Strong gravitational lensing provides fundamental insights into the understanding of the dark matter distribution in massive galaxies, galaxy clusters and the background cosmology. Despite their importance, the number of gravitational arcs discovered so far is small. The urge for more complete, large samples and unbiased methods of selecting candidates is rising. A number of methods for the automatic detection of arcs have been proposed in the literature, but large amounts of spurious detections retrieved by these methods forces observers to visually inspect thousands of candidates per square degree in order to clean the samples. This approach is largely subjective and requires a huge amount of eye-ball checking, especially considering the actual and upcoming wide field surveys, which will cover thousands of square degrees. In this paper we study the statistical properties of colours of gravitational arcs detected in the 37 deg^2 of the CARS survey. We have found that most of them lie in a relatively small region of the (g'-r',r'-i') colour-colour diagram. To explain this property, we provide a model which includes the lensing optical depth expected in a LCDM cosmology that, in combination with the sources' redshift distribution of a given survey, in our case CARS, peaks for sources at redshift z~1. By further modelling the colours derived from the SED of the galaxies dominating the population at that redshift, the model well reproduces the observed colours. By taking advantage of the colour selection suggested by both data and model, we show that this multi-band filtering returns a sample 83% complete and a contamination reduced by a factor of ~6.5 with respect to the single-band arcfinder sample. New arc candidates are also proposed.