Research On Random Stacking Forest For Financial Risk Early Warning

In recent years,fintech has developed rapidly at home and abroad.The technology based on fintech is enabling the informatization and intelligence of the financial system in an all-round way.There are huge research space and value in financial intelligence risk control,quantitative trading,asset management and other fields.Financial risk early warning is an important research issue in the field of financial intelligent risk control.It can monitor changes in business operations and financial status of companies,help government regulators,investors,or business owners to find potential risks of enterprises as early as possible,so as to take corresponding measures early.It is of practical significance to the survival of every enterprise and the steady development of the market economy.In view of the existing domestic financial risk early warning research,there are problems such as biased sample selection,short time span,and weak generalization ability.This article combines machine learning,probability statistics,finance,financial management and other cross-domain knowledge.From the perspective of differentiated integration and cluster-assisted classification,we carried out research on financial risk early warning models.Firstly,this paper constructed financial risk early warning data sets,designed and collected15367 pieces of financial risk data in the past 10 years from 2006 to 2016,involving 2,550listed companies and 6 first-level early warning factors and 96 second-level early warning factors.Secondly,a two-stage hybrid ensemble learning method based on clustering and stacking generalization is proposed—random stacking forest for financial risk early warning.This model first calculates the dimensional constraints,then discretizes features by randomly clustering the random feature subspaces of multi-dimensional continuous features and generates high-dimensional features.At the same time,considering that the original features and randomly generated features must have redundancy,multi-layer cross-integration learning is used to ensure the generalization of the model,and the layer-by-layer process is used to perform feature pruning and feature enhancement.The dynamic evaluation of the change of AUC to determine the depth of the stack,so as to achieve a better classification effect.Finally,considering that the financial warning data set is a non-public data set and in order to further verify the effectiveness of the model,in addition to the financial risk early warning experiment,this paper also chooses a similar but open source financial risk control data set-ULB credit card fraud data set as an additional supplementary experiment.In the experiment,AUC and minority F₁ scores were used as evaluation indicators.Random stacking forests were evaluated from four aspects,and compared with mainstream integrated learning algorithms such as Bagging,Random Forest,Adaboost,DBDT,XGBoost,Lightgbm,Catboost and Stacking.The classification effect of the model is improved by 1%-3%,which confirms the effectiveness of the model and the applicability of the research method in this paper in the early warning of financial risk.