Table A4. Comparative overview of... | Emerald Publishing

Table A4.

Comparative overview of classification methods

Method	Basic idea	Key assumpitons	Strengths	Limitations
Random forest	Ensemble of bootstrapped trees with feature randomness; votes aggregated	Weakly correlated, moderately strong trees; enough trees for stability	Strong out-of-box accuracy; robust to noise and outliers; little tuning; variable importance available	Less interpretable than single tree; slower with many trees; probability calibration sometimes needed
Decision tree	Greedy recursive splits to reduce impurity	None on distribution/scale; assumes meaningful splits exist	Interpretable rules; handles nonlinearity and interactions; invariant to monotone rescaling	Unstable; prone to overfit; lower accuracy vs ensembles
Gradient boosting	Sequentially adds small trees to fit residuals (boosts weak learners)	Additive tree model; learning rate and depth govern bias–variance tradeoff	State-of-the-art tabular accuracy; captures subtle interactions; flexible	More tuning sensitive; can overfit without early stopping; slower than RF
Support vector machine	Finds a maximum-margin boundary (linear or kernelized)	Margin separability in transformed space; appropriate kernel choice; scaled features	Strong on high-dimensional data; handles complex boundaries with kernels	Harder to tune; no native probabilities (needs calibration); slower on very large N; scaling required
Logistic regression	Fits a logistic link between features and the probability of bankruptcy	Linear log-odds; additivity; limited multicollinearity; well-specified features	Simple, fast, well-understood; baseline for odds ratios; calibrated probabilities	Misses nonlinearity and interactions unless engineered; sensitive to scaling/collinearity; underfits complex patterns

Method	Basic idea	Key assumpitons	Strengths	Limitations
Random forest	Ensemble of bootstrapped trees with feature randomness; votes aggregated	Weakly correlated, moderately strong trees; enough trees for stability	Strong out-of-box accuracy; robust to noise and outliers; little tuning; variable importance available	Less interpretable than single tree; slower with many trees; probability calibration sometimes needed
Decision tree	Greedy recursive splits to reduce impurity	None on distribution/scale; assumes meaningful splits exist	Interpretable rules; handles nonlinearity and interactions; invariant to monotone rescaling	Unstable; prone to overfit; lower accuracy vs ensembles
Gradient boosting	Sequentially adds small trees to fit residuals (boosts weak learners)	Additive tree model; learning rate and depth govern bias–variance tradeoff	State-of-the-art tabular accuracy; captures subtle interactions; flexible	More tuning sensitive; can overfit without early stopping; slower than RF
Support vector machine	Finds a maximum-margin boundary (linear or kernelized)	Margin separability in transformed space; appropriate kernel choice; scaled features	Strong on high-dimensional data; handles complex boundaries with kernels	Harder to tune; no native probabilities (needs calibration); slower on very large N; scaling required
Logistic regression	Fits a logistic link between features and the probability of bankruptcy	Linear log-odds; additivity; limited multicollinearity; well-specified features	Simple, fast, well-understood; baseline for odds ratios; calibrated probabilities	Misses nonlinearity and interactions unless engineered; sensitive to scaling/collinearity; underfits complex patterns