Member-only story
Causal Popularity: A New Model for Artificial General Intelligence (AGI)
The research at Estimand is currently centered around Causal Popularity, a combination of cause and effect detection with measurement of popularity. While still in its infancy, this is a critical first step toward AGI.
Why it matters: The current means to build AGI using statistical methods is contrary to how humans think and is making things more complicated than they need to be.
It started with causal detection
While building a causal detection model, many Data Scientists declared it was impossible without a human expert. We disagreed and began experimenting. Using a form of backtesting to prove that X and Y are connected with the same equation across time, we declared success. Training on financial data, we realized multiple unexpected connections, such as banana prices and interest rates.
Root cause issue
The problem with our first approach is that it does not always discover the root cause. Instead, it can falsely detect an indicator as the actual cause. To find the root cause, we must use graph theory and the notion of popularity. This is similar to the PageRank algorithm of Google Fame [Easley & Kleinberg 2010].
For example, consider the banana prices and interest rate connection. The reason for this relationship is the price of bananas being pegged at $0.60 per pound unless a significant economic event occurs [Moses 2024]. However, the price change in bananas does not cause an interest rate change. Instead, it is an indication of another event occurring. What Judea Pearl, the father of Causality, refers to as Causal Irrelevance [Pearl 2022].
The easiest way to determine the root cause is to find the most popular connection to the indicator. In the case of bananas, this is the strength of the US Dollar (USD). Thus, the value of USD is the most significant influence on interest rates in the United States. This brings us closer to the required empirical evidence for a cause-effect relationship.
