Forecasting financial outcomes using variable selection techniques

HG Finance

Zhang, Ping ; Tsoukas, Serafeim

University:University of Glasgow

Department:Adam Smith Business School

关键词: HG Finance;   

Others  :  http://theses.gla.ac.uk/41019/1/2019ZhangPhD.pdf

来源: University of Glasgow

【 摘 要 】

Since the activities of market participants can be influenced by financial outcomes, providing accurate forecasts of these financial outcomes can help participants to reduce the risk of adjusting to any market change in the future. Predictions of financial outcomes have usually been obtained by conventional statistical models based on researchers’ knowledge. With the development of data collection and storage, an extensive set of explanatory variables will be extracted from big data capturing more economic theories and then applied to predictive methods, which can increase the difficulty of model interpretation and produce biased estimation. This may further reduce predictive ability. To overcome these problems, variable selection techniques are frequently employed to simplify model selection and produce more accurate forecasts. In this PhD thesis, we aim to combine variable selection approaches with traditional reduced-form models to define and forecast the financial outcomes in question (market implied ratings, Initial Public Offering (IPO) decisions and the failure of companies). This provides benefits for market participants in detecting potential investment opportunities and reducing credit risk. Making accurate predictions of corporate credit ratings is a crucial issue for both investors and rating agencies, since firms’ credit ratings are associated with financial flexibility and debt or equity issuance. In Chapter 2, we attempt to determine market-implied credit ratings in relation to financial ratios, market-driven factors and macroeconomic indicators. We conclude that applying variable selection techniques, the least absolute shrinkage and selection operator (LASSO) and its extension (Elastic net) can improve predictive power. Moreover, the predictive ability of LASSO-selected models is clearly better than that of the benchmark ordered probit model in all out-of-sample predictions. Finally, fewer predictors can be selected into LASSO models controlled by BIC-type tuning parameter to produce more accurate out-of-sample prediction than its counterpart AIC-type selector. Next, the LASSO technique is further applied to binary event prediction. A bank’s decision to go public by issuing an Initial Public Offering (IPO) is the binary object in Chapter 3, which transforms the operations and capital structure of a bank. Much of the empirical investigation in this area focuses on the determinants of the IPO decision, applying accounting ratios and other publicly available information in non-linear models. We mark a break with this literature by offering methodological extensions as well as an extensive and updated US dataset to predict bank IPOs. Combining the least absolute shrinkage and selection operator (LASSO) with a Cox proportional hazard, we uncover value in several financial factors as well as market-driven and macroeconomic variables, in predicting a bank’s decision to go public. Importantly, we document a significant improvement in the model’s predictive ability compared to standard frameworks used in the literature. Finally, we show that the sensitivity of a bank’s IPO to financial characteristics is higher during periods of global financial crisis than in calmer times. Moving to another line of variable selection techniques, Bayesian Model Averaging (BMA) is combined with reduced-form models in Chapter 4. The failure of companies is closely related to the health of the whole economy, since the beginning of the most recent global crisis was the bankruptcy of Lehman Brothers. In this chapter, we forecast the failure of UK private firms incorporating with financial ratios and macroeconomic variables. Since two important financial crises and firm heterogeneities are covered in our dataset, the predictive powers of candidate models in different periods and cross-sections are validated. We first detect that applying BMA to the discrete hazard models can improve the predictive performance in different sub-periods. However, comparing the results with classified models, it should be noted that the Naive Bayes (NB) classifier provides slightly higher predictive accuracy than BMA models of discrete hazard models. Moreover, the predictive performance of the discrete hazard model and its BMA version are more sensitive to adding time or industry dummy variables than other competing models. Considering financial crisis or firm heterogeneity can influence the predictive power of each candidate model in the out-of-sample prediction of failure.