§ Improving the Gini Coefficient Formula's Accuracy

§ 2021-08-02

Just a few days ago, to my big surprise, Vitalik released a blog post about a topic critical to the functioning of Rug Pull Index. A blog post titled: "Against overuse of the Gini coefficient."

While I'm super interested to learn more about V's thoughts - I haven't had any time to read it yet. Still, a friend of mine, and also ex-BigchainDBer, Ryan Henderson, first made me aware of an inaccuracy in RPI's Gini coefficient calculation for small and/or extreme case populations as e.g. a population of two with incomes of 0 and 1. He, too, had recently used the Gini coefficient in a paper on neural networks.

Originally using Wikipedia's formula, for a maximally inequal populalation e.g. of 0 and 1, the Gini coefficient calculation ended up being G=12G=\frac{1}{2}, where as we'd expect it to be G=1G = 1.

However, after having some more discussions with Ryan and my friend Jost Arndt, we settled on a more intuitive derivation of the Gini coefficient for Rug Pull Index that should also produce reasonable results for small or extreme populations. For anyone interested, I updated the /specification page with the latest formula.