|
Guest article from Kris Longmore, Founder and Head of Quantitative Research at Quantify: In the context of systematic trading research, data mining bias refers to the risk of attributing significance to a result that was in fact due to chance. I refer to it as an "insidious threat" because it creeps into the research process naturally and quietly and can have disastrous results if not accounted for. In the extreme, this leads to the decision to trade a losing strategy, but at the very least, neglecting data mining bias leads to inflated and unrealistic performance expectations.
Data mining bias arises due to two fundamental characteristics of trading system research and development. The first is randomness, and the second is sequential comparison, which is the search for the 'best' parameter set.
In any series of trades arising from a given system, the realised performance will have an element of randomness as well as an element due to the inherent edge of the system. If we examine only a single system variant, we have no way of knowing the ratio of the two components. The random component is equally likely to be positive or negative, and in some cases, may be extreme. In any event, this random component results in variation around the long-term expectation of the system.
Sequential comparison and selection leads to the system variant with the best performance in the historical simulation being selected. Consider this selection process in the c...................... To view our full article Click here
|
|