I am researching a use of language models for making predictions.
I have this question: if I need GPT to predict something (like a sport event outcome or a future price of an investment) from lots of data, what do you think is a better way to use GPT:
- to ask for a precise prediction and then simplify it to whether the value like number of goals or the Bitcoin price will be above certain threshold, or
- to ask directly whether the value will be above the threshold?
Which way of thinking is better suited for GPT?
GPT is a Large Language Model.
If you are want to make predictions using numerical data you are in the wrong place. You could maybe use GPT with another AI / Algorithm, but it will not be your main driver.
Thanks, that’s what I am researching, what the quality of the predictions can be. And I am using mostly textual data, sorry for not mentioning that.
Here I want mostly answers to my question. I think producing an exact number and comparing the data with the threshold are two distinct tasks and GPT must be better at one of one than at the other.
Ok. You are in a sports store asking if you should buy a bass guitar or cello.
GPT, and any LLM does not produce or work reliably with numbers. Because they are Large Language Models. They are Generative AI.
If you are headstrong in using GPT your best bet is using Code Interpreter to handle the logic, and then using GPT to try and make sense of it.
If you want results. You should start your journey learning about “Predictive AI”
You can combine the two together to create a very potent machine. But using GPT to work with numbers is a huge no-go
Linear Regression came in mind for such cases or I am in the same store?
I am predicting from mostly textual data, so I can’t use machine learning models.
What I am asking is whether GPT can better predict an exact number, or decide whether the result will be above or below a threshold. I think these are quite distinct ways of predicting, it’s like regression × classification, so GPT might be better at one of them.
Can we say that the classification variant will be higher quality as it minimizes the working with numbers?
You want to predict a value based off of text? Not numbers?
By itself, no. You would need to somehow correlate these numbers to previous text examples using GPT.
100%. You could use the values output by GPT for your algorithm
I am newbie at ML than I just follow the story, I tried anyway to start a PoC for anomaly detection on stocks and I used linear regression (of course just numbers or anything transformed to a number again).
Ok thanks for your answer. I totally get the problem with numbers for LLMs but they have also a strength of understanding language. Predictive AI models can’t understand language but are good with numbers. So no option is good at both. As the numerical models are often used for predictions, I am researching using LLMs for predictions.
I also think that there is a need for minimizing the maths in the process. So it makes sense that asking whether Bitcoin price will be above today’s price in a week will be more successful than asking what the price will be. I think the second would make GPT repeat some stats rather than think about the prompt data. (This is just an example, I think there are better ways of predicting btc price.)
If you believe that you have semantic data that can determine if the price is going up or down in the future then yes, GPT would be a good choice as a classifier. Not for reading or outputting numerical data, especially exact numerical data.
that being said, it can predict some things, describe or send 2 frames of a person tripping, and ask where the person will be in the next frame… a little creativity and you can leverage llm for lots of things besides making text that appears smart. there’s quite of bit encoded in these things.
I’ve only used GPT to test the overall economy at the sector level. It’s only quality data. and got satisfactory results *When making the prediction, I didn’t want the prediction result. I need an explanation of the cause and effect that affects the securities. To be used to assess problems that will occur to the organization rather than Numerical data If there is enough information required I mostly do it myself.
If you have numerical data and still try in GPT You could make it quality data first, it might help.