Financial statement fraud is a costly problem for society. Detection models can help, but a framework to guide variable selection for such models is lacking. A novel Fraud Detection Triangle (FDT) framework is proposed specifically for this purpose. Extending the well-known Fraud Triangle, the FDT framework can facilitate improved detection models. Using Benford's law, we demonstrate the posited framework's utility in aiding variable selection via the element of surprise evoked by suspicious information latent in the data. We call for more research into variables that measure rationalisations for fraud and suspicious phenomena arising as unintended consequences of financial statement fraud.