An Evaluation of the Survival Model's Contribution to Thrift Institution Distress Prediction (*)

Abstract

Previous thrift studies use a variety of methodological approaches to predict financial institution distress over a fixed time horizon, including discriminant analysis (Altman, 1977; Pantalone and Platt, 1987), conditional probability models such as logit and probit (Barth et al., 1985; Rudolph and Hamdan, 1988; Barth et al., 1990; Cole, 1993), recursive partition model (McKee and Greenstein, 2000) and expert systems (Elmer and Borowski, 1988; Booth et al., 1989). While many of these models perform reasonably well, opportunities exist for improvements in their predictive ability. Recent studies (e.g., Cole and Gunther, 1995; Helwege, 1996) suggest that survival-time models offer several advantages over previous research methodologies in identifying distressed banks and savings and loan institutions (S&Ls, thrifts) and the related causes of the financial distress. Proponents of these models argue that such methods incorporate a greater degree of time series variation into parameter estimates and also address sample selection problems caused by censoring. In fact, Platt and Platt (1990) argue that temporal variations in economic environments can affect prediction accuracy in bankruptcy studies. Yet little attention has been devoted to temporal issues associated with financial institution failure. Lane et al. (1986), Whalen (1991), and Henebry (1996) use the Cox proportional hazard model (PHM) to estimate the probable time to failure in banks. Cole and Gunther (1995) use a split-population survival-time model to identify differences that may exist between the determinants of bank failure and its timing. While both of these approaches offer valuable insight into the timing of bank failure, they are founded on the stringent assumption that explanatory variable values remain constant over the time horizon implicit in the model specification. Consequently, Helwege (1996) uses a time-varying PHM to identify the determinants of S&L failure and suggests its superiority to statistical methods commonly used in the existing distress literature. Yet these studies have not evaluated the survival model's predictive ability relative to more commonly used methodological approaches, nor have they considered the impa ct of misclassification costs or varying distress definitions on prediction accuracy. Therefore, this study specifically compares the performance of survival-time models with those of two probit models in predicting distress in the S&L industry between 1976 and 1987. The thrift industry during this period provides a particularly rich setting in which to evaluate the survival model methodology as the period's economic and legislative changes yielded large numbers of financially distressed institutions that are critical to model estimation and evaluation. Furthermore, analyzing model effectiveness using this time period's data avoids the structural instability caused by high merger activity in the 1990s. The results of this study indicate that survival models can identify distressed S&Ls with greater accuracy than probit models. More importantly, these findings are robust to different definitions of financial distress, and appear relatively insensitive to costs of misclassification. Although probit models generally yield fewer type II errors (misclassification of a financially non-distressed institution as distressed) and higher total classification accuracy, survival models display higher total classification accuracy when distress is defined broadly. Researchers will find this study interesting because it illustrates how survival-time models can be used to investigate distress and how such methods perform relative to more commonly used statistical approaches. In doing so, the article demonstrates how the timing of financial distress can be predicted using accounting information and also shows how time-varying independent variables are associated with an institution's survival probabilities. …

Cite this work

No public copy is stored for this record, so only the metadata is cited.

Stored only when both papers are in the corpus — 14 citations exist upstream.

Papers that cite the same references — the relational fallback while embeddings are still backfilling.