Comparison of linear and nonlinear models for panel data forecasting: Debt policy in Taiwan

Hsiao-Tien Pao*, Yao Yu Chih

*Corresponding author for this work

Research output: Contribution to journalArticlepeer-review

10 Scopus citations


This paper discusses the time-series cross-sectional (TSCS) regression and the prediction ability of the artificial neural network (ANN) by examining the panel data of debt ratios of the high tech industry in Taiwan. We build models with these two methods and eight determinants of debt ratio and compare the forecast performances of five models, two ANN nonlinear models and three traditional TSCS linear models. The results show that the sign of each determinant in linear models is the same as that in ANN models. In addition, the insignificant determinants in linear models have low relative sensitivities in ANN models. It seems that these two methods show consistent results for the capital structure determinants. Researchers and practitioners can employ either ANN or traditional statistical model to analyze the important determinants of the capital structure of their firms. The results of comparing the out-of-sample forecasting capabilities of the two methods indicate that: (1) the proposed ANN with 1-year lag model shows better forecast performance than the other three linear models in spite of high or low debt ratio; (2) the debt ratios of the present year are highly related to those of the previous year; and (3) the ANN model is capable of catching sophisticated nonlinear integration effects. Consequently, the ANN method is the more appropriate one between the two methods to be applied to build a forecasting model for the high tech industry in Taiwan.

Original languageEnglish
Pages (from-to)525-541
Number of pages17
JournalReview of Pacific Basin Financial Markets and Policies
Issue number3
StatePublished - 1 Sep 2005


  • Artificial neural networks
  • Capital structure
  • Forecasting
  • Panel data
  • TSCS regression

Fingerprint Dive into the research topics of 'Comparison of linear and nonlinear models for panel data forecasting: Debt policy in Taiwan'. Together they form a unique fingerprint.

Cite this