To utilise high occupancy vehicle lanes better, high occupancy toll (HOT) lanes are introduced to counter against congestion on the urban highways. In such system, low occupancy vehicles (LOVs) are allowed to pay a toll and access to HOT lanes from general purpose (GP) lanes, so that the toll rate plays a key role in dynamically allocating LOVs over the HOT and GP lanes to improve the overall system performance. First, this study presents an improved random forest (RF) method to build the lane choice behaviour prediction model. Then, by using the 5 min historical traffic and toll data collected from Interstate 405 in the USA, the improved RF combined with cross-validation and grid search shows the highest accuracy of 88.7%, which is better than other four methods. Furthermore, a novel nested model with two levels is proposed to optimise the toll rates under different real-time traffic conditions. For the nested model, the experimental results show that the proposed dynamic pricing strategy can decrease the total delay and improve the efficiency significantly. To realise the pricing strategy, some Intelligent Transportation System technologies for HOT lane systems are described in detail and designed as the fundamental of the pricing strategy.