Diabetic Macular Edema (DME) is a leading cause of vision impairment in diabetic patients, requiring early diagnosis and progression monitoring. Optical Coherence Tomography (OCT) is often employed for detecting and monitoring DME, although standard retinal fundus images have been recently shown useful for this purpose too. Ultra-wide field (UWF) fundus imaging, an extension of conventional retinal imaging, offers a comprehensive view of the retina, providing a rich source of data for computational diagnostic models, potentially enabling the training of machine learning systems with high predictive performance. However, reliable predictions for clinical use require not only accurate models but also well-calibrated ones that provide trustworthy uncertainty estimates. This paper presents a solution to the DIAMOND challenge, a competition held in conjunction with MICCAI 2024, dealing with the task of DME progression forecasting from UWF images. Since one of the goals of DIAMOND is to assess model calibration, we explore various calibration techniques such as temperature scaling, deep ensembles, test-time augmentation or margin-based label smoothing, to improve predictive confidence estimates, a combination of which was used in our final submission. As DIAMOND did not release the training data (requiring code submission instead), we perform our analysis on the equally challenging task of glaucoma detection on a large-scale dataset, and report results for DIAMOND on a hidden validation set.