Deep Learning Augmented Detection of Earnings Manipulation in African Listed Firms: A Multi-Model Comparative Study

Abstract This study evaluates the effectiveness of a gradient boosting deep learning proxy, augmented with machine learning fraud risk scores and audit quality signals, in detecting earnings manipulation among African listed firms, benchmarked against traditional logistic regression and Jones-model accruals predictors. A balanced panel of 800 firm-year observations across 80 listed firms in South Africa, Nigeria, Kenya and Ghana, covering 2015 to 2024, was analysed. Four competing models were developed: a traditional logistic regression using Jones-model proxies, a machine learning augmented logistic regression, a random forest, and a gradient boosting model. Model performance was benchmarked using ten-fold cross-validation area under the ROC curve and average precision, supplemented by out-of-sample hold-out testing and country-stratified validation, and feature attribution analysis identified the dominant predictors of manipulation risk. The machine learning augmented logistic regression achieved the highest out-of-sample area under the curve of 0.745, outperforming the traditional Jones-model proxy, which achieved 0.709. The machine learning fraud risk score and book tax difference were the most important predictors in the ensemble models, with return on assets the dominant feature in gradient boosting (importance of 0.223) and the machine learning fraud risk score second (0.118). Country-stratified area under the curve values revealed substantial heterogeneity, from 0.696 in Ghana to 0.458 in Kenya, underscoring the influence of institutional context on model transferability. African audit regulators, exchange commissions and Big Four firms should consider incorporating machine learning derived fraud risk scores alongside traditional accruals models in audit planning procedures. This is among the first studies to systematically evaluate and compare machine learning augmented detection models for earnings manipulation in an African multi-country panel context, directly addressing a gap in the forensic accounting and audit quality literature on emerging market applicability.