Impact of various kernel and length scale choices on a GP fit of a ZTF SN Ia light curve. If the uncertainties in the data are underestimated, not including a white noise kernel will result in unrealistic wiggles in an attempt to fit all the points (top left). Even if the uncertainties are not underestimated in the data, a lack of a white noise kernel can result in an underestimation of the uncertainty of the GP fit, as is visible between 10–30 d in the light curve shown in the top-right panel. With limited data at the edge of the fitting regions, not including a constant kernel can result in deviating behaviour at the edge regions (bottom left). We optimise the length scale during the fit in the bottom-right panel, which in this case resulted in ℓ = 12.3 d. By setting a length scale, the likelihood is higher to either over-fit if the length scale is too short (ℓ = 3 d), or to under-fit and miss important features if the length scale is too long (ℓ = 30 d). The ideal value for the length scale will depend on the cadence of the data, which varies across our sample.