The one-on-one pair trading strategy is commonly used for arbitrage. The pair trading targets are usually chosen from two products or companies with highly business-related in the same industry/market, and then obtaining profits from arbitrage according to the information of the fundamentals and stock price changes of the two investment targets. The arbitrage strategies are highly based on the assumption that the deviations from the relative stock prices changes are mean reverting. However, the life cycles of products are getting shorter and shorter, and the development of new technologies are quite rapid. The deviations between the originally stable relative prices may not be always mean reverting. This paper utilizes the threshold unit root model developed by Caner and Hansen (2001) to find the threshold values and the appropriate arbitrage strategies. We use TSMC and UMC as the sample companies to compare the performances among a variety of pair trading strategies. The sample period covers 2,743 daily trading data starting from January 1st, 2000 to December 31th, 2010. The trading strategies include: (1) the technical trading strategy based on the price ratio of the two investment targets, (2) the modified technical trading strategy based on both the price ratio and the fundamental information, (3) the cointegration trading strategy executed when the values of the error correction term excessively deviate from its mean value if the two companies' share prices are cointegrated; (4) the threshold unit root trading strategy executed when the deviations of the stock prices are nonlinearly unit root based on the threshold unit root model .This article investigates the performance of the above four trading strategies. The empirical results show that the strategy based on the threshold unit root model is the most profitable among the four pair trading strategies.