This paper shows that performing an XOR operation between the outputs of parallel arbiter PUFs generates a more secure output at the expense of reduced stability. In this work, we evaluate the security and stability of XOR PUFs using 1,000,000 randomly chosen challenges, applied to 10 custom-designed PUF chips, tested for 100,000 cycles per challenge, under different voltage and temperature conditions. Based on extensive hardware data, we propose a practical method for selecting challenges that will produce stable responses. A linear regression approach based on soft responses collected during enrollment phase was used to build accurate models for each individual arbiter PUF. Hardware data from fabricated chips verify that the approach is highly effective.