September 13, 2019 at 11:17 pm #5847
We have a big update coming soon.
Anomaly detection gives us the ability to counter trade, which has been a goal of ours for a while now. Practically, this means our ML should reliably spot pending reversals in price (anomalies).
We just had a significant breakthrough in AD that generates next-level results.
Our current ML has lived in the swing-trading space (momentum rather than reversal), but this anomaly classification has major ramifications for day trading as well as long-term investing.
Needless to say, we’re excited to share. Look for updates and this to be rolled out ASAP…
September 14, 2019 at 7:56 am #5856
Looking forward to see how this affects performance!
September 16, 2019 at 9:04 pm #5866
Great! Yes…we’re just doing some final quality validation to really run it through the ringer and see how it performs on new and different data sets. But this is certainly the biggest advance since the start of Crypto-ML. 🙂
September 30, 2019 at 10:01 pm #5995joe1prattParticipant
Any information you can share on the release date for version 4.0 with anomaly detection?
October 1, 2019 at 7:29 pm #5999
Hi @joe1pratt, this is such a big departure from the current models, we’re really putting it through the rigor and testing in an expanding set of scenarios.
To be straight forward, we’re finding it is extremely capable–meaning it *can* generate results that are multiples better. However, it is also less dependable–meaning there are scenarios in which it really falters.
As an analogy, if this was autonomous driving, we have something that can blast through freeway traffic at 120 MPH. But when it goes to park, it keeps denting its fender on light poles. Unacceptable on a high-end car.
To address this, we’re generating our biggest data set ever. It is time consuming to get some of the historical data points at the frequency we need. And it’s also a massive amount of data to process.
But the expectation is this will train our model in a sufficiently broad set of scenarios…and we should avoid those darn bumper dents.
If I step back and look at our current models, they are a little more cautious and prefer to stick a little closer to the speed limits. But they’re dependable and get us where we want to go in a broad set of conditions. As a side bonus, this massive new data set we are generating will, at minimum, make our current model stronger and more robust.
To sum…4.0 anomaly detection is our future and our focus. Whether it makes up the core of our models or it becomes another tool for our current models to leverage is yet to be seen. I expect we are weeks (not months) away from getting something live.
October 18, 2019 at 5:40 pm #6139
Everyone, I’m extremely happy to report that we have the model ready.
Going with the analogy above, we are no longer running into light poles!
Functionally, this means we have a new anomaly detection model that can take the same parameters and work across different time periods and different cryptos. It also works across split data tests and the other fundamental analyses we put our models through.
It considers price, volume, month, day of week, day of month, hour of day, and 5 proprietary variance calculations. Based on these, it predicts average price for a period in the next day. If that value exceeds certain metrics, it generates a signal.
And let me tell you…it’s a game changer.
The next step is to get this up on the site. It’ll be on the Member Dashboard with alerts. It will be listed as a BTC Beta signal. Since it hasn’t been giving live trades to our customers, we feel most comfortable releasing it as beta. But rest assured, we are personally trading off of it already.
You can expect a full write up shortly.
Now time for the weekend! Thanks for the support, everyone.
October 19, 2019 at 12:05 am #6140martyfParticipant
congrats! Can’t wait to see it live!
October 19, 2019 at 1:29 am #6141matthew55Participant
Looking forward to seeing how it performs!
May I ask how many signals we can expect from it?
And, maybe, the possible gain compared to the current strategy?
October 21, 2019 at 10:08 am #6155
Hey, that’s great news, congrats! Really looking forward to trying it out!!
October 22, 2019 at 9:28 am #6157
Hi @matthew55 and @kw, thanks!
In terms of performance, the biggest difference is it is more robust. That means the same exact model with the same exact parameters works in a much broader set of scenarios. We’ll cover some numbers in the release comparing performance, but the real test will be how it performs over the next year or so.
We’ll also cover the frequency. It can be a bit choppier than the current models. That means it might open a trade and then close it 50 minutes later. Or it may keep a trade open for a very long time.
The current models seem to have a feel to them in terms of trade length/frequency, but I think we’ll not be able to expect that with the new one. That is inherent in the approach–it’s looking for unusual data (anomalies), which can occur at any time.
This also means we cannot support the concept of a Strength Meter. We’re working on other options to provide some form of visualization.
We should have a post out tomorrow or Thursday covering everything.
October 22, 2019 at 12:33 pm #6173martyfParticipant
> That means it might open a trade and then close it 50 minutes later.
Hrm…. this means I might need to open a binance US account after all. Those coinbase fees are getting a bit steep.
Although I guess the first iteration of this isn’t going to be on auto-trade, so we won’t have to worry about the bot getting tons of fees.
Looking forward to the post! Can you give us any teasers? Did the new model call the swings of this recent up and down? 🙂
October 23, 2019 at 7:46 am #6178
Saw this on Twitter.
It’s the last Friday of the month. That means futures expire in two days and like clockwork bitcoin fell about 7% to about $7500.
I looked into it, and there definitely seems to be a pattern there.
And Justin mentioned this about anomaly detection:
It considers price, volume, month, day of week, day of month, hour of day
I wonder if the new ML model picks up on a similar pattern?
October 23, 2019 at 1:02 pm #6181matthew55Participant
More important, would be nice to know if we can expect a recovery pretty soon… Or if every month we have to expect a loss of 7%
Also because Crypto ML called a Long just when it should have called a Short, since the position was closed very quickly with the mentioned loss…
October 23, 2019 at 5:53 pm #6183
@kw, yes…you’re amazing. This is actually making it in the write-up that’s coming out tomorrow. There are absolutely patterns when it comes to time.I need to double-check the numbers, but I think time-based factors make up around 30% of model, which is big. Good eye. 🙂
@matthew55, yes…there was clear consolidation happening and based off traditional technical analysis, there was certainly an impending breakout (positive or negative). Given that we’re in a bear market, it seems natural to assume a negative breakout, but our current signal did call a long.
I don’t think it’s as simple as expecting a loss every month at this time.
Since the forthcoming anomaly model will consider these types of events, the expectation is it will continue to call the right direction on these more often than not.
Great questions/comments, guys. This sort of discussion makes us all (particularly me) smarter and helps ensure Crypto-ML considers the right factors in our data sets.
October 23, 2019 at 7:31 pm #6186
Check this out. The blue line is the last Friday and each month. The red arrow shows where it dumps a couple days before.
October 25, 2019 at 9:27 am #6199
@kw, that is a great chart. It’s one thing to see it in the numbers, but another to see it visualized like that. Great stuff.
I’ll say it again–these are the types of discussions that make us all more knowledgable and capable. Thanks for sharing.
- You must be logged in to reply to this topic.