In everyday travel, U.S. commuters will each spend 38 h a year stuck in traffic and waste over $800 in fuel (TTI, 2015). Yet, despite this statistic, the regular commute of drivers is often predictable, leading many federal projects to aim at alleviating congestion through traveler information and intelligent transportation systems (e.g., INFLO, Queue WARN, CACC, EnableATIS, ATIS2.0). Short-term destination prediction is a developing field of research that can improve these approaches through real-traveler information, such as route, traffic incidence, and congestion levels. The short-term destination prediction problem consists of capturing vehicle Global Positioning System (GPS) traces and learning from historic locations and trajectories to predict a vehicle's destination. Drivers have predictable trip destinations that can be estimated through probabilistic modeling of past trips. To study these concepts, a database of GPS driving traces (260 participants for 70 days) was collected. To model the user's trip purpose in the prediction algorithm, a new data source was explored: point of interest (POI)/land use data. An open source land use/POI dataset is merged with the GPS dataset. The resulting database includes over 20,000 trips with travel characteristics and land use/POI data. From land use/POI data and travel patterns, trip purpose was calculated with machine learning methods. To take advantage of this data source, a new prediction model structure was developed that uses trip purpose when it is available and that falls back on traditional spatial temporal Markov models when it is not. For the first time, there is an understanding of “why” a trip is taken (not just “where” and “when”), allowing the use of “why” in the prediction model. This paper explores the baseline model followed by the inclusion of trip purpose. First, a baseline tiered time origin model was developed using the Markov Chain approach. This modelling structure allows for a short training period of current modeling techniques. The other major advantage to this structure is it allows for easy implementation of the trip purpose module. Then, a machine learning technique derived the trip purpose on 5-, 15- and 30-trip learning sets, followed by results organized by purpose, time, and origin. The machine learning technique does not require future land use data and is feasible for applicable use. This model is the first to use trip purpose to make a short-term destination prediction in pseudo real-time. Results show improved accuracy and speed over the current start-of-trip destination prediction models.