As research endeavours and commercial applications demand more of the quadrotor, it is only natural to develop models which can facilitate this. Currently, analytical descriptions of the quadrotor are rudimentary and most data-driven quadrotor models are identified from flight data collected indoors. Therefore, existing quadrotor models are constrained to a narrow region of the flight envelope due to the inherent restrictions of these indoor spaces. In contrast, the flexibility afforded by outdoor spaces facilitates the exploration of the majority of the quadrotor's flight envelope. Subsequently, to enable the development of high-fidelity quadrotor models, a modular system identification pipeline that is compatible with outdoor high-speed flight and aggressive manoeuvring is created in the present work. Drawing inspiration from current state-of-the-art quadrotor models, polynomial step-wise regression, artificial neural networks (ANNs), and a hybrid approach fusing the merits of the two techniques are implemented in the pipeline. As a proxy for reliability, the confidence of the identified models' predictions are encapsulated through accompanying prediction intervals (PIs). As there are no analytical formulations for obtaining the PIs associated with ANN predictions, various ANN-based PI estimation techniques are also numerically validated. Using the system identification pipeline, for the first time, high-fidelity quadrotor models which accurately capture high-speed flight of up to 19 m/s and aggressive manoeuvres such as punch-outs, flips, and barrel rolls, are identified from real outdoor flight data, despite the contamination of unknown wind. Contrary to expectations, the simulations demonstrate that the polynomial model consistently produces the most feasible and useful models. The dense architecture of the employed ANNs, including those for the hybrid approach, appear to promote and propagate instabilities arising from the wind contamination. The associated PIs are found to grow in tandem with these instabilities and increasing uncertainty in the system, accentuating the utility of such PIs.