What structure should a predictive neural network have?

Question

Neural networks

Hello!
As a thesis, I decided to create a neural network that (theoretically) could predict the dollar rate based on news articles. I understand that this is basically impossible and the result will be close to random, but this is a diploma, so the result is not so important :) I want to ask for advice. I must say right away that I started reading about neural networks just a couple of days ago, so I ask you not to throw slippers much.
Now I have parsed about 147,000 articles over the past 9 years, highlighted the keywords, processed them lexically, and removed unnecessary ones in my opinion. Thus, out of 90k keywords, I got about 40k.
As for neural networks, I understood only an approximate principle of their work: we submit some data to the input, and at the output we get the probability of a positive or negative answer. At least, this principle was described in the guides that I read) But networks with only 5-10 inputs were considered there, 0/1 was applied to each input, the output was a probability of one. I figured out how to create such networks and how to train them. But, unfortunately, I did not find articles describing neural networks that would solve my problem (or close to it).
At the moment I have 2 questions:
1. How to create a neural network that could process keywords? The only thing that is spinning in my head is to create 40k inputs from the network. But it is logical that this is wrong. Moreover, at some point, keywords can be added.
2. How to make the neural network give out not the probability of an event, but the direction of the series (the rate will go up, strongly up, remain the same, down, etc.)?
I do not ask you to describe the algorithm in full, just tell me, please, the names of algorithms, techniques, etc., articles that could help me :)
Thank you in advance

Answer 1

As a thesis, I decided to create a neural network that (theoretically) could predict the dollar rate based on news articles

it doesn't work like that
Not only in relation to this currency, but in general
Also don't think that this idea came to you first
Also there was a feeling that you don't understand how ML works in principle in the real world

But, unfortunately, I did not find articles describing neural networks that would solve my problem (or close to it).

You haven’t found how fact extraction, sentiment analysis is done, attempts to write programs for forecasting rates and find formulas for making a naive buy / sell forecast on investing.com
My advice is to change the subject while there is time
How to predict the ruble exchange rate

Answer 2

There is such a thing as "technical analysis" in finance. They predict the course without any external sources at all, purely according to the history of fluctuations.
Words\algorithms for working with text: word embeddings, RNN, LSTM, CNN, Transformer (the topper now)
But in general, if I were you, I would still try a task that is closer to the ground. Last year, a rather curious architecture / article Neural Ordinal Differential Equations appeared, which, among other things, you can try to apply to exchange rates / stocks.