Reward-based crowdfunding is a form of innovative financing that allows project creators to raise funds from potential backers to start their ventures. A crowdfunding project is successfully funded if and only if the predetermined funding goal is achieved within a given time. We consider a model where backers arrive sequentially at a crowdfunding project. Upon arrival, a backer makes her pledging decision by taking into account the expected success of the project. We characterize the dynamics of a project's pledging process. In particular, we show that there exists a "cascade effect" on backers' pledging, which is mainly driven by the all-or-nothing nature of crowdfunding projects. According to our data collected from the most popular online crowdfunding platform, Kickstarter, the majority of projects fail to achieve their goals. To address this issue, we propose three contingent stimulus policies, namely, seeding, feature upgrade and limited-time offer. We show that the optimal stimulus policies have a cutoff-time structure. Then we propose simple heuristics derived from the deterministic counterpart of the stochastic model and show that they are asymptotically optimal when the problem is scaled up. However, for limited-time offer, we show that profit loss from the heuristic has a magnitude with an order higher than the square root of the scale parameter, which is the typical order of magnitude in loss from deterministic heuristics in revenue management. This result underscores the importance of contingent policies in crowdfunding. Lastly, we show that the benefit of contingent policies is greatest in the middle of crowdfunding campaigns. Testing with the data set of Kickstarter, we obtain empirical evidence that the projects' success rates improve by 14.6% on average with updates in the middle of the campaign and when the pledging progress is lagging.