Comparisons with Canonical Techniques


In the 1940s the biggest problem facing physics was how to deal with infinite dimensional spaces. In the industrial world, in finance and engineering, there are often an infinity of possibilities. Each financial forecast problem can be mapped to a quantum field theory! There are numerous benefits of a field theoretic approach over neural networks and more canonical AI. The sheer power and scope of our algorithms have surprised even ourselves. With a small fraction of the funding received by ML, we have already outperformed it in time series problems such as equity and demand forecasting. We are eager to widen the range of applicability of these approaches to pioneer a post-AI paradigm. One, free of overfitting, frustrating black boxes, slowness and arbitrariness.   


The table above draws comparisons between field theory and other techniques used for time series. ‘ARMA ‘ denotes classical methods including GARCH, ARIM, SARIMA and generalisations. ‘ML’ denotes all techniques involving neural networks, clustering, PCA, etc. FMI stands for ‘Field Machine Intelligence ‘, an acronym describing all field theory based algorithms. While it can break linearity, ML has various inbuilt assumptions such as stationarity, IID, regime dependence. It is a black box and overfits. While classical methods can break stationarity and IID, they are by definition linear. Field theory has none of these problems or assumptions. Indeed it can be shown that FMI reduces to autoregressive methods and neural networks and even chaos theoretic methods in different limits.