Taxi time predictions are used by air traffic controllers to optimally release aircraft from the gate such that efficiency losses due to queuing are minimized, while runway capacity is maintained. More accurate taxi times can therefore result in improved airport surface operations and reduce air traffic controller workload. This article proposes a methodology to develop taxi time predictor and applies this methodology to Schiphol airport. The methodology combines novel data-driven predictors with different improvements and extensive performance evaluation. One such improvement involves using recent taxi time prediction errors to improve upcoming taxi time predictions. During evaluation, this article extends conventional analysis by analyzing different prediction horizons and performance metrics. Applying the methodology at Schiphol airport resulted in a predictor that increased the fraction of flights with a taxi time error of less than two minutes from 64.41% to 67.91% compared to the currently operational manual decision tree predictor.