Circular trading of goods is a carefully designed scam ubiquitous among fraudulent business dealers all around the world. Dealers involved in this scheme create an artificial trading network by issuing doctored sales-invoices amongst themselves without any movement of goods. In practice, it is observed that almost all cases of circular trade involve two or three dealers. Here, we work towards predicting circular trade involving three dealers. For the same, we built four different classification models consisting of feature variables tailored for predicting any plausible circular trade amongst three dealers. In particular, the logistic regression model gave the best performance among all the four different models with a prediction accuracy of 80%. Interestingly, we observe that a feature variable formed by using the personalised PageRank technique significantly improves the model over the state of the art link prediction variables. Predicting a future circular trade from a huge network of sales-transactions data is of significant importance to the tax enforcement officers. In addition to automating the process of detecting circular trading, which is manually impossible, this model helps them to target on a set of plausible evaders and take appropriate preventive measures. This model have been developed for the Commercial Taxes Department, Government of Telangana, India, using their first two quarter's tax returns dataset. © 2021 ACM.