Travel time prediction is important for freight transportation companies. Accurate travel time prediction can help these companies make better planning and task scheduling. For several reasons, most companies are not able to obtain traffic flow data from traffic management authorities, but a large amount of trajectory data were collected everyday which has not been fully utilised. In this study, we aim to fill this gap and performed travel time predictions for freight vehicles at individual level using Gradient Boosting Regression Tree (GBRT) models. All the features were extracted or composed from vehicles' temporally sparse trajectory data. Three routes were selected for the prediction experiments. Bayesian optimisation was adopted for model fitting while the results show that both pre-start (before trip starts) and post-start (after trip starts) predictions accuracies reach above 80%. The results also show that the prediction performance can be gradually improved by adding more mean speed estimates of traveled distance from the first 5 minutes as the real-time information. And the prediction performance can be further improved by about 2% by adding more mean speed estimates even if an unusual and non-recurring events occurred at a location of a route segment. This study shows the feasibility of both pre-start and continuous post-start prediction with limited amount of temporally sparse trajectory data for real-world practice.