It presumably relies on having multiple phones recording the temperature, averaging out person-to-person variations in attire (thickness of pants, etc.). If those variations are correlated among people for a given day (everyone is wearing thicker pants), it's probably colder, and that would have to be included in the algorithm.