I’ve written a neural network trading platform, and to train the networks I manually go back on a graph and select buy and sell ranges as input to train the networks.
I’ve tried various ways of automating this, mainly around using psar to figure out what would have been a historical buy or sell entry point and range. I’ve also tried adding in ADX to that mix to signify a trend with the psar value, but both options produce too much noise.
Is there a better way to automatically figure out historical ranges that, if they were traded on would produce less noise?
you could try donchian channels, they represent the high and low of the last n periods…if you want historical price ranges i would plot a DC of e.g 52 on weekly charts, then you get the range of last year (not the calendar year, the last 52 weeks)…you can play with different periods on different timeframes to find what you are looking for…
I’ve managed to rectify the auto training now after thinking of it a bit more. Basically taking the PSAR values, and then calculating for the PSAR period of “BUY” if the price of the start candle is lower than the price of the end candle. This would be a good period as an input to a NN.
I’ve just hit the simulate button on 20 instruments, so I’ll let you know what kind of results I get.