Catch Me If You Can: Improving the Scope and Accuracy of Fraud Prediction

The Cost of Fraud Prediction Errors

“…The Chakrabarty et al (2020) Adjusted Benford Score Chakrabarty et al (2020) develop a fraud prediction model that is in spirit close to the MAD-Score calculated in the previous section. However, rather than directly including the raw FSD-Score in the model, they evaluate the performance of a range of alternative models that further include several standardized variations of the raw score, which adjust the measure for systematic variation across firms, financial statement lengths, industries, and over time.…”

“…They conclude that although the model that makes use of Benford's Law data only ensures the broadest applicability, a model that incorporates both the F-Score variables and the variables based on Benford's Law has the highest accuracy, Hence, in our analyses, we evaluate the performance of the adjusted Benford Score that utilizes both the F-Score and Benford's Law variables. Specifically, we calculate the Adjusted Benford Score following the coefficients in Chakrabarty et al (2020):…”

Taking the hunch out of the crunch: A framework to improve variable selection in models to detect financial statement fraud

“…'s (2010) support vector machine‐based model as implemented by Alawadhi et al. (2023); (iii) the F‐score developed using logistic regression (Dechow et al., 2011); (iv, v) two variants that incorporate Benford's law of digit distribution (Amiram et al., 2015; Chakrabarty et al., 2022); (vi) Alawadhi et al. 's (2023) misrepresentation model; and (vii) Bao et al.…”

“…Beneish and Vorst (2022) empirically evaluated seven existing financial fraud detection models using a cost-based measure that evaluates the benefits of correctly detecting fraud relative to the cost of incorrectly alleging fraud (a false positive). The seven models included: (i) the M-score based on unweighted probit (Beneish, 1999); (ii) Cecchini et al's (2010) support vector machine-based model as implemented by Alawadhi et al (2023); (iii) the F-score developed using logistic regression (Dechow et al, 2011);(iv, v) two variants that incorporate Benford's law of digit distribution (Amiram et al, 2015;Chakrabarty et al, 2022); (vi) Alawadhi et al's (2023) misrepresentation model; and (vii) Bao et al's (2020) boosted ensemble model often referred to as a machine learning model. Beneish and Vorst (2022) justified the choice of these seven models in their second footnote.…”