We propose a distributed XML stream filtering system that uses a large number of subscribers? profiles, written in XPath expressions, to filter XML streams and then publish the filtered data in real-time. To realize the proposed system, we define XPath expression features on XML data and utilize them to forecast the servers? loads. Our method is realized by combining methods to share the total transfer loads of each filtering server and to equalize the sum of overlap size between filtering servers. Experiments show that the rate at which the publishing time increases with the number of XPath expressions is three times smaller in the proposed system than in the round-robin method. Furthermore, the overhead of the proposed method is quite low.