Hands-On Machine Learning for Algorithmic Trading

Outlier control

Data preparation before analysis typically includes treatment of outliers, example, by winsorizing, or clipping, extreme values. The challenge is to identify outliers that are truly not representative of the period under analysis, as opposed to extreme values that are an integral part of the market environment at that time. Many market models assume normally-distributed data when extreme values are observed more frequently, as suggested by fat-tailed distributions.

The solution involves careful analysis of outliers with respect to the probability of extreme values occurring and adjusting the strategy parameters to this reality.