Under the coverage of legitimate commerce, criminals money-launder their illicit incomes through the payment gateways provided by Payment Service Providers (PSP). In order to do money-laundering forensics in transactions of PSP companies, a new method was proposed by Hojati et al which is done through detecting deviations from class behavior based on peer group analysis (PGA) method. Our experiments showed that using the proposed method for money laundering detection leads to a false positive rate of about 13%. In this paper, we improved the proposed method and reduced the false positive rate to less than 1%. To achieve this, we analyzed the amount of financial transactions of sellers along with the number of visitors to their websites in PGA. Based on the number of visitors, we estimated the volume of transactions for each seller. If the volume of sales was much higher than expected, we considered it abnormal. We achieved a higher detection accuracy by using a restricted Boltzmann machine to separate out-of-class transactions. We also reduced rate of false negative alarms by the help of CBR method. Our proposed system detects money laundering online using a four-week sliding window. The experimental results confirmed the detection accuracy of 99% for our proposed system.