I backtested 15 Neural Network trading strategies with TOP20 Crypto Coins

[deandree]

Hi, all!

I saw quite active trading discussion in this forum, so I thought I could share my research and maybe start a discussion about Neural Networks and trading strategies.

For the last few months I've been testing lots of trading strategies and sharing my results and analysis in Medium posts.
For my setup, I'm using Gekko open source platform. I would NOT call it BOT, it's more like a platform that let's you execute and test strategies that YOU have created.

So in the last month, I collected 15 best open source Neural Network strategies that I could find and backtested them against TOP coins, for September. For example, here is how results look like:



After that, I created another part, where I backtest July, August and October, month by month. To keep it short - there are good months and then there are bad ones, but I'm sure traders here already know that. Good thing is, even in this bear market, there is still a chance.

In my latest post (Part6), I get started on prediction analysis by trying to predict single event - huge price drop that started from November 14th. I draw the prediction lines on chart, so we can compare the accuracy, for example, here is neataptic strategy predictions with 5m candles and 900 history:


Often times I analyze buy/sell decisions made by strategies that accomplished best results, by drawing indicators on chart. For example, here is full Year 2018 results for RSIBB strategy from my upcoming (Part7) post:


Let me know if this is useful to you! Thanks!

[1714077481]

1714077481

[1714077481]

1714077481

[Marbigoad]

Thanks for making this! Can we expect a part with November backtests? Very curious how strategies handled the huge drop in price.

[butka]

I see from the URL that you are performing your tests using the 4 hours candles. Have you tried going to the daily candle size? In general, how big a difference can you see while varying the candle sizes in the Neural Network strategy? Is it like most of the other trading strategies: the higher the time frame the more reliable the strategy?

I've tried running Gekko once with a sub-optimal strategy (I think it was BUL BEAR something on an 1 hour candles) and the results were not so good.

[deandree]

@Marbigoad Thanks, for the next part I have decided to dig deeper into NN, after that maybe backtest November/December.

@butka I'm using 8h 6h 4h 2h 1h 30m 10m candles. I have not tried going higher than that, maybe in the future. Results vary by a large margin. Higher candles sizes are better, and my guess is because:
1) less noise in data
2) less trades therefor less paid in comissions

Do you mean RSI_BULL_BEAR? That strategy was quite good in beginning of the year but since then the market has adopted.

[butka]

Do you mean RSI_BULL_BEAR? That strategy was quite good in beginning of the year but since then the market has adopted.

Yes, that's the one. Do you want to say that the nowadays market is not so good for RSI_BULL_BEAR any more?

[deandree]

Yes, based on my research that seems to be true.

I've ran RSI_BULL_BEAR against June/July/August and results are quite bad, see here:
https://medium.com/@deandree/can-you-profit-in-crypto-markets-with-not-so-basic-strategies-part2-6cbc97a5f960
https://medium.com/@deandree/can-you-profit-in-crypto-markets-august-update-part-3-de43970a5128

On the other hand, I remember running it it March/April and results were much much better (profit). Unfortunately I don't have that documented in Medium.

Also, could be that just the parameters no longer work, not the strategy. You can always use some optimizer, like Genetic Algo. Haven't tried with RSI_BULL_BEAR though.

[paper-eddy]

neural network could be the thing I have been searching for. I would definitely investigate more on these

[Kotone]

neural network could be the thing I have been searching for. I would definitely investigate more on these

It also made me curious, this is the first time I heard that in tradind. I thought it's just a technical jargon in the tech background that same goes with AI and machine learning  I'll have it more investigated as well.

[deandree]

Well, Neural Nets are just one type of machine learning and AI is based on Neural Nets, so these are very related terms.
Neural Nets are relatively advanced trading technique, if you want to compare them against more basic strategies, I have 2 post about that here and here

[jeppex11]

If you figure out any that works on lower time frames please make a post about it Would be very interesting to see that

[deandree]

Lower timeframes have more noise and less signal which makes it much harder to create profitable strategies, because you are just trading against random ...
But yes, will keep posting my research

[deandree]

Everybody wants prices to go up, but is there really a reason? For that to happen, there must be more demand. And for demand, we need real world mainstream use cases. But I still don't see any.
My suggestion is to learn how to short, otherwise you will always be a slave to market only going up, by shorting you separate from direction and only care about trend, doesn't matter up or down.

[deandree]

Hey all, I was planning on writing new piece where I backtest full Year 2018 with top strategies, would someone be interested in Medium post like that?