Report the Baseline: Reassessing Machine Learning on Bangladeshi Road Accident Records

Studies of machine learning on Bangladeshi road accident records routinely report accuracies above 95%, and occasionally above 99%. We show that on the national Accident Research Institute (ARI) dataset such figures can be produced without the models learning anything of substance, because the trivial baseline is rarely stated. Using 43,087 crash records from 2001-2020 we evaluate two tasks against explicit baselines. For four-class severity prediction, the best of three models reaches 73.86% accuracy against a majority-class baseline of 73.49%, a gain of 0.37 points; macro F1 improves from 0.212 to 0.276, and no model attains non-zero F1 on the Simple-injury class. For district-year accident-count regression, gradient boosting reaches R2 = 0.820, but predicting each district's historical mean - a model with no learned parameters - already reaches R2 = 0.758 and beats gradient boosting on mean absolute error. Both tasks look successful on their headline metric and are close to vacuous once the right comparison is drawn. We trace this to the data: 73.5% of recorded crashes are fatal, which reflects which accidents enter a police report rather than what happens on the road, and driver age is unrecorded in 63% of records. We argue that baseline reporting should be a minimum standard for this literature, and release all code and results.