During preliminary research runs on WEKA I use 100 candle windows, this is too large on a daily timeframe. However that is not too big of an issue as the window size can be easily modified in configuration of the range feature extractor. What is really an issue is that, say for example, I get "not range" for one window... and then the next decision later i get "is range".... the subsequent decision is based on a new window that only shifted one candle from the beginning candle of the previous window. To be more clear, that subsequent decision is determined from 99/100 candles already used in the precvious decision. Therefore, the model's decisions are bleeding into each other. In other words, the decisions are based on overlapping data. This issue exposes a design flaw.
During preliminary research runs on WEKA I use 100 candle windows, this is too large on a daily timeframe. However that is not too big of an issue as the window size can be easily modified in configuration of the range feature extractor. What is really an issue is that, say for example, I get "not range" for one window... and then the next decision later i get "is range".... the subsequent decision is based on a new window that only shifted one candle from the beginning candle of the previous window. To be more clear, that subsequent decision is determined from 99/100 candles already used in the precvious decision. Therefore, the model's decisions are bleeding into each other. In other words, the decisions are based on overlapping data. This issue exposes a design flaw.