Monitoring the motor fluctuations and the severity of symptoms over time in Parkinson's disease (PD) patients is crucial for quantifying the progression of the disease and the adjustment of personalized therapy. The widespread availability of wearable sensors enables remote tracking of patients and the development of digital biomarkers for motor-related symptoms derived from the kinematic data acquired from these devices. Despite the progress in remote monitoring of PD symptoms, most research has been conducted on controlled behavior in the clinic, which departs considerably from individual patients' everyday behaviors and daily routines. This manuscript describes our top-performing algorithm in the Biomarker & Endpoint Assessment to Track Parkinson’s Disease DREAM Challenge, funded by the MJFF, for predicting self-labeled PD symptom severity from free-behavior sensor data. To account for the self-labeled nature of the dataset and to capture each patient's subjective perception, we applied personalized automatic prediction algorithms consisting of ensembles of multiple random forest models followed by a predictability assessment of each patient. The results highlight the gradual approach required to develop new solutions in this field and constitute an important step forward in generating automatic and semi-automatic techniques that can facilitate the treatment of PD patients.