Neural Network strategy backtest comparison
(11-13-2018, 06:44 AM)deandree Wrote: Yes, starting time is very important. I've seen quite huge (x2/x3) difference on 10m candles, on 120/240/etc it could be much higher.

Due to already long time it takes to run my tests for each Medium part, I have not included multiple starting time in my tests. If I wanted to do that,  I could try a few offsets (like for 240m candles take 0h 1h 2h 3h offset) and calculate average. If people want to see that, I can make special article about that.

Wat do you mean by "starting time"?
1) the history buffer Gekko watches before trading kicks in? (which is measured in the amount of candles)
2) the size of the candles? Ie 10m vs 1hr candles?
3) literally the start date+time when starting the bot (July 8th 2019 17:33)
By starting time I mean "from" variable. In this specific context, I mean all the different unique points you can start with, based on your candle size. E.g. if your candle size is 5min, you can start at 00:00 00:01 00:02 00:03 00:04. If your strategy shows huge difference with each starting point, you can be pretty sure it's overfitting.
I've made a new post in my Medium series - [Part 9] Crypto Trading 2019 Half Year Review: 17 Advanced + 15 Neural Net strategies tested.

I have refreshed my coin list to TOP 30 coins and I backtest first half of 2019 - Jan 1st to Jul 1st.
hm i'm kinda in doubts of these neural networks.
the problem is training them, by nature they can follow a patern and adapt to it easily
resulting in excelent on paper trading results, because they learned the curve.
idealy though they should forget that, and learn the trading patterns instead.
ea candle stick history of 50 candles to predict a candle.
not yet seen that.. (oh i wish this was python so easy to create a tensorflow NN in that language)
Respect   to  article author , some  wonderful information  . Read More

Forum Jump:

Users browsing this thread: