In this work, a load-based yaw misalignment observer was tested and validated with turbine and mast data collected during a wake steering and characterization campaign. A shallow feed-forward neural network was used to map the relation between the yaw misalignment and the in-and out-of-plane blade load harmonics for a 3.5 MW machine, and its performance was analysed over about 108 full days of useful data. Confirming previous findings, this simple neural network was able to accurately estimate the yaw misalignment, with an average 10-minute absolute error of at most 4°.The performance of the yaw misalignment observer was compared to the one of the standard onboard wind vane during the wake steering campaign. Results indicate that the wind vane may significantly overestimate the misalignment for large angles, possibly on account of the wake rotation and flow distortion effects caused by the nacelle. The observer on the other hand, sampling the flow at the rotor disk and not behind it, is not affected by such phenomena and could therefore provide a more accurate measurement of the misalignment angle, possibly improving the performance of wake steering. When the turbine is already equipped with load sensors, this is obtained without the need to install and maintain extra hardware, which instead is the case with spinner-mounted anemometers or lidars.