2020 left us with plenty of time to fill: while some decided to spend it productively (learning Chinese or knitting, depending on who you ask), most of us followed the path of least resistance and decided to enjoy the endless stream of entertainment provided by our smart TVs.
Now that 2020 feels like a distant memory, if you are looking for a cultural break between the latest episode of Bridgerton and the first episode of Friends (which by now you must have watched for the tenth time), you might want to check out one of the many docuseries available on Netflix. Particularly fascinating is the series “Connected”, by Latif Nasser. It was while watching the fourth episode of that series, called “Digits”, that we had a breakthrough idea for how to improve forecasting models in financial services and beyond.
In a nutshell, the episode covers Benford’s law: a statistical law that predicts the distribution of the first digit in series of natural occurring numbers. So far, this law has been used on historical data sets, to identify instances where such datasets might possibly have been manipulated (essentially: fraud detection). But what if we add a twist? Instead of using the law to check data from the past, why not use it to evaluate how well we can predict the future? In other words, if we can use the law to see whether there is something odd in a data set, the law can also be used to identify how well predictive models can forecast future behaviour. As liquidity and funding experts, we successfully tested our hypothesis on stress-testing models. But we see that possible applications are unlimited when it comes to checking the output of predictive models. Imagine being able to evaluate two apparently equivalent AI algorithms, based on Benford testing, for example.
In our latest white paper, we address this in detail, and we pave the way for Benford’s law to be applied to other predictive models in the future.
The latest “ace up your sleeve” in the search for the optimal liquidity stress testing model: Benford’s law testing
If your work has anything to do with predictive models, stress tests or similar things, this publication will provide you with some great insights into how Benford’s law can be used to better evaluate the robustness of your models and to understand how realistic their output is.