Predicting discrete traits in evolving systems

Discrete data are prevalent across biology, from nucleotide base pairs to mating systems or petal count on a flower. Predictions using discrete data are central to imputation, forecasting, and hypothesis testing in biology. How much accuracy and confidence they deliver together, and how this varies across scenarios and taxa, has not been quantified systematically. Here, using accuracy and log loss, we benchmark these models across scenarios varying in tree size, evolutionary rate, and trait coevolution. Against non-phylogenetic baselines, they increase accuracy by 23 percentage points and lower log loss by 92% on average; different Markov implementations perform similarly. Log loss rises as terminal branch lengths elongate, a direct consequence of the Markov process. Last, we provide a tutorial and empirical demonstration for the evolution of unexpected character states (outliers). Our study provides a framework for imputing missing data, reconstructing ancestral traits, and testing evolutionary hypotheses.