IMO, Market manipulations and Strong sentiments will cause pumps and dumps
You can't get away with the systematic risk like stock market crash or KOL's negative speech.

But you can front run some market manipulation.
Since bitcoin is highly leveraged, when there's imbalance long-short ratio + high open interest
Big players will likely to trade against retail investors, and make profit from liquidations.

For example:
When long-short ratio and open interest is much higher than average.
That means a lot more people are longing bitcoin.



You can see from the image i share in above link, the long-short ratio is super high before it get crash.

You can also check orderbook + trades history,
when orderbook get thinner and imbalance. (There's too much order accumulated on one side)
And trade volume is low, it could be a good time for market manipulation.

For finding long-short ratio and open interest, there's free data available from Binance.
https://www.binance.com/en/futures/funding-history/4

*long-short ratio = accounts with long position / accounts with short position
*open interest = opening contract amount nominated in USD.

Let me know if this is helpful for you.

Hi,

I think this is quite difficult to get a good result with AI.

There's 2 ways to do this

1) Using Deep learning to figure out what's going on next

If you are using DL, you will need to write another scripts to control the trading logic.
And the addon script could be super complicated since you cannot know whether the model is giving you right answer.

Another thing is your the time scale of your training data. And what's the future you are asking the model to predict.
If you are going to feed ultra short time data in your training data. Then your predicting result couldn't be too far from current.

I am doubting this is useful for normal trader, unless you are trading with high volume.

And if you are feeding training model with long-time data, then you model may lack sufficient to get a promising result.


2) Using reinforcement learning to train a bot.

If you are using RL, there's a problem of learning spaces.

How can you let the bot learn if there's isn't such a history?

For example, if your bot is training with data till year 2019.

It won't be able to understand why the market can goes to 65k.

And depends on different learning model (or agent in RL)

it will do random things to verify itself.

Another thing is that, your RL (or DRL) learning process will cost you a lot of dollars.

The computing power will cost a lot more whenever you expand your environment.

And i think it is a must for you to contain the PNL status for the bot (otherwise it doesn't know if it's doing right)
Plus OHCLV, you got 6 dimensions. And i have already tried with this. There's no promising result i can say.

In conclusion , Please do it for fun, rather for profit....