/vt/ - Virtual Youtubers

I'll start making funny graphs again when I find a good normalization for all the data.
>>2881238
LSTM stands for Long Short Term Memory. The idea is that you use whole batches of correlated data as input. This allow for a certain degree of prediction. Keep in mind that I'm absolutely new to this, I only have a little experience in the mathematics behind from the beginning of my degree.
My first model was from a simple ARIMA regression, I got the formula for each holo and since the adjustment only happens once a day, I just had to train the damn thing every day OR normalize the adjustment. Truth be told, I never actually used it since my main problem comes from scrapping the data and I only managed to into websockets 2 days ago. Before that I used the API but it gave weird numbers when recreating the queue from the history, probably due to immediate sell <-> buy transactions. The model was linear in the following parameters:
>number of coins in circulation
>coins bought and sold during the cycle
>>2881283
>>2879027
Regression is now useless on the short term, hence why I will try LSTM.