We present progress towards a new strategy to improve the precision of
PDFs in the small $x$ extrapolation region. In particular, we will show how a
Gaussian Process can be used to model fake deep inelastic scattering data in
the small $x$ extrapolation region, and how, by treating this synthetic data
as regular experimental data, it might be used in a PDF fit to control the
uncertainties of PDFs in the extrapolation region. Finally, we will discuss
current obstacles and possible solutions to these obstacles.