Positioning is an essential feature in Narrow-Band Internet-of-Things (NB-IoT) systems. Observed Time Difference of Arrival is one of the supported positioning techniques for NB-IoT. It utilizes the downlink NB positioning reference signal (NPRS) generated based on a length-31 Gold sequence. Although a Gold sequence has good auto-correlation and cross-correlation properties, the correlation properties of NPRS in NB-IoT are still sub-optimal. The reason is mainly due to two facts: the number of NPRS symbols in each subframe is limited, and the featured sampling-rate is low. In this paper, we propose to modify the NPRS generation by exploiting the cross-correlation function of the NPRS. That is, for each orthogonal frequency division multiplexing (OFDM) symbol we generate the first NPRS symbol as specified in the current standard, i.e., a Gold sequence; while the second OFDM symbol is set to the additive inverse of the first one. Our simulation results show that the proposed NPRS sequence results in improving the correlation properties, particularly with respect to the cross-correlation property. Furthermore, 15% -30% positioning-accuracy improvements can be attained with the proposed method, compared to the legacy one under both Additive White Gaussian Noise and Extended-Pedestrian-A channels. The proposed NPRS sequence can also be applied to other similar systems, such as long-term-evolution (LTE).