Prediction Error Estimation Research Articles

The temporal dynamics of human decision under risk

Event Abstract Back to Event The temporal dynamics of human decision under risk Laurence Hunt1, 2*, Matthew Rushworth3 and Tim E. Behrens3 1 University of Oxford, FMRIB Centre, United Kingdom 2 University of Oxford, Experimental Psychology, United Kingdom 3 University of Oxford, United Kingdom The functional neuroanatomy of human value-based decision has been widely investigated using functional MRI (fMRI). A ventromedial portion of prefrontal cortex (VMPFC) has been isolated as especially important, although its precise contribution remains debated. One view holds that VMPFC codes stimulus values (1) that are then compared elsewhere in the brain (2). Another suggests VMPFC codes the value of the option that will be become chosen (3), perhaps to enable subsequent computation of a prediction error (4). In yet other studies, VMPFC signals the difference in value between chosen and unchosen options specifically at choice, not outcome, time (5,6) - arguing for a central role for VMPFC in the formation of a decision. This would suggest that an initial representation of stimulus value in VMPFC may evolve into chosen and unchosen value signals as an option is selected. However, the temporal evolution of this signal may occur too quickly to be resolved using fMRI. To test this hypothesis, we measured simultaneous magnetoencephalography and electroencephalography (M/EEG) in 19 subjects as they performed a simple economic decision task. Subjects weighed the probability of receiving monetary reward on two options with the magnitude of reward that could be attained. Subject choices matched predictions from descriptive economic models of behaviour, and reaction times correlated negatively with the difference in value between the two options. We used trial-by-trial regression to investigate the effects of several variables on M/EEG data at different timepoints through the trial. At the time decision stimuli were presented, and at the time of button-press, we looked for the effect of ’value difference’ (chosen-unchosen values) but also of overall ’stimulus value’ (chosen+unchosen values). We employed the multiple sparse priors (MSP) source reconstruction algorithm implemented in SPM8 to investigate the neuroanatomical basis of these signals. Aligned to the presentation of the decision, ’value difference’ signals localised to VMPFC, peaking approximately 800ms after stimulus onset. Crucially, however, this signal was immediately preceded by a ’stimulus value’ signal, also localised to VMPFC, representing the sum of all options available to the subject (500-800ms). Aligning instead to the time of subject response, ’value difference’ signals were found preceding the button-press, approximately 300ms before the action was executed. In contrast to the stimulus-aligned signals, these localised to SMA/cingulate motor area and primary motor cortex, but were not preceded by a ’stimulus value’ signal, and so might relate to motor preparation rather than the decision process per se. These results suggest that as a decision is made, the signal within VMPFC evolves from representing the stimulus value of all available options to representing the decision that has been made. The role of VMPFC may therefore go beyond the signalling of the chosen option for subsequent computation of a reward prediction error, and may be fundamental to the process of selecting an option during value-based decision.

Optimal classifier selection and negative bias in error rate estimation: an empirical study on high-dimensional prediction.

BackgroundIn biometric practice, researchers often apply a large number of different methods in a "trial-and-error" strategy to get as much as possible out of their data and, due to publication pressure or pressure from the consulting customer, present only the most favorable results. This strategy may induce a substantial optimistic bias in prediction error estimation, which is quantitatively assessed in the present manuscript. The focus of our work is on class prediction based on high-dimensional data (e.g. microarray data), since such analyses are particularly exposed to this kind of bias.MethodsIn our study we consider a total of 124 variants of classifiers (possibly including variable selection or tuning steps) within a cross-validation evaluation scheme. The classifiers are applied to original and modified real microarray data sets, some of which are obtained by randomly permuting the class labels to mimic non-informative predictors while preserving their correlation structure.ResultsWe assess the minimal misclassification rate over the different variants of classifiers in order to quantify the bias arising when the optimal classifier is selected a posteriori in a data-driven manner. The bias resulting from the parameter tuning (including gene selection parameters as a special case) and the bias resulting from the choice of the classification method are examined both separately and jointly.ConclusionsThe median minimal error rate over the investigated classifiers was as low as 31% and 41% based on permuted uninformative predictors from studies on colon cancer and prostate cancer, respectively. We conclude that the strategy to present only the optimal result is not acceptable because it yields a substantial bias in error rate estimation, and suggest alternative approaches for properly reporting classification accuracy.

Flexible Regression Model Selection for Survival Probabilities: With Application to AIDS

Clinicians are often interested in the effect of covariates on survival probabilities at prespecified study times. Because different factors can be associated with the risk of short- and long-term failure, a flexible modeling strategy is pursued. Given a set of multiple candidate working models, an objective methodology is proposed that aims to construct consistent and asymptotically normal estimators of regression coefficients and average prediction error for each working model, that are free from the nuisance censoring variable. It requires the conditional distribution of censoring given covariates to be modeled. The model selection strategy uses stepup or stepdown multiple hypothesis testing procedures that control either the proportion of false positives or generalized familywise error rate when comparing models based on estimates of average prediction error. The context can actually be cast as a missing data problem, where augmented inverse probability weighted complete case estimators of regression coefficients and prediction error can be used (Tsiatis, 2006, Semiparametric Theory and Missing Data). A simulation study and an interesting analysis of a recent AIDS trial are provided.

Stochastic characterization of the Montalto Uffugo research site (Italy) by geostatistical inversion of moment equations of groundwater flow

Summary We assess the applicability and performance of a methodology of inverting stochastic mean groundwater flow equations to characterize the spatial variability of (natural) log-transmissivity (Y) of a heterogeneous aquifer. The methodology, originally proposed by Hernandez et al., 2003 , Hernandez et al., 2006 , relies on a nonlinear geostatistical inverse algorithm for recursive approximations of steady-state (ensemble) mean groundwater flow that allows estimating jointly the spatial variability of Y, the underlying variogram parameters, and the variance–covariance of the estimates. Estimates of prediction errors of hydraulic heads and fluxes are then calculated a posteriori, upon solving equations satisfied by the corresponding covariances. Here, we extend the methodology to quasi-steady state flow conditions and present its first field application by using information collected during a pumping test performed at the Montalto Uffugo research site (Italy). Log-transmissivity is parameterized geostatistically on the basis of an available measured value and a set of unknown values at discrete pilot points. Best estimates of Y at the measurement location and at the pilot points are obtained by a maximum likelihood fit of computed and measured heads. These posterior estimates are then projected onto the computational grid by kriging. Information on head drawdowns is provided through self-potential signals recorded by 47 surface electrodes during the test. The maximum likelihood-based objective function includes a regularization term reflecting prior information about Y. The relative weight assigned to this term is evaluated separately from other model parameters to avoid bias and instability. We explore the effectiveness of both a zero- and a second-order closure of the mean flow equation at providing a proper geostatistical characterization of the log-transmissivity field. The parameters of the variogram of Y are estimated a posteriori using formal model selection criteria. The adoption of a second-order mean flow model renders the sharpest definition of the regularization term and of the Y variogram parameters.

Rapid on-site evaluation of transbronchial aspirates: randomised comparison of two methods

The value of different staining methods for rapid analysis of transbronchial needle aspirates during bronchoscopy has not been explored. In the present study, we compared a Papanicolaou-based rapid stain, prepared by a technologist and read by a cytopathologist, and a Wright-Giemsa-based rapid stain, prepared and read by a cytopathologist alone. Gold standard was the final laboratory report issued on each aspirate. We harvested 827 aspirates from 218 target sites in 126 consecutive patients. At least one positive aspirate was found in 99 (79%) patients. In those 99 patients, 288 of 574 (50%) aspirates were positive for neoplastic (83%) or non-neoplastic (17%) disease. False-negative aspirates and target sites were more frequent with the rapid Wright-Giemsa than with the rapid Papanicolaou stain (14.2 versus 7.3%, p = 0.008, and 13.7 versus 3.6%, p = 0.021, respectively). The sensitivity of the Wright-Giemsa-based and Papanicolaou-based rapid stains for detecting diagnostic material was 93 and 100% in patients, 83.1 and 95.5% in target sites, and 72.8 and 84.9% in aspirates, respectively. Specificity was 100% for both methods in patients and target sites, and 90.4 and 95% in aspirates. We concluded that a Papanicolaou-based stain has superior yield and accuracy to a Wright-Giemsa-based stain for rapid on-site evaluation of transbronchial needle aspirates.

Estimation of prediction error for survival models

When statistical models are used to predict the values of unobserved random variables, loss functions are often used to quantify the accuracy of a prediction. The expected loss over some specified set of occasions is called the prediction error. This paper considers the estimation of prediction error when regression models are used to predict survival times and discusses the use of these estimates. Extending the previous work, we consider both point and confidence interval estimations of prediction error, and allow for variable selection and model misspecification. Different estimators are compared in a simulation study for an absolute relative error loss function, and results indicate that cross-validation procedures typically produce reliable point estimates and confidence intervals, whereas model-based estimates are sensitive to model misspecification. Links between performance measures for point predictors and for predictive distributions of survival times are also discussed. The methodology is illustrated in a medical setting involving survival after treatment for disease.

Genetic variation in dopaminergic neuromodulation influences the ability to rapidly and flexibly adapt decisions

The ability to rapidly and flexibly adapt decisions to available rewards is crucial for survival in dynamic environments. Reward-based decisions are guided by reward expectations that are updated based on prediction errors, and processing of these errors involves dopaminergic neuromodulation in the striatum. To test the hypothesis that the COMT gene Val(158)Met polymorphism leads to interindividual differences in reward-based learning, we used the neuromodulatory role of dopamine in signaling prediction errors. We show a behavioral advantage for the phylogenetically ancestral Val/Val genotype in an instrumental reversal learning task that requires rapid and flexible adaptation of decisions to changing reward contingencies in a dynamic environment. Implementing a reinforcement learning model with a dynamic learning rate to estimate prediction error and learning rate for each trial, we discovered that a higher and more flexible learning rate underlies the advantage of the Val/Val genotype. Model-based fMRI analysis revealed that greater and more differentiated striatal fMRI responses to prediction errors reflect this advantage on the neurobiological level. Learning rate-dependent changes in effective connectivity between the striatum and prefrontal cortex were greater in the Val/Val than Met/Met genotype, suggesting that the advantage results from a downstream effect of the prefrontal cortex that is presumably mediated by differences in dopamine metabolism. These results show a critical role of dopamine in processing the weight a particular prediction error has on the expectation updating for the next decision, thereby providing important insights into neurobiological mechanisms underlying the ability to rapidly and flexibly adapt decisions to changing reward contingencies.

Friction-induced vibration: Quantifying sensitivity and uncertainty

The limited success of predictive models of friction-induced vibration can, in part, be attributed to the inherent sensitivity of friction-coupled systems to variations, often uncontrolled, in parameter values such as the friction coefficient. This paper explores the sensitivity and uncertainty of predictions from a modal point of view, using models of a realistic complexity. A method for efficiently estimating prediction error bounds is presented and validated using representative parametric uncertainties. Measurement uncertainties are quantified providing an input for the error-bound analysis. Taken together, this forms the foundation for a direct comparison of predictions with experimental results from sliding contact tests.

Estimation of prediction error by using K-fold cross-validation

Estimation of prediction accuracy is important when our aim is prediction. The training error is an easy estimate of prediction error, but it has a downward bias. On the other hand, K-fold cross-va...

Editorial [Hot topic: Genomic Signal Processing: Part 1 (Guest Editors: E.R. Dougherty, X. Cai, Y. Huang, S. Kim and R. Yamaguchi)

Genomic Signal Processing (GSP) has been defined as the analysis, processing, and use of genomic signals for gaining biological knowledge and the translation of that knowledge into systems-based applications, where by genomic signals we mean the measurable events, principally the production of mRNA and protein carried out within the cell. Owing to the defining role of DNA in the production of mRNA, the structural characterization of DNA is inevitably a part of GSP and, interestingly, signal processing methods are utilized in understanding DNA structure. A key goal of translational genomics is to discover families of genes or gene products that can be used to classify disease, thereby leading to molecular-based diagnosis and prognosis. A deeper goal is to characterize genomic and proteomic regulation, thereby leading to a functional understanding of disease and the development of systems-based medical solutions. GSP is growing in importance as an ever larger community is recognizing that accomplishing these goals requires various disciplines within or related to signal processing, including pattern recognition, prediction/estimation theory, information theory, dynamical systems, control theory, network modeling, and communication theory. In sum, systems biology and systems medicine demand deep understanding of systems theory. This inevitably entails the theory and methods of signal processing, which have been so successful in areas such as communications, and the related theory pertaining to the characterization and control of dynamical systems, without which one cannot even imagine our contemporary technological society. The purpose of this special issue is to bring some of the key developments in GSP to the wider genomics community. Owing to its grounding in systems theory and stochastic processes, GSP often requires mathematics beyond the level of that studied in undergraduate electrical engineering, or even undergraduate mathematics and statistics, and, therefore, as originally published in the scientific literature, is not accessible to many researchers in biology and medicine. Because systems biology and systems medicine will, ipso facto, have to rely on mathematical systems theory, this dichotomy is a problem that will have to be addressed in the future, from both educational and research perspectives; nonetheless, in a review format it is possible to communicate many of the basic ideas without recourse to the kind of full rigorous mathematical analyses required in original research. In this issue, the guest editors have tried to accomplish this aim in two ways: first, by specifying the kinds of issues we would like contributors to address, with the requirement that the mathematical details be kept to a minimum (with references to the relevant literature); and, second, by working with the contributors to achieve the goals of the issue through the review process. Although the final scope has been determined by the submissions and acceptances, we believe that a good breadth of subject matter is included in the issue, including the inference, analysis, and control of gene regulatory networks, mass spectrometry for proteomics, sequence analysis, metagenomics, MicroRNA target prediction, clustering, and feature selection and error estimation for classification, where in the cases of classification and clustering the papers pay close attention to performance in the context of small samples, a ubiquitous situation in which many methods in the literature perform poorly.

Construction of a five-marker protein-based model for stage-independent assessment of prognosis in breast cancer

11013 Background: While the TNM method for assessment of stage in breast cancer is simple and robust, molecular methods for limited patient subsets are gaining popularity (i.e Oncotype Dx or Mammaprint). We hypothesized that multiplexed quantitative measurement of proteins known to be involved in breast cancer signaling pathways of growth, proliferation, survival, and metastasis can enhance clinical methods of predicting prognosis in all patients. Methods: We assessed the expression of twenty-three proteins (ER, PR, EGFR, HER2, HER3, HER4, ERK, PTEN, PI3Kp85α, PI3Kp110α, p27/Kip1,EIF4E, FOXO3,AKT1, AKT2, AKT3, MYC, cyclinD1, FOXO1, mTOR, p70S6Kb, NFkB and BCL2) in four subcellular compartments by automated quantitative analysis of protein expression (AQUA) on tissue microarrays of the archival Yale breast cancer cohort (n=676). To project future performance of these markers including clinicopathological parameters, we constructed univariate and multivariate logistic regression models using leave-one-out cross-validation and calculated prediction error (PE) estimates of each model's value to predict a binary endpoint of 10 year survival. In addition, we constructed univariate and multivariate Cox models of ten year disease specific survival (DSS). Results: By Cox univariate analysis, ER, PR, PTEN, and BCL2 were directly correlated with DSS, while FOXO1, HER2, HER3, and PI3Kp110α were inversely correlated with DSS. A five-variable logistic regression model of 10 year survival including nuclear AKT1, BCL2, nuclear FOXO1, cytoplasmic mTOR, and nuclear p70S6Kb (prediction error= .274) surpasses performance of TNM staging (PE=.367) and the Nottingham Prognostic Index (PE=.326). The same model is associated with 10-year DSS by Cox proportional hazard (p=<.00001) independent of TNM stage and NPI. Conclusions: Our protein-based, multiplexed approach to prognostic classification was superior to traditional methods (TNM or NPI) and single biomarkers in this retrospective cohort. Current outcomes are influenced by modern therapies, limiting the direct impact of this analysis. However, molecular profiling of primary tumor linked to outcome paves the way for the incorporation of new prognostic models into prospective studies. [Table: see text]

Stimulus–outcome learnability differentially activates anterior cingulate and hippocampus at feedback processing

Memory systems are known to be influenced by feedback and error processing, but it is not well known what aspects of outcome contingencies are related to different memory systems. Here we use the Rescorla-Wagner model to estimate prediction errors in an fMRI study of stimulus-outcome association learning. The conditional probabilities of outcomes for a given stimulus are manipulated so that associations are either learnable or unlearnable (pseudorandom). The delay between stimulus and outcome is jittered so that we can separately compare activity for either stimulus processing or feedback processing. We find that hippocampus and anterior cingulate are differentially active primarily at feedback processing: Learnable associations are correlated with significantly more hippocampal activity and significantly less anterior cingulate activity than unlearnable associations. We also find that positive prediction errors modulate feedback processing in the midbrain for both types of associations. We suggest that learnable associations use more declarative memory, unlearnable associations involve more uncertainty monitoring, and, in both kinds of associations, positive prediction errors provide a reinforcement signal.

An alternative deterministic method for the spatial interpolation of water retention parameters

Abstract. There are two approaches available for mapping water retention parameters over the study area using a spatial interpolation method. (1) Retention models can be first fitted to retention curves available at sampling locations prior to interpolating model parameters over the study area (the FI approach). (2) Retention data points can first be interpolated over the study area before retention model parameters are fitted (the IF approach). The current study compares the performance of these two approaches in representing the spatial distribution of water retention curves. Standard geostatistical interpolation methods, i.e., ordinary kriging and indicator kriging, were used. The data used in this study were obtained from the Las Cruces trench site database, which contains water retention data for 448 soil samples. Three standard water retention models, i.e., Brooks and Corey (BC), van Genuchten (VG), and Kosugi (KSG), were considered. For each model, standard validation procedures, i.e., leave-one-out cross-validation and split-sample methods were used to estimate the uncertainty of the parameters at each sampling location, allowing for the computation of prediction errors (mean absolute error and mean error). The results show that the IF approach significantly lowered mean absolute errors for the VG model, while also reducing them moderately for the KSG and BC models. In addition, the IF approach resulted in less bias than the FI approach, except when the BC model was used in the split-sample approach. Overall, IF outperforms FI for all three retention models in describing the spatial distribution of retention parameters.

Boosting for high-dimensional time-to-event data with competing risks

For analyzing high-dimensional time-to-event data with competing risks, tailored modeling techniques are required that consider the event of interest and the competing events at the same time, while also dealing with censoring. For low-dimensional settings, proportional hazards models for the subdistribution hazard have been proposed, but an adaptation for high-dimensional settings is missing. In addition, tools for judging the prediction performance of fitted models have to be provided. We propose a boosting approach for fitting proportional subdistribution hazards models for high-dimensional data, that can e.g. incorporate a large number of microarray features, while also taking clinical covariates into account. Prediction performance is evaluated using bootstrap.632+ estimates of prediction error curves, adapted for the competing risks setting. This is illustrated with bladder cancer microarray data, where simultaneous consideration of both, the event of interest and competing events, allows for judging the additional predictive power gained from incorporating microarray measurements. The proposed boosting approach is implemented in the R package CoxBoost and prediction error estimation in the package peperr, both available from CRAN.

Shark depredation rates in pelagic longline fisheries: a case study from the Northwest Atlantic

Abstract MacNeil, M. A., Carlson, J. K., and Beerkircher, L. R. 2009. Shark depredation rates in pelagic longline fisheries: a case study from the Northwest Atlantic. – ICES Journal of Marine Science, 66: 708–719. A suite of modelling approaches was employed to analyse shark depredation rates from the US Atlantic pelagic longline fishery. As depredation events are relatively rare, there are a large number of zeroes in pelagic longline data and conventional generalized linear models (GLMs) may be ineffective as tools for statistical inference. GLMs (Poisson and negative binomial), two-part (delta-lognormal and truncated negative binomial, T-NB), and mixture models (zero-inflated Poisson, ZIP, and zero-inflated negative binomial, ZINB) were used to understand the factors that contributed most to the occurrence of depredation events that included a small proportion of whale damage. Of the six distribution forms used, only the ZIP and T-NB models performed adequately in describing depredation data, and the T-NB and ZINB models outperformed the ZIP models in bootstrap cross-validation estimates of prediction error. Candidate T-NB and ZINB model results showed that encounter probabilities were more strongly related to large-scale covariates (space, season) and that depredation counts were correlated with small-scale characteristics of the fishery (temperature, catch composition). Moreover, there was little evidence of historical trends in depredation rates. The results show that the factors contributing to most depredation events are those already controlled by ships' captains and, beyond novel technologies to repel sharks, there may be little more to do to reduce depredation loss in the fishery within current economic and operational constraints.

Repeated double cross validation

AbstractRepeated double cross validation (rdCV) is a strategy for (a) optimizing the complexity of regression models and (b) for a realistic estimation of prediction errors when the model is applied to new cases (that are within the population of the data used). This strategy is suited for small data sets and is a complementary method to bootstrap methods. rdCV is a formal, partly new combination of known procedures and methods, and has been implemented in a function for the programming environment R, providing several types of plots for model evaluation. The current version of the software is dedicated to regression models obtained by partial least‐squares (PLS). The applied methods for repeated splits of the data into test sets and calibration sets, as well as for estimation of the optimum number of PLS components, are described. The relevance of some parameters (number of segments in CV, number of repetitions) is investigated. rdCV is applied to two data sets from chemistry: (1) determination of glucose concentrations from near infrared (NIR) data in mash samples from bioethanol production; (2) modeling the gas chromatographic retention indices of polycyclic aromatic compounds from molecular descriptors. Models using all original variables and models using a small subset of the variables, selected by a genetic algorithm (GA), are compared by rdCV. Copyright © 2009 John Wiley & Sons, Ltd.

Parallelized prediction error estimation for evaluation of high-dimensional models

There is a multitude of new techniques that promise to extract predictive information in bioinformatics applications. It has been recognized that a first step for validation of the resulting model fits should rely on proper use of resampling techniques. However, this advice is frequently not followed, potential reasons being difficulty of correct implementation and computational demand. This is addressed by the R package peperr, which is designed for reliable prediction error estimation through resampling, potentially accelerated by parallel execution on a compute cluster. Its interface allows easy connection to newly developed model fitting routines. Performance evaluation of the latter is furthermore guided by diagnostic plots, which helps to detect specific problems due to high-dimensional data structures. http://cran.r-project.org, http://www.imbi.uni-freiburg.de/parallel. Supplementary data are available at Bioinformatics online.

Theoretical characterization of gas–liquid chromatographic stationary phases with quantum chemical descriptors

Quantitative structure–property relationship (QSPR) solvent model has been developed for the McReynolds constants (prototypical solutes) on 36 gas–liquid chromatographic stationary phases. PM6 semiempirical quantum chemical calculations combined with conductor-like screening model (COSMO) has been utilized. From 276 descriptors considered, forward stepwise variable selection, followed by best subset selection, yielded linear regression models containing six purely quantum chemical and two hybrid, topologically based descriptors. Internal (leave-one-out and bootstrap) as well as external validation methods confirmed the predictive power of these structure-driven models across all 10 McReynolds constants, with 40 Kováts-index units overall root-mean-square prediction error estimate.

Handling Certain Structure Information in Subspace Identification

The prediction-error approach to parameter estimation of linear models often involves solving a non-convex optimization problem. In some cases, it is therefore difficult to guarantee that the global optimum will be found. A common way to handle this problem is to find an initial estimate, hopefully lying in the region of attraction of the global optimum, using some other method. The prediction-error estimate can then be obtained by a local search starting at the initial estimate. In this paper, a new approach for finding an initial estimate of certain linear models utilizing structure and the subspace method is presented. The polynomial models are first written on the observer canonical state-space form, where the specific structure is later utilized, rendering least-squares estimation problems with linear equality constraints.

Aggregate Economy Risk and Company Failure: An Examination of UK Quoted Firms

Considerable attention has been directed in the recent finance and economics literature to issues concerning the effects on company failure risk of changes in the macroeconomic environment. This paper examines the accounting ratio-based and macroeconomic determinants of insolvency exit of UK large industrials during the early 1990s with a view to improve understanding of company failure risk. Failure determinants are revealed from estimates based on a cross-section of 369 quoted firms, which is followed by an assessment of predictive performance based on a series of time-to-failure-specific logit functions, as is typical in the literature. Within the traditional for cross-sectional data studies framework, a more complete model of failure risk is developed by adding to a set of traditional financial statement-based inputs, the two variables capturing aggregate economy risk - one-year lagged, unanticipated changes in the nominal interest rate and in the real exchange rate. Alternative estimates of prediction error are obtained, first, by analytically adjusting the apparent error rate for the downward bias and, second, by generating holdout predictions. More complete, augmented with the two macroeconomic variables models demonstrate improved out-ofestimation-sample classificatory accuracy at risk horizons ranging from one to four years prior to failure, with the results being quite robust across a wide range of cut-off probability values, for both failing and non-failed firms. Although in terms of the individual ratio significance and overall predictive accuracy, the findings of the present study may not be directly comparable with the evidence from prior research due to differing data sets and model specifications, the results are intuitively appealing. First, the results affirm the important explanatory role of liquidity, gearing, and profitability in the company failure process. Second, the findings for the failure probability appear to demonstrate that shocks from unanticipated changes in interest and exchange rates may matter as much as the underlying changes in firm-specific characteristics of liquidity, gearing, and profitability. Obtained empirical determinants suggest that during the 1990s recession, shifts in the real exchange rate and rises in the nominal interest rate, were associated with a higher propensity of industrial company to exit via insolvency, thus indicating links to a loss in competitiveness and to the effects of high gearing. The results provide policy implications for reducing the company sector vulnerability to financial distress and failure while highlighting that changes in macroeconomic conditions should be an important ingredient of possible extensions of company failure prediction models.