Stochastic field theory for time series

Description

Machine learning has so far underperformed in time series prediction (slowness and overfitting), and classical methods are ineffective at capturing nonlinearity. Revant Nayar shares an alternative approach that is faster and more transparent and does not overfit. It can also pick up regime changes in the time series and systematically captures all the nonlinearity of a given dataset.

Revant begins with an introduction to field theory and explains why it is expected to be useful in time series forecasting and analysis, with examples from physics. Revant compares this approach to canonical methods—in particular, ARMA and generalizations, ML-based methods, and Bayesian and Monte Carlo methods—and describes some of the properties of time series in general in relation to how these methods account for them. Revant then examines two datasets, one from commodity pricing and the other from the equity exchange, and demonstrates the performance of field theory versus some of the other methods. Revant briefly discusses possible trading strategies that could be created using the results and dives a bit deeper into field theory to explore how the machinery works in an intuitive manner, without the use of equations. Revant outlines some of the tools in the field theoretic toolbox and the different situations they could be applicable to and shares resources to learn more about or implement some of the techniques. Revant ends with a lamentation on the dire state of algorithmic hedge funds and explains how the situation could be alleviated using some of the techniques described in this talk.