Абай атындағы ҚазҰпу-нің хабаршысы, «Физика-математика ғылымдары» сериясы, №3 (7 9 ), 2022 150 мрнти
The Process of Building a Valuation Map Model
жүктеу/скачать 1,21 Mb. Pdf көрінісі
|
вестник КазНПУ 2
| The Process of Building a Valuation Map Model
Over the years, several different modeling techniques have been observed to implement credit scoring. They range from parametric or non-parametric, statistical or machine learning to supervised or unsupervised algorithms. the latter methods involve rather difficult spreads involving hundreds or thousands of different models, different test structures, and ensemble methods with several learning methods to achieve greater accuracy. Ignoring this diversity, the method of modeling is different - the credit rating model. Commonly referred to as a normal scorecard, it is based on logistic regression as the underlying model. In comparison with other modeling methods, this method meets almost all the claims, in fact, which prepares it with the desired alignment between practitioners and is used by almost 90% of the creators of indicator charts. A feature system model is simple to build, understand, and implement, and can be nimbly implemented. Being a hybrid of statistics and machine learning, its predictive accuracy is comparable to other more complex methods, and its estimates have every chance of being applied precisely as probability estimates and, therefore, for direct data entry for risk-based pricing. This is quite fundamental for lenders that comply with the Basel II regulatory framework. While instinctive and straightforward to interpret and justify, scorecards are mandated by regulators as the exclusive way to model credit risk in some countries. The business intelligence test uses these methods as a measurement data test, a method of combining univariate and multivariate statistics, and all kinds of data visualization methods [5]. Correlation, cross- tabulation, scatter, timeline test, and supervised and unsupervised segmentation test are considered common methods. Segmentation is special because it determines when a certain number of scorecards are needed. The choice of variables based on the results of business intelligence analysis is introduced with the division of the data mining view as a minimum number into 2 different sections: the section for study and testing. Variable selection is a set of model candidate variables whose significance is tested during model learning. Candidate model variables are still popular as autonomous variables, predictors, attributes, model moments, covariates, regressors, functions, or properties [6]. The main task is to find the right set of variables so that the scorecard model can not only rank buyers based on the likelihood of their having bad debts, but also consider the possibility of their having bad debts. Typically, this means choosing statistically important variables in the predictive model and having an equilibrium set of predictors (usually 8-15 is a good balance) to approximate a 360-degree view of buyers. In addition to buyer- specific risk features, we still need to consider the likelihood of connecting regular risk moments to account for financial drift and volatility [7]. When choosing variables, there are a few restrictions: - For starters, the model typically has some high predictor variables that are prohibited by legal, ethical, or regulatory rules. - Secondly, some variables have every chance of being unattainable or of low quality at the modeling or manufacturing steps. Apart from this, there is every chance of being significant variables that were not recognized as such, for example, due to the periodic oversight of the population selection or because of such that their model effect would be inconsistent due to multicollinearity. - And, in the end, the last text every time will be the case, and he can insist on connecting only variables that are significant for the business, or to insistently ask for uniformly growing or decreasing effects. All these limitations are considered likely sources of periodic misses, which in fact makes it difficult for data scientists to minimize periodic selection misses. Common preventive measures in variable selection include: - Collaboration with experts in the given field to identify significant variables; - Awareness of every dilemma related to data source, reliability or measurement error; - data cleaning; - Introduction of control variables to account for unresolved variables or certain activities such as financial drift. It is important to understand that the choice of variables is an iterative process that happens throughout the entire process of building a model. - It occurs before model fitting by reducing the number of variables in the data mining view to a manageable set of candidate variables; - lasts in the process of learning the model, where subsequent reduction is performed as a result of statistical insignificance, multicollinearity, small contributions, or penalties to avoid overfitting; - Lasts during the evaluation and testing of the model; ВЕСТНИК КазНПУ им. Абая, серия «Физико-математические науки», № 3 (7 9 ), 2022 г. 153 - Ends during business assertion, where the readability and interpretability of the model play a significant role [8]. The selection of variables is completed after reaching the "golden spot" - this means that further improvement in terms of model accuracy cannot be achieved. The iterative nature of the variable selection process is shown in Fig.1. Figure 1. Iterative nature of the variable selection process In addition to some general recommendations for solving this problem, the data specialist should offer the best approach to converting the signature of the client's data into a powerful information artifact - a representation of data mining. This is probably the most creative and most challenging aspect of the data scientist role, as it requires a solid understanding of the business in addition to statistical and analytical skills. Very often, the key to creating a good model is not the power of a particular modeling method, but the breadth and depth of derived variables that represent a higher level of knowledge about the phenomena being studied. жүктеу/скачать 1,21 Mb. Достарыңызбен бөлісу:
|