Title: Probabilistic orbits and dynamical masses of emission-line binaries

Abstract: The observed orbits of emission-line stars may be affected by systematics owing to their broad emission lines being formed in complex and extended environments. This is problematic when orbital parameter probability distributions are estimated assuming radial-velocity data are solely comprised of Keplerian motion plus Gaussian white noise, leading to overconfident and inaccurate orbital solutions, with implications for the inferred dynamical masses and hence evolutionary models. We present a framework in which these systems can be meaningfully analysed. We synthesise benchmark datasets, each with a different and challenging noise formulation, for testing the performance of different algorithms. We make these datasets freely available with the aim of making model validation an easy and standardised practice in this field. Next, we develop an application of Gaussian processes to model the radial-velocity systematics of emission-line binaries, named marginalised $\mathcal{GP}$. We benchmark this algorithm, along with current standardised algorithms, on the synthetic datasets and find our marginalised $\mathcal{GP}$ algorithm performs significantly better than the standard algorithms for data contaminated by systematics. Finally, we apply the marginalised $\mathcal{GP}$ algorithm to four prototypical emission-line binaries: Eta Carinae, GG Carinae, WR 140, and WR 133. We find systematics to be present in several of these case studies; and consequently, the predicted orbital parameter distributions, and dynamical masses, are modified from those previously determined.