Аннотация:Gradient boosted decision tree algorithms only make it possible to interpolate data. Therefore, the prediction quality degrades if one of the features, such as time, lies outside the boundaries of training data set. It is hard to eliminate these types of features because they can affect other features indirectly. Our new quick and robust extrapolation algorithm, ExtraBoost, for binary classification and regression problems can handle these dependencies and perform functional analyses, e.g., derivative analysis. In this article we describe the mathematical background of ExtraBoost, perform a running time analysis of the learning and inference stages in terms of big-O notation and evaluate the Tensorflow GPU-friendly implementation of the algorithm. Due to the lack of public binary classification data sets with indirect time dependencies, we propose a method of generating such data sets as well as an example of a data set based on data from a commercial search engine.