Negative explained variance in random forests

The random forests algorithm is known for being relatively robust to overfitting. The reason for this is that, in random forests, many (thousands) of tree-like models are grown on bootstrapped samples of the data. Tree-like models split the data repeatedly into groups, by the predictor variable and value that lead to the most homogenous post-split groups. Random forests further de-correlates the tree-type models by allowing each tree to choose only from a small sub-selection of predictors at each split.