Earnings manipulation detection: How machine learning hunts for financial mutations in African markets
Source PublicationScientific Publication
Primary AuthorsIssah, Zubeiru, Doe-Dartey et al.
"Spotting financial fraud using traditional methods is like checking a student's maths homework by only looking at the final answer. Machine learning is like a teacher who looks at the eraser marks, the time it took to complete, and the student's previous grades to figure out if they actually did the work or just copied it."
Is there a hidden elegance within the apparent chaos of biological systems? In nature, organisms constantly adapt their genetic code to survive, hiding mutations within a dense structure of DNA.
We see a strangely similar evolution in the corporate world. Companies, striving to survive and attract investors, sometimes alter their financial DNA. They hide poor performance behind clever accounting. Spotting this deception is notoriously difficult. But what if we treat financial data like a genome, searching for the hidden mutations of fraud?
A new era of earnings manipulation detection
Researchers recently analysed 80 listed firms across South Africa, Nigeria, Kenya, and Ghana. They examined a decade of data from 2015 to 2024. Their goal was simple. They wanted to see if modern computer modelling could outsmart traditional accounting checks.
For years, auditors relied on standard formulas to flag suspicious numbers. It works, but it has limits. The researchers built four competing models to test a new approach. They pitted classic logistic regression against machine learning tools like random forests and gradient boosting.
Let us take a brief philosophical detour. Why would evolution organise a genome with so much junk DNA, hiding the essential instructions deep within? It provides a buffer against fatal errors and allows for rapid adaptation. In a corporate ecosystem, complex accounting rules act much like this junk DNA. They provide a dense cover that allows companies to adapt their reported earnings without explicitly breaking the law. Finding the true financial health requires sorting the vital signals from the noise.
Sorting the signal from the noise
The study measured how well each model caught anomalies. The winner was a hybrid. A machine learning-augmented logistic regression model scored an accuracy rating of 0.745. This beat the traditional model's 0.709.
What gave the game away? The models heavily weighted a company's return on assets and the difference between its book and tax records. These features acted like glowing markers on a mutated gene.
However, the data suggests that geography matters. The models performed reasonably well in Ghana but struggled in Kenya. This implies that local regulations and institutional behaviour heavily influence how companies massage their numbers. What works in one environment might fail in another.
These findings suggest that African audit regulators and major accounting firms could benefit from adding machine learning to their toolkit. By evolving their defence strategies, they may finally keep pace with the ever-changing nature of corporate deception.