Predictive structure dispersed across many weak, heterogeneous signals

Pervasive nonlinear interactions and threshold effects

Variance reduction through aggregation improves generalisation

Strong performance on noisy, tabular accounting data (financial ratios, governance indicators, journal-entry attributes)

Models complex interactions without manual specification

Robust to multicollinearity and outliers

Opaque ensemble structure complicating interpretability

Difficult to justify in audit documentation without post-hoc tools (such as feature importance and partial dependency plots)

Less effective for highly imbalanced label settings unless tuned carefully

Residual errors contain systematic information

Complex relationships can be approximated through sequential error correction

Combines many simple models to build a highly flexible, nonlinear prediction function

State-of-the-art performance in distress, fraud, misstatement, and going-concern prediction

Handles high-dimensional feature spaces and subtle interaction structures

Highly tuneable for different accounting tasks

Sensitive to label noise – problematic for fraud, misstatement, and GCO datasets, i.e. if the label is wrong, the residual is wrong – and the model learns from the error itself

Risk of overfitting without careful regularisation

Interpretability limited because it relies on many small trees; explanations rely on approximations

Requires explicit handling of class imbalance (weighting, resampling or focal loss) for rare-event accounting outcomes such as fraud and bankruptcy

Misclassified observations are especially informative

Outcome labels are sufficiently reliable for reweighting

Decision boundaries can be sharpened by focusing on “hard” cases

Effective when labels are observable and accurate (e.g. internal control weaknesses, restatements)

Performs well on structured, moderately complex tabular datasets where weak learners can reliably extract signals

Poor compatibility with noisy or strategically distorted labels

Amplifies misclassifications in fraud and GCO contexts

Less robust than GBM or RF for high-stakes predictions

Data are separable (linearly or via kernel transformation)

Optimal classification boundary maximises margin

Good performance on smaller, structured datasets

Effective in high-dimensional spaces when kernels are well ⁠specified

Difficult to scale to large transactional datasets

Kernel choice strongly influences performance

Limited interpretability and compatibility with audit documentation

Predictive relationships are hierarchical, distributed, and nonlinear

Representations can be learned directly from raw or minimally processed input

Large datasets needed for generalisation

State-of-the-art for textual and narrative accounting data (MD&A, earnings calls, ESG disclosures)

Captures tone, sentiment, obfuscation, and semantic context

Flexible across sequential and unstructured data types

High opacity; internal logic difficult to document or defend to regulators

Data-hungry and prone to overfitting in typical accounting sample sizes

Risk of hallucination and instability in generative tasks

Normal behaviour can be learned from majority patterns

Distance/density metrics correspond to economic similarity

Latent structure reflects meaningful categories or behaviours

Can detect anomalous journal entries, unusual transactions, and client-risk segmentation

Do not require labelled outcomes – useful where fraud or misstatements are rarely observed

Discovered patterns may not correspond to accounting constructs

High false-alarm risk; anomalies may reflect benign events

Interpretation requires substantial professional judgement

Decisions shape future states and rewards

Optimal policies can be learned through repeated interaction

Organisational/compliance objectives are representable in reward functions

Conceptually promising for adaptive forecasting, audit effort allocation, and internal control optimisation

Models complex sequential decision processes

Limited real-world adoption due to data scarcity and regulatory constraints

Reward functions difficult to specify in normative accounting contexts

Trial-and-error learning inappropriate for high-stakes environments

Language encodes latent semantic and institutional structure

Attention-based contextual embeddings capture meaning

Scale produces emergent capabilities

Good at analysing narrative disclosures, contracts, and ESG reports

Useful for summarising, extracting terms, drafting workpapers

Transforms linguistic information into quantifiable signals

Susceptible to hallucination and unverifiable reasoning

Limited audit admissibility due to lack of transparent reasoning chains

Requires strict governance for privacy, accuracy and accountability