The Mean Trap 2010-2020 Learning Model and the World at the End of the Line

The Mean Trap 2010-2020 Learning Model and the World at the End of the Line

Machine learning continues to grow in sophistication, but much of it is predicated on trust in an average world. Methods that measure performance or risk by standard deviation or a single metric, while effective in a stable environment, rapidly become impotent in a domain dominated by variability at the end of the road.

In forecasting financial markets or infectious disease outbreaks, models that worked just before suddenly lose their meaning. This is not a failure of the technology, but the result of a design that ignores the nonlinear nature of the world. Averages provide reassurance, but signs of crisis arise outside of them.

Against this background, recent years have seen an emphasis on models that deal with multiple states and contexts simultaneously. What Taleb's philosophy demonstrates is robustness rather than predictive accuracy. A design that can withstand extreme events is the condition for survival in a world at the end of the line.