- July 15, 2020 at 10:31 pm #9304
BCH has now also been upgraded to 5.40.
- July 16, 2020 at 2:48 am #9305ppierpaoloParticipant
Hi, do you have backtests for the new mechanics and models? It would be nice to see how they perform compared to the previous mechanics and models (e.g. the difference in USD returns over the past 2 years).
- July 18, 2020 at 12:21 pm #9310
Yes. The backtesting we perform is complex and not easy to share. It involves multiple steps, including:
2. General backtesting
3. Backtesting by market condition
4. Forward testing
This link provides some insight into how this may be viewed: https://crypto-ml.com/blog/machine-learning-upgrade-to-5-0-deep-neural-networks/#Statistical_Measures
In general, any enhancement improves overall backtesting results while also better handling exceptions and edge cases.
- July 23, 2020 at 4:45 am #9407
Nice to hear about the upgrades!
I’m wondering how people are finding the autotrading? Are you getting the same results as on the site? What about costs (e.g., is there a minimum amount so that costs don’t eat up the earnings)?
Basically, before moving some of my hard earned crypto into autotrading, I’d like to know what others’ experience has been like.
- July 24, 2020 at 2:18 am #9416matthew55Participant
Hello Crypter, I started using Crypto-ML on Dec. 2019, (auto) trading only BTC at first…
After a couple of months I switched to ETH because the performance appeared to be better…
Right now (Jul. 2020) I accumulated a 72% profit in a bit more than 7 months…
I’m pretty impressed, honestly 🙂
- July 23, 2020 at 10:27 am #9408LeoTheLionParticipant
I have appreciated it. There will always be those trades that happen in the middle of the night with losses if not followed. I use binance us that has low fees that don’t eat up too much. Coinbase takes way too much with respect to fees. Minimal difference in prices vs website here, gains still there.
- July 24, 2020 at 12:07 am #9415
- July 24, 2020 at 2:35 am #9417
Thanks for letting me know, Matthew! That’s like 8% per month, compounded.
I appreciate the sharing of your experiences, Matthew and Julian!
Jonathan (aka ‘Crypter’)
- July 25, 2020 at 2:31 am #9419matthew55Participant
Are you able to quantify the increase of performance (in %) after these last changes (particularly the mechanics ones)?
I see that all of the 3 crypto (BTC, ETH, BCH) have now reached a much higher value compared to the moment when their respective LONG positions were closed… so I’m wondering whether the previous model would have behaved differently and returned higher gains in the same conditions…
Thank you for clarifying
- July 25, 2020 at 7:55 am #9420martyfParticipant
yeah, I would be interested in seeing how the new models perform compared to the old models. Surely these data can be obtained just by running the old code against price action.
- July 15, 2020 at 11:34 am #9299
Release 5.40 of our machine learning trading is in the process of being released.
– BTC-USD: Mechanics and model have already been upgraded.
– USD-BTC: Mechanics and model have already been upgraded.
– ETH-USD: Mechanics have been upgraded.
– BCH-USD: Upgrade will take place within 24 hours.
In order to better manage risk and capture profit, the following changes have been released as part of 5.40:
– Cool Down Improvement: This improvement helps avoid rapidly issuing a new trade after a previous trade closed. The goal is to avoid whipsaws and this is a general best practice in trading.
Previously, if our models issued a loss, they would avoid issuing a new trade until a certain “Cool Down” period elapsed. Now, this “Cool Down” occurs after both a gain and a loss. The duration of this period is determined by our machine learning optimizers.
– Slide Improvement: This improvement helps avoid keeping trades open during stagnating or fading markets. This also creates a “time limit” for a trade to build positive momentum and follows the “triple barrier exit” approach to trading.
Previously, once a position was opened, if the trade went negative and remained negative for a certain period of time, the trade would close, even if it did not hit the stop point. Now, any time price begins fading (even if the trade is positive), the trade will increment toward expiration and closure. This is calculated based on the number of periods price has faded and the current neural network prediction value. Once these combined values exceed a threshold determined by our machine learning optimizers, the trade will close.
– Profit Target Improvement: This improvement increases the likelihood of closing out trades for a positive return. It is designed to avoid getting into a trade that initially moves positive, then later drops negative before the trade is closed.
Profit targets were previously allowed to increase over the life of the trade. Now, the goal of the trade is established once the trade is opened. If that original goal is hit, the profit will be captured. After the appropriate “Cool Down” period, a new trade can be opened if the neural network prediction is still sufficiently positive.
– Improved Risk/Reward Management: This improvement attempts to further shift the risk/reward ratio of any trade in the favor of the trader.
Previously, certain trades could be opened even if the reward potential was lower than the risk potential (based on the predictions). This now has new controls so that only trades with a sufficient risk/reward ratio will result in an open. The “sufficient” level is established by our machine learning optimizers.
– Minor bug fixes: Additional minor fixed have also been deployed for USD-BTC, which is on target to exit beta by the end of July.
The neural network model for USD-BTC and BTC-USD was also updated along with the 5.40 release. The recent market performance is factored in and provides for higher accuracy across a wide variety of tested market conditions.
The neural networks are continuously trained and may be updated at any point.
For any questions or clarifications, please respond below or join us on Telegram.
- You must be logged in to reply to this topic.