Topic: Custom indicators for use in predictive models (Read 453 times)

I believe NBT is a 'stable coin', the value of 1 NBT is pegged to 1 USD. If we could predict the value for this we could predict bitcoin, which unfortunately our models can't do, it appears to be too efficient for our models to to find any inefficiencies. This goes for DASH, LTC and ETH, they appear to be to efficient for our models

Our back test results tell a different story.

According to my experiences, some issues like weak decentralization can be interesting or even impressing for some people.

I see what you did with the Ethereum

Dump prediction indeed seems problematic.

The reduced profit in coins like DASH and ripple, NBT could imply that the coin itself has fundamental issues, like premine or the coin rules could be less fair (weak decentralization, other issues).


However the thing that surprised me was Litecoin.

That's amazing. I wonder if shill activity on social media that we observed could be quantified and plugged in as an indicator.

I don't believe that TA will work for crypto markets when the value is so small. A little bit of btc can manipulate the market up or down so TA is useless here. Unlike stocks where the market cap is big, TA works much better there.

I work with a couple of AI / machine learning PhDs, we're currently adapting their models to the crypto markets. These markets seem to be young and fairly inefficient, one model we're focusing on is producing very interesting backtesting results.For anyone that's interested: https://www.mediafire.com/?57qzdvn71rqaacl

The model in question is based on recurrent neural networks and hawkes processes, the premise of it is to take TA indicators then use ML model to auto weight the importance of these indicators to accurately predict the direction of the price into the future. It continuously learns over time and updates its parameters given changing market condition


Currently we're using TA-lib, a python library of 150 TA indicators as the models inputs. Currently it seems to be quite good at investing at early stages of pumps,  however it's very poor at predicting price reversals ie knowing when sell. Obviously this isn't something trivial, events such as dumps can be completely random.

We want to improve the models performance and optimize it to act more efficiently in these markets. One way to do this is to use custom indicators that have been designed to work with specifically with these markets.

Which get's me to the topic of this post. It seems the loan rates on poloniex are much higher for coins before they dump, although they are high at other times as well, by making an indicator from:
a. the LoanOffer rates
b. the amount of active loans

We could feed this into the model to determine if it can find any patterns between how these metrics change with what the price does.

Has anyone had any experience with this ? I'm trying to figure out how the indicator would be programmed. Currently i'm thinking making an oscillator going from the minimum interest rate to the max ( 5% / day on polo). The rate of loans being taken and the amount of loans currently out for loan could also be used in another indicator somehow

This is just one idea I had, if anyone has written other custom indicators they use in these markets and are interested in what we're doing, please feel free to drop me a pm.

Thanks for your time
[email protected]