Fraudulent transactions affect the different entities involved in the payment pipeline: (i) the merchant/vendor at which the transaction is performed, (ii) the authorizing bank, (iii) the card-holder, and (iv) the payment processing network/gateway. While fraud transactions result in the loss of billions of dollars globally, they may also result in reputational damage to the involved parties. Detecting fraudulent transactions is thus of utmost importance for the business and it also enhances customer experience in the financial domain. Predicting fraud transactions involves identifying suspicious transactions at the time of authorization and raising an alert to the decision-making authority. This research proposes a novel Event-aware Multi-component (EMl) loss for fraud prediction which incorporates key fraud-specific characteristics for learning a robust and accurate fraud prediction model. Specifically, the proposed loss incorporates key domain characteristics of fraud modeling such as focusing more on recent transactions, optimizing for an ideal event (fraud) rate, and maximizing the net benefit (or fraud savings) seen by the fraud prediction model. Further, the proposed loss is agnostic to the model architecture and can be utilized with different backbone architectures. Experimental results and analysis on multiple datasets demonstrate the efficacy of the proposed loss, where it achieves improved detection performance while optimizing for the above-mentioned industry requirements.