Our systematic large-scale comparison study performed using 243 real datasets on different prediction tasks shows the good average prediction performance of random forest (compared to logistic regression) even with the standard implementation and default parameters, which are in some respects suboptimal. This study should in our view be seen both as (i) an illustration of the application of principles borrowed from clinical trial methodology to benchmarking in computational sciences—an approach that could be more widely adopted in this field and (ii) a motivation to pursue research (and comparison studies!) on random forests, not only on possibly better variants and parameter choices but also on strategies to improve their transportability.
