Traditional approaches to systematic investing have focused on long-term investing in stationary risk premia, with models typically constructed as a function of price or long-term macro-economic data. However many emerging data sources are starting to become available that can offer a much more detailed view of the world – such as high-frequency power consumption, shipping or weather forecast data. Unlike traditional financial inputs, such as price or yield curve changes, the nature of the relationship between this data and resulting price move is often idiosyncratic and can vary significantly across markets. In addition, for many of these markets traditionally stable market relationships are at risk of disruption from technological change – potentially making traditional modelling approaches unreliable. In this talk we’ll describe recent work at Cantab that attempts to apply machine learning techniques such as dynamic bayesian networks, and neural networks to dynamically model relationships in this data and identify trading opportunities. Compared to existing approaches this can potentially yield more adaptable models and require fewer explicit assumptions about the dynamics of individual markets.