfolios generally use historical company fundamental data in portfolio construction. What if we could forecast, with a small margin of error, the forward-looking company fundamentals? Using best practices from the science of forecasting and machine learning techniques, namely Random Forests and Gradient Boosting, I try to build a value composite model to sort portfolios based on forecasted fundamentals. I use the in-sample data to train the models to predict forward looking earnings, free cash flow, EBITDA, and Net Operating Profit After Taxes. The combined value portfolio out of sample did not produce statistically significant outperformance verses the equal weight portfolio (as a comparison to the long only value composite) or versus cash (for the long/short portfolio).
{"title":"Gazing into the Future: Using Ensemble Techniques to Forecast Company Fundamentals","authors":"Steven Downey","doi":"10.2139/ssrn.3580018","DOIUrl":"https://doi.org/10.2139/ssrn.3580018","url":null,"abstract":"Quantitative factor portfolios generally use historical company fundamental data in portfolio construction. What if we could forecast, with a small margin of error, the forward-looking company fundamentals? Using best practices from the science of forecasting and machine learning techniques, namely Random Forests and Gradient Boosting, I try to build a value composite model to sort portfolios based on forecasted fundamentals. I use the in-sample data to train the models to predict forward looking earnings, free cash flow, EBITDA, and Net Operating Profit After Taxes. The combined value portfolio out of sample did not produce statistically significant outperformance verses the equal weight portfolio (as a comparison to the long only value composite) or versus cash (for the long/short portfolio).","PeriodicalId":406435,"journal":{"name":"CompSciRN: Other Machine Learning (Topic)","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2020-04-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"130404064","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
We find economically and statistically significant gains when using machine learning for portfolio allocation between the market index and risk-free asset. Optimal portfolio rules for time-varying expected returns and volatility are implemented with two Random Forest models. One model is employed in forecasting the sign probabilities of the excess return with payout yields. The second is used to construct an optimized volatility estimate. Reward-risk timing with machine learning provides substantial improvements over the buy-and-hold in utility, risk-adjusted returns, and maximum drawdowns. This paper presents a new theoretical basis and unifying framework for machine learning applied to both return- and volatility-timing.
{"title":"Machine Learning Portfolio Allocation","authors":"Michael Pinelis, D. Ruppert","doi":"10.2139/ssrn.3546294","DOIUrl":"https://doi.org/10.2139/ssrn.3546294","url":null,"abstract":"We find economically and statistically significant gains when using machine learning for portfolio allocation between the market index and risk-free asset. Optimal portfolio rules for time-varying expected returns and volatility are implemented with two Random Forest models. One model is employed in forecasting the sign probabilities of the excess return with payout yields. The second is used to construct an optimized volatility estimate. Reward-risk timing with machine learning provides substantial improvements over the buy-and-hold in utility, risk-adjusted returns, and maximum drawdowns. This paper presents a new theoretical basis and unifying framework for machine learning applied to both return- and volatility-timing.","PeriodicalId":406435,"journal":{"name":"CompSciRN: Other Machine Learning (Topic)","volume":"80 9","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2020-03-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"132579839","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
We propose a novel approach to the anonymisation of datasets through non-parametric learning of the underlying multivariate distribution of dataset features and generation of the new synthetic samples from the learned distribution. The main objective is to ensure equal (or better) performance of the classifiers and regressors trained on synthetic datasets in comparison with the same classifiers and regressors trained on the original data. The ability to generate unlimited number of synthetic data samples from the learned distribution can be a remedy in fighting overtting when dealing with small original datasets. When the synthetic data generator is trained as an autoencoder with the bottleneck information compression structure we can also expect to see a reduced number of outliers in the generated datasets, thus further improving the generalization capabilities of the classifiers trained on synthetic data. We achieve these objectives with the help of the Restricted Boltzmann Machine, a special type of generative neural network that possesses all the required properties of a powerful data anonymiser.
{"title":"Data Anonymisation, Outlier Detection and Fighting Overfitting with Restricted Boltzmann Machines","authors":"A. Kondratyev, Christian Schwarz, Blanka Horvath","doi":"10.2139/ssrn.3526436","DOIUrl":"https://doi.org/10.2139/ssrn.3526436","url":null,"abstract":"We propose a novel approach to the anonymisation of datasets through non-parametric learning of the underlying multivariate distribution of dataset features and generation of the new synthetic samples from the learned distribution. The main objective is to ensure equal (or better) performance of the classifiers and regressors trained on synthetic datasets in comparison with the same classifiers and regressors trained on the original data. The ability to generate unlimited number of synthetic data samples from the learned distribution can be a remedy in fighting overtting when dealing with small original datasets. When the synthetic data generator is trained as an autoencoder with the bottleneck information compression structure we can also expect to see a reduced number of outliers in the generated datasets, thus further improving the generalization capabilities of the classifiers trained on synthetic data. We achieve these objectives with the help of the Restricted Boltzmann Machine, a special type of generative neural network that possesses all the required properties of a powerful data anonymiser.","PeriodicalId":406435,"journal":{"name":"CompSciRN: Other Machine Learning (Topic)","volume":"21 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2020-01-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"131812088","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Differential machine learning (ML) extends supervised learning, with models trained on examples of not only inputs and labels, but also differentials of labels to inputs. Differential ML is applicable in all situations where high quality first order derivatives wrt training inputs are available. In the context of financial Derivatives risk management, pathwise differentials are efficiently computed with automatic adjoint differentiation (AAD). Differential ML, combined with AAD, provides extremely effective pricing and risk approximations. We can produce fast pricing analytics in models too complex for closed form solutions, extract the risk factors of complex transactions and trading books, and effectively compute risk management metrics like reports across a large number of scenarios, backtesting and simulation of hedge strategies, or capital regulations. The article focuses on differential deep learning (DL), arguably the strongest application. Standard DL trains neural networks (NN) on punctual examples, whereas differential DL teaches them the shape of the target function, resulting in vastly improved performance, illustrated with a number of numerical examples, both idealized and real world. In the online appendices, we apply differential learning to other ML models, like classic regression or principal component analysis (PCA), with equally remarkable results. This paper is meant to be read in conjunction with its companion GitHub repo https://github.com/differential-machine-learning, where we posted a TensorFlow implementation, tested on Google Colab, along with examples from the article and additional ones. We also posted appendices covering many practical implementation details not covered in the paper, mathematical proofs, application to ML models besides neural networks and extensions necessary for a reliable implementation in production.
{"title":"Differential Machine Learning","authors":"Antoine Savine, B. Huge","doi":"10.2139/ssrn.3591734","DOIUrl":"https://doi.org/10.2139/ssrn.3591734","url":null,"abstract":"Differential machine learning (ML) extends supervised learning, with models trained on examples of not only inputs and labels, but also differentials of labels to inputs. Differential ML is applicable in all situations where high quality first order derivatives wrt training inputs are available. In the context of financial Derivatives risk management, pathwise differentials are efficiently computed with automatic adjoint differentiation (AAD). Differential ML, combined with AAD, provides extremely effective pricing and risk approximations. We can produce fast pricing analytics in models too complex for closed form solutions, extract the risk factors of complex transactions and trading books, and effectively compute risk management metrics like reports across a large number of scenarios, backtesting and simulation of hedge strategies, or capital regulations. The article focuses on differential deep learning (DL), arguably the strongest application. Standard DL trains neural networks (NN) on punctual examples, whereas differential DL teaches them the shape of the target function, resulting in vastly improved performance, illustrated with a number of numerical examples, both idealized and real world. In the online appendices, we apply differential learning to other ML models, like classic regression or principal component analysis (PCA), with equally remarkable results. This paper is meant to be read in conjunction with its companion GitHub repo https://github.com/differential-machine-learning, where we posted a TensorFlow implementation, tested on Google Colab, along with examples from the article and additional ones. We also posted appendices covering many practical implementation details not covered in the paper, mathematical proofs, application to ML models besides neural networks and extensions necessary for a reliable implementation in production.","PeriodicalId":406435,"journal":{"name":"CompSciRN: Other Machine Learning (Topic)","volume":"27 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2020-01-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"126581989","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
We study an approach to regression that we call objective-aligned fitting, which is applicable when the regression model is used to predict uncertain parameters of some objective problem. Rather than minimizing a typical loss function, such as squared error, we approximately minimize the objective value of the resulting solutions to the nominal optimization problem. While previous work on objective-aligned fitting has tended to focus on uncertainty in the objective function, we consider the case in which the nominal optimization problem is a two-stage linear program with uncertainty in the right-hand side. We define the objective-aligned loss function for the problem and prove structural properties concerning this loss function. Since the objective-aligned loss function is generally non-convex, we develop a convex approximation. We propose a method for fitting a linear regression model to the convex approximation of the objective-aligned loss. Computational results indicate that this procedure can lead to higher-quality solutions than existing regression procedures.
{"title":"Objective-Aligned Regression for Two-Stage Linear Programs","authors":"Alexander S. Estes, Jean-Philippe P. Richard","doi":"10.2139/ssrn.3469897","DOIUrl":"https://doi.org/10.2139/ssrn.3469897","url":null,"abstract":"We study an approach to regression that we call objective-aligned fitting, which is applicable when the regression model is used to predict uncertain parameters of some objective problem. Rather than minimizing a typical loss function, such as squared error, we approximately minimize the objective value of the resulting solutions to the nominal optimization problem. While previous work on objective-aligned fitting has tended to focus on uncertainty in the objective function, we consider the case in which the nominal optimization problem is a two-stage linear program with uncertainty in the right-hand side. We define the objective-aligned loss function for the problem and prove structural properties concerning this loss function. Since the objective-aligned loss function is generally non-convex, we develop a convex approximation. We propose a method for fitting a linear regression model to the convex approximation of the objective-aligned loss. Computational results indicate that this procedure can lead to higher-quality solutions than existing regression procedures.","PeriodicalId":406435,"journal":{"name":"CompSciRN: Other Machine Learning (Topic)","volume":"79 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2019-10-14","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"131389884","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：