Gekko Forum
State of Neural Networks strategies - Printable Version

+- Gekko Forum (https://forum.gekko.wizb.it)
+-- Forum: Gekko (https://forum.gekko.wizb.it/forum-13.html)
+--- Forum: Strategy Development (https://forum.gekko.wizb.it/forum-12.html)
+--- Thread: State of Neural Networks strategies (/thread-57965.html)



State of Neural Networks strategies - rdeman - 07-09-2019

It appears to me that the development of Neural Network based strategies for Gekko have gone a bit... stale?


The best article so far on this topic has been written by Deandree Blake at Medium.com (see: Medium.com "Crypto Trading 2018 in Review: 17 Advanced + 15 Neural Net strategies tested" https://medium.com/coinmonks/crypto-trading-2018-in-review-17-advanced-15-neural-net-strategies-tested-part-7-d308963d13e9, also referred to on this forum "Neural Network strategy backtest comparison": https://forum.gekko.wizb.it/thread-57766.html ) So far the most helpful resources I have seen, yet this all is about 2018.

Deandree but also others on various postings seem to grab their strategies from the infamous xFFFFF Gekko-Strategies Github repository:
https://github.com/xFFFFF/Gekko-Strategies: Latest commit was on 24 May 2018. Not very active....

As Deandree points out in his Medium.com article: "All strategies share some common traits: they use some kind of NN lib. Most (13/15) use convnetjs. The other 2 use LSTM type lib — synaptic or neataptic."

So 13/15 of the Neural Net strategies in the xFFFFF repository are using convnetjs (https://github.com/karpathy/convnetjs): Latest commit on 24 Nov 2016. The most popular neural network library has not been maintained in 3 years... Dodgy
1/15 strat (gekkoNeatapticIndicator) uses Neataptic: https://github.com/wagenaartje/neataptic Last commit on 9 Jun 2018 over a year without maintenance

1/15 strat (LSTM_MACD_RSI_V3) uses Synaptic: https://github.com/cazala/synaptic Last commit 14 Feb (2019 that is, I assume, reasonably active)

So here we have it: Only LSTM_MACD_RSI_V3 uses a neural network library that is up to date (and gets points for clean code) but sadly, LSTM_MACD_RSI_V3 does not score very well in Deandree's backtesting...

Is this really the best we have? Does anyone know of more up to date information? 

NOTE: It looks like zuki_nn is so far one of the best neural networks out there and it relies on a 2016 neural network library... so yes I do newer does not necessarily mean better. However, it seems things have gone stale since the 2018 crypto bear market. I am just curious to see if anything significant has happened since.