Home Community Trade Signal Discussion ML bot is not handling this chop well

Viewing 6 reply threads
  • Author
    • #5575

      is there any concern over the recent activity of the algo? The bot has been faked out I think 3 times now as this triangle is coming to a close. Personally I’m in a no-trade zone as I wait for this thing to resolve.

    • #5577

      I’m keen to know a little more too. My guess is that the new SL functionality has kicked in here and the bot is actually just trying to trade through this tightening triangle/wedge or whatever you wanna call it – and we just hope it’s on the right side of the trade when BTC makes it move

    • #5584

      Great question and we are definitely in a highly uncertain period. A few things have me being very cautious:

      1. The Market Index is dropping pretty quickly and is approaching bear territory.

      2. The signals are very conflicted. Many times, you’ll see the “short” signal closed and all of the “long” signals open around the same time. But right now, we’re seeing them all over the place.

      3. There are a lot of very large moves happening. This has occurred in the past and makes things very unpredicatble. Here is a read on this: https://finance.yahoo.com/news/massive-bitcoin-transactions-few-explanations-144159601.html

      How we’ve seen the ML respond in the past is it is opening positions at pivotal points. It’s possible we’ll see a breakout up, but these continue to be failed launches.

      As it says in the article: The market is ready to go way up. Or way down. 🙂

      One answer to this is we are continuing to feed data to the ML and let it absorb, understand, and find patterns in the current situation. We are all in on this. It certainly feels as though this could be handled better.

      The second piece of this is trading is a numbers game. Even though we have a string of losses here, that is historically how it can go. We updated the Trade History page today. You’ll see that we’ve gone through phases of losing trades (which are kept in check) but then get followed by winners. This works out in the long run, but ideally, we want this to get to a point where we don’t get these losses.

      In terms of recommendations, I think it’s smart to be in the “no trade zone” or reduce the amount you are trading.

      A couple of weeks back we made a similar recommendation (based on indicators from across our platform) here: https://crypto-ml.com/crypto-market-analysis-bearish/

    • #5591

      Is the bot trained to recognize these consolidations? I would imagine the bout could be trained to recognize the early and middle stages of consolidation and avoid giving trending/breakout signals?

    • #5592

      > Is the bot trained to recognize these consolidations? I

      Based on my understanding of how this ML is arranged, I would bet that its not. Getting the bot to recognize a consolidation structure is effectively forcing a structure onto the ML parameters, whereas the entire point of ML is for the parameters to emerge from the data itself.

      But who knows – it could be possible for it to emerge.

      I’d hazard a guess that because the ML is trained to perform over a 30 day period, it may be missing this almost 3 month consolidation.

      “So here is how that plays out. If the Market Index correctly predicts the 30-day direction of the market, it receives a biscuit. If it doesn’t, no biscuit.”


      Granted, I don’t know how the open/close signals work within the market index. Yeah, it seems that Auto-ML is a different beast than Crypto-ML.


      Also, @Justin, I’m curious about this statement:

      “From a machine learning perspective, Auto-ML has been running in the background since January of 2018. ”

      Does this mean that it was only trained on data from January 2018? I would assume that you could feed in historical data to train – that it doesn’t necessarily need live data.

    • #5599

      Hello, yes–there have been consolidations in the past, so it should have good perspective on them.

      The behavior we tend to see is when price moves toward the top of a triangle (*resistance* line), it opens a trade. Right or wrong, it is likely “looking” for a breakout–which doesn’t always materialize.

      We’ve seen that behavior consistently across each of the crypto.

      But as I’m writing this, you can see our models have closed all signals prior to prices just now dropping below the support line resulting in a huge drop and now BTC is below $10K.

      So even though there were losses, we are also getting some clear protective notices here (especially considering the rapidly dropping Market Index).

      The opposite approach would be buying at the *support* lines of the descending triangle type formation. But this can be extremely dangerous. While you might get a positive bounce, there’s a real chance it will break through down extremely rapidly. I suspect that is why our models avoid it and prefer to instead is target breakouts on the high/resistance side.

      Does that make sense?

      And @martyf, in regard to the training data, we are building our own proprietary data set. Because of this, we can’t go grab years of back data. But our models do get to consider fully rich data going back to January of 2017.

      It looks at the big picture but weighs recent developments a bit heavier.

    • #5601

      Here is a visual of the descending triangle pattern (from traditional technical analysis) being referenced:

Viewing 6 reply threads
  • You must be logged in to reply to this topic.