Gekko Forum
Gekko-GeneticAlgorithm - Printable Version

+- Gekko Forum (
+-- Forum: Gekko (
+--- Forum: Strategy Development (
+--- Thread: Gekko-GeneticAlgorithm (/thread-78.html)

Gekko-GeneticAlgorithm - susitronix - 01-25-2018

Good day

I tryed install the Gekko-GeneticAlgorithm from GekkoWarez (on Github) or
Enyone tryed this yet? This sound good.

Could not get it to work even with the node v7.1 as recomended
The installation guide says open two comand shells:

In terminal 1:
1) go to gekko/web
2) node –max-old-space-size=8192 server.js
Now the Gekko api server is running we can run the GA.

My OS: Win7-64

CMD: This is not recognized: "–max-old-space-size=8192"
instead i tried : "node server.js" and a Gekko API Server was startet.

Then in the second CMD

In terminal 2:

1) go to gekko/gekkoga

2) node run -c config/your-new.js
At the end of every epoch you will see the winner for that epoch, and you will see the overall winner across all the epochs running in that session. This will all appear in terminal 2.

node run -c config/your-new.js   //this comand also dos not work.

goto gekko/gekkoga   // i guess he meen: cd gekko/gekkoga??

Is this because the TaLib is not working on win7..? or not
my bad i gave up.

RE: Gekko-GeneticAlgorithm - Gryphon - 01-26-2018

Have you copied the sample-config.js in gekkoga/config into a file called your-new.js? That's the first step. (file name can be anything, so long as you use the matching name in the command)

GA needs a config to load from - they provide a sample for a 'customMACD' strategy, which you more than likely don't have.

To get running quickly I copied the gekko MACD strategy into custom.js. Next issue was that the settings didn't work - gekkoga doesn't seem to like mutating settings that are in a sub-array, i.e. [thresholds], so in the strategy I set it to look for the thresholds in settings, instead of settings.thresholds, and updated the GA config accordingly: eg. up:randomExt.float(20,0).toFixed(2)

I hadn't heard of gekkoga before your post, thanks for making me aware. The poor little RPi I have Gekko on has needed an extra heatsink after running GA for a few hours. Definitely giving results I wouldn't have tried myself, but so far I've found them to be incredibly specific to the training data and not very transferable.

MACD, USD:BTC 04Oct'17-01Jan'18
Candle Size: 120 min
short = 15
long = 16
signal = 11
down = -16.07
up = 7.82
persistence = 3

Gives ~350% profit, beating the market, but shift it to older data and it does nothing

RE: Gekko-GeneticAlgorithm - susitronix - 01-26-2018

Wow this sounds good, Installation wise. Ill try later today. Thanks

Its kind of strange that a working Strategie BackTest Result can vary by that much when changing the TimeFrame for Backtest Simulation.

In short time Backtest i recognized the beginning and the last Trade seems important.
If the lastTrade is a buy while the Trend goes Down, it will Hold a Bag. Therfoe the last Trade (Loss) could be subtracted.

In the documentation under github he describes:
Thresholds problem
Some strategies (~MACD) use threshold values that can vary a lot from coin to coin. Some coins might need values between 0.05 and -0.05 while others will need them between 0.00000005 and -0.00000005. This makes finding the correct Min and Max values time consuming and expensive. These thresholds are linked to the individual markets currency value (I think).
Rather use the PPO strategy since I've been informed that it is MACD but with percentages.

Im busy now but exited to try later: Thanks

My only Strategie that workt is:
-CandleSize 18min   //(((TRY ALSO 20min)))but 18 works better Backtesting sometimes..... Tongue
-interval=2   //<--strange i know

I Manually made BackTests for All Stategies included, but thats the only setting that looks profitable.

Also Recognized: if i Restart the Gekko/Backtest by Cmd-shell the same data (7-14days for Example) could bring a diffrent Result???
Maybe its also because the Backtest-Importer data dos only contain the CandleData.
This makes some nice Strategies usless for Backtest because rely on diffrent data

RE: Gekko-GeneticAlgorithm - susitronix - 01-27-2018

I shortly tested your MACD Strat setting.
dosent make positive trades. Itested only 1month

The settings looks pretty scrambled.
After my searching for a possibility to somehow predict the TrendDirection,
i found that the indicator settings used by professional Daytraders (PennyStocks) do work much better than my own "newbee" nervous Attempt.

Please if You know all this sikip the following, as a Beginner i learnd:

There is diffrent Volatility in the Markets therfor longer or shorter Trendlines/IndicatorSettings, dependent on faster/longer Swings.

Long Storry short. I guess the MACD crossing...
Example sellSignal: ...if at a upwards Trend the 9-line cossing the 21-line ----> They Sell
buySignal:    the 9-line is below. If it crosses the 21-line again/opposite ----->They Buy

The Point: I guess if the two lines short/long are close together, it could not create a clean Trend-reversal-crossing.
Also the persistance=3 mean: gecko will wait 3-candles befor a new Trend becoms acknowledged.
With a 120min Candle this take 6 hours after Trendreversal.
If a crossing occurs then MACD-Strat waits until a Threshold (%) has reached beyond the Crossing,
in order to filter out some nois that otherwise could trigger a to early desicion (Trend not prooved yet).

For me somehow it should work best with some of these professional-Optimized-Strategie-settings.
The Cryptomarket seems to be bit different (my experience binance 4weeks):
If at the TotalMarketCap 50G$ being pulled out in less then an hour then the Market just sinks down artificially.
The Panic Sells comes in delayed.
Then these people buy off cheep while on the other hand maybe push back 40G$ in another 40min...

Would be Twice as efficient if there is a dynamic Stoploss or Trend relatet WaitStateLogic.
I noob cant code this addition yet but its best time to get in to it.

i got Stratrunner going for 24h, it made one Roundtrip but now for 20h its doing nothing at all exept reciving Candles.
mabe i disrupted the process by using the laptop. Fixed older Laptop fan now got two. Think restart the Strat again and never touch this comp.
unforgenatly i got only one account yet. otherwise with altcoins could be intresting

-CANDLE        =9     //10
-HISTORY      =80   //80
-OPTINTIME   =13   //15
-CANDLESIZE =9    //10
-UP                 =31  //30
-DOWN          =22   //20

second Column might be weaker Result. cheers.all.philipp

RE: Gekko-GeneticAlgorithm - Gryphon - 01-30-2018

(01-27-2018, 02:58 AM)susitronix Wrote: I shortly tested your MACD Strat setting.
dosent make positive trades. Itested only 1month

The settings looks pretty scrambled.

Scrambled is the word - that's what I'd meant by results I wouldn't have tried myself. The issue applying machine learning to a data set is it gives a result that's highly specific to that data set. I was using USD:BTC from Bitfinex. I ran the same settings on USDT:BTC from Binance over the same time frame and it didn't make a single trade.

I thought running GA on data from a bear market, and setting it's target to score (Based on profit and sharpe ratio) instead of profit might give a result making better trades... tried it on some LTC history that was half bear, half bull and it started getting good results. I looked into it and found that it had set the required history to a large number that essentially stopped trades in the bear half of the data, then it traded in the bull half and made a profit.

Again, it managed the objective in a way very specific to the training data. I've no doubt it has potential to find really effective settings, but we need to learn what to train it on first.

RE: Gekko-GeneticAlgorithm - susitronix - 01-30-2018

This sound like a long term trader.
My objective is more towards dayTrader-Strat, where it would buy the Lows/sell the Highs around 4-6 times a day.
Making week Trend independent trades