Forecasting Model And Optimization Algorithm Based On Variable Selection

In our real life,the causal relationship between each specific result and its associated characteristics is often impossible to form in a short period of time,and there is usually a certain time lag effect.For example: the prediction result of the default risk of listed companies mainly depends on the change of the explanatory variables of multiple categories and multiple windows during the observable period or even the delay period.In this paper,we combine the unconstrained distributed lag model(DLM)with common forecasting models,such as logistic regression model and SVM model,to propose a new framework for early warning systems that can deal with the effects of delay lag.However,in the actual operation process,a large amount of high-dimensional data needs to be used for the learning and training of the prediction model,and in order to avoid the multicollinearity between indicators,we introduce the Lasso penalty term to construct the Lasso-Logistic distribution lag model and the Lasso-SVM distribution lag model,which eliminate potential redundant variables while realizing parameter identification.Secondly,in order to improve the accuracy and robustness of variable selection and prediction effect of the model,different coefficients in the Lasso-Logistic distributed lag model are compressed with different penalty weights,and we propose an adaptive Lasso-Logistic model.Finally,considering the different frequencies of data windows appearing in actual problems,in order to make full use of various mixing data(such as: listed company monthly,quarterly and annual report data,etc.)to improve the prediction accuracy of the model,we propose a bi-convex prediction model with mixed data.In addition,we also use alternating direction multiplier method for solving the above models,which can obtain global optimal solutions for smooth convex optimization,non-smooth convex optimization and non-convex optimization problems.Under the real background of the forecast of the listed companies' financial distress,the financial ratio indicators and macroeconomic factors under the continuous time window are introduced into each model to obtain the best estimates of their coefficients.Finally,we carry out a series of comparative analysis to test the predictive performance of each model,and to detect the impact of early continuous changes inside and outside the listed company on its financial status.Through empirical research,the result shows that the Lasso-Logistic distributed lag model,the Lasso-SVM distributed lag model and the adaptive Lasso-Logistic model all have good feature selection capabilities,and the adaptive Lasso-Logistic model is more prominent.In terms of prediction accuracy,the model proposed in this paper is superior to the traditional Logistic and SVM models in a single time window,and its prediction accuracy is 96.00%.This means that compared with the existing single-window model,the multi-window indicator data model can convey more effective information in predicting financial distress.