Control of the crystallographic texture is an essential part of the processing of dual phase Ti alloys like Ti64. In these alloys, texture leads to unwanted anisotropy and affects the fatigue life in service. Texture development depends strongly on the process conditions, which will vary throughout a component. In an open die forging, for example, the temperature, strain rate and total deformation vary with position and time and so will the texture. Although crystal plasticity modelling has long been recognised as a powerful method of predicting the texture development during processing, accurate texture predictions require high fidelity and computationally demanding simulations, using crystal plasticity finite element modelling for example. This computational cost makes texture predictions prohibitive in a digital twin of a forging. Here, we investigate the capability of recently developed statistical techniques to improve the efficiency of crystal plasticity texture predictions during beta processing of a Ti alloy. In the first part, we have applied multilevel Monte Carlo (MLMC) estimation to make predictions using low fidelity models corrected by high fidelity predictions. MLMC is applicable to any kind of fidelity and here it was applied to 3 cases: one where the number of orientations used in the simulations was increased, one where the numbers of material points representing each simulation was increased between levels, and another where the simulation approach was varied, starting off with a Taylor approximation, scaling up to a self-consistent approximation and finally with a full-field model. Using the Fourier coefficients to represent the texture, we show that, for rolling, predictions from the very fast Taylor model can be corrected to match those of the high-fidelity full field model using this approach. In the second part, we applied Gaussian process estimation to build a surrogate mode of the full field crystal plasticity model. Using the output of a process simulation of a forging, 100 predictions of texture were made for different positions in the forging. These were then used to train a Gaussian process based surrogate model, using principal component analysis of the orientation distribution function to represent the texture. The surrogate model enables near instant prediction of the texture at any point of choice in the forging. This work shows that statistical methods developed in the field of data science can be successfully adopted to make efficient texture predictions during thermomechanical processing of light alloys.