Home Community Crypto, Trading, Investing, and Technology BTC Beta – Stop Loss Has Tightened

Viewing 12 reply threads
  • Author
    • #6862

      the “relaxing” of the stop loss parameters in bullish conditions will be something that you’ll have to do manually? Or the system is now better able to understand when the market behavior changes and adapt accordingly?

      Because I understood it was already able to automatically tighten or relax its parameters, even though apparently it didn’t work out very well recently 🙂

      Thank you for the clarification

    • #6863

      Point of an ML system is to eliminate human interference and let it adapt on its own. I have no doubt in its ability to remain profitable in the long term.
      By the way, i haven’t seen any disclaimer anywhere, but are the statistics in the trade history going back all the way from the start of 2018 just the backtesting data or are they done with live forward trading in real conditions?

    • #6864

      Hey guys, I’ll chime in here.

      , our approach with BTC Beta has been to let it come to the model and solution entirely on its own with no constraints. In other words, we didn’t say you have to have x many trades per day, week, or month. We didn’t put limits on stop losses, etc…

      The output is a very aggressive model that, given enough time, performs extremely well (in our testing) across pretty much any market condition. But it does produce losses. And those losses probably exceed most people’s tolerance level.

      As such, we posed the problem–can it generate similar results while also being a little more conservative. And the answer is yes. This more conservative approach actually would have done exceptionally well in 2019 (beating the more aggressive model) but slightly worse (albeit still great) in other years. We’ll need to monitor.

      *Here’s my token analogy- This would be the same as telling a self-driving car to accelerate and brake a little less harsh. The core model is still intact and it still drives well, but it’s now considering the human factor of comfort.

      We prefer to be as hands-off as possible, but there are times that it is appropriate to challenge our models in different ways.

      If you want a deep dive into this model–and you haven’t read it yet–I highly encourage you to take a look at this post: https://crypto-ml.com/bitcoin-trading-with-machine-learning-anomaly-detection/

      Regarding the Trade History, every single trade listed there is an actual alert delivered by our platform to real customers. It does *not* include any backtesting or hypothetical data.

      , great question–and yes, the systems do have the ability to adjust these levels on the fly. But it is keeping this aggressive stop loss value. In fact, all of our models seem to be pegged as low as they want to go–we haven’t seen the values change in quite some time.

      The manual intervention was to see if it can find a similarly-performing set of parameters while being a little more conservative, which it was able to do. Hopefully that makes sense.

    • #6870

      @JUSTIN Thanks for the clarification! It really depends on the individual risk profile, whether or not they want to experience more volatility in their portfolio in order to potentially generate greater returns. I have seen other great mechanical, non-ML based systems that are able to beat the market systematically like opening range breakouts (ORB) systems with entry filters. But as they always say, the markets are always changing and with that, the system that used to work for you does not always mean it will work in the future.

      That is why I strongly believe in ML systems. To me, ML systems perform self-optimizations but in a more mechanical manner, eliminating curve over-fitting as much as possible compared to human interventions. I am very curious as to what you guys, the developers of ML systems think of what the future of trading would be; would ML systems eventually rule the trading world or human crafted mechanical systems still have a place many years from now?

    • #6963

      Thanks for the update team. Curious how you arrived at 3.54?

      Have you considered running a model/exercise to explore pairing market index and SL together and subsequently whether any paired/categorical features could assist in improving model performance?

    • #6967

      The 3.54% was just derived by the machine, as was the 6%.

      The first part of the machine learning is to feed data in so that it can build a model that predicts what you’re trying to predict. This answers: what factors cause price to change?

      The second step is to turn that into an actionable system that delivers alerts, which is done based on certain parameter levels. To determine those levels, a second round of machine learning gets applied which optimizes those parameters. This answers: when those factors change, when should I take action?

      That second part is what optimizes the stop loss.

      And yes, the anomaly detection does consider the Market Index. Since BTC Beta was introduced, the markets have dropped 15%, which poses a challenge regardless. We will have a write-up out shortly on this, the markets, a 2019 wrap-up, and what to expect going forward.

    • #6968

      When btc beta goes from open to closed, is that the sell signal or is open with a bearish signal the sell signal? I am a bit confused by the documentation

      • #6973

        BTC Beta can close under two conditions:

        1. When a down pulse occurs (bearish anomaly detected)
        2. When the stop loss is hit

        Looking at the visualization now, you do not see a down pulse, but that is because it was closed via a stop loss. It would be ideal to have a visualization of when the stop occurs. I have submitted an enhancement request for the same.

    • #7733

      Good work, but maybe i think tight stop losses are good.

    • #7755

      Yes, absolutely. It’s a balance…too tight and it’s easy to get prematurely stopped out; too loose and it’s easy to take unacceptable losses.

    • #7775

      The recent ETH long got stopped out because the AI thought that a reversal for ETH is going to happen, and it just so happens that the price crosses up the threshold, maybe just by a slight margin, which dictates opening a long right there and then. But alas price dropped thereafter and the downtrend continued. But i believe that we are still in a bull market and uptrend will resume some time in the near future. I have faith that the AI will learn from this mistake and be even better in the future. The main point is, you can be right just around 50% of the times or even less, but when you are right, you ride out big profit margins. That is what this intelligent trend-trading AI is all about. In the long run as shown in trade history, the AI will be very profitable compared to HODLING!

    • #7777

      Hey @glennseer, I totally agree. The level looked interesting for a reversal, but it also seemed fairly risky to go after a bounce.

      As for the winning trade ratio, you’re exactly right…in fact, we pulled the stats on our trades and posted those here: https://crypto-ml.com/blog/bitcoin-trading-with-machine-learning-anomaly-detection/#Wins_and_losses

      And we’re actually witnessing one of the less-obvious value points of the ML right now–with the positions closed, we’re avoiding this drop. Missing these bigger drops makes a big difference over time.

    • #7778

      The portfolio management cannot come soon enough. I cannot keep up with these ETH trades! And some of them are coming while I’m sleeping.

      • #7782

        Hi @kw, absolutely and noted.

    • #7779


      This time round, ETH opened long just before the pump. It’s moments like these that makes me strongly believe in the AI’s ability to capture the edge that traditional technical indicators cannot. Let’s hope this pump does not lose momentum and result in a continuing downtrend 🙂

    • #6860

      Please note that the BTC Beta Model parameters have optimized for more bearish conditions. The stop loss has moved tighter from approximately -6% to -3.54%. Other parameters have likewise shifted.

      This optimization nets 32.7% better results when tested against 2019 data. This change also tests well against other time frames and market conditions. But as and when bullish conditions return, these parameters will once again relax.

      Even with the change, losses are still produced and expected. However, this change will better manage adverse impact from losses.

Viewing 12 reply threads
  • You must be logged in to reply to this topic.