Popov wrote:Actually I will spend the next two days in making the Walk Forward more powerful.
Since you're working on improving WFA, I have a suggestion:
As it is now, it's running the walk forward process on market data that the strategy development process has already been exposed to (the same period from where the strategy was discovered). So, technically the WFA is suffering from data snooping, because it's "walking" on the same data used to develop the strategy. For example: if 2005-2010 is used in the "Data Horizon", the Generator is looking for strategies that work on 2005-2010 period, and then the Reactor runs WF validation on the same period.
To simulate the "true" walk forward concept, the in-sample period used to discover and optimize the strategy should be completely independent of the out-of-sample, since this is what will happen when we are developing a strategy to run in the future (we have data from XXX to June-2018, and the key challenge is to be profitable after that).
So, my suggestion is: if the user selects walk forward validation in Reactor, and 2005-2010 is selected in "Data Horizon", only the first in-sample period should be available to Generator. Let's say that in-sample period is 1 year, then the data mining process should only run from Jan-2005 to Dec-2005, and the remaining period (2006-2010) used for walk forward validation.
Awesome work you're doing here!!!