Obsolescence is highly complex problem due to the influence of many factors such as competitive market pressure, technological advancement and short life cycle of technological components. Basically, obsolescence problems are often sudden and not planned, causing delays and extra cost. To overcome this problem, forecasting appears to be one of the most efficient solutions. Indeed, many studies have been conducted to create models that can effectively forecast obsolescence. In addition, applying machine learning techniques have attracted many attentions and have been widely used to predict obsolescence risk and life cycle. Popular algorithms such as random forest, has been reporting satisfactory performance. To improve the accuracy of machine learning algorithms for obsolescence forecasting, this paper proposes a new optimization approach for obsolescence forecasting based on random forest (RF) and Particle Swarm Optimization (PSO). In fact, parameters optimization and features selection of RF have an important effect on it is predictive accuracy and PSO presents one kind of effective method for RF parameters and features choosing. To examine the effectiveness of this approach, this paper presents a comparison between PSO-RF and GA-RF (random forest based on genetic algorithm). Experimental results show that PSO-RF outperformed GA-RF with 96% of accuracy.