The ability to forecast machinery failure is vital to reducing maintenance costs, operation downtime and safety hazards. Recent advances in condition monitoring technologies have given rise to a number of prognostic schemes (both model-based and data-driven) that attempt to forecast machinery health by constructing health propagation models for the underlying systems. In particular, algorithms that use the data-driven approach learn models directly from the data, rather than using a hand-built model based on human expertise. This paper introduces a novel architecture for data-driven Failure Prognosis of complex non-linear systems using Least Squares Support Vector Regression Machines (LSSVR). An adaptive recurrent LSSVR machine is proposed and augmented with a Bayesian Inference scheme that allows probabilistic estimates of future health deterioration. Extensions for efficient multi-step long-term prognostics and Remaining Useful Life (RUL) calculation are suggested. Data from a seeded fault test for a UH-60 planetary gearbox plate is used to test the online performance of the prognostics algorithm.