Parkinson’s disease (PD) is associated with an increased risk of falls. For people living with PD, experiencing a fall can lead to multiple injuries and significant health decline. An accurate falls risk assessment is therefore needed for care planning and targeted fall prevention programmes. Traditional fall risk assessment relies on clinical assessment which may be subjective and time-consuming. More recent approaches utilised sensor data but so far offered only up to one year of follow-up. We tested whether a short sensor-based assessment can predict the first fall in up to five years. 104 PD patients without a history of falls were recruited and assessed using six wearable sensors. Kinematic data from a two-minute walking and standing tasks were collected and analysed using three machine learning classifiers (Random Forest, Support Vector Machines, Elastic Net). Participants were then followed up for 5 years. Models were tasked with predicting the risk of developing the first fall at different time points. Falls indices were sourced and confirmed through self-reports, clinics, telephone interviews and healthcare record surveys. Machine learning models showed high capability of predicting fall risk at various time points. Random forest achieved the best performance with classifier accuracy ranging 97.2% to 86.2% for 12 to 60 months. Sensor-based features pertaining to walking variability were the most consistent predictor of future falls. Clinically assessed motor and quality of life scores were also different between fallers and non-fallers.