Prediction prompt - is it really a prediction (forecast)

Hi,

I have a question that bothers me. So I’ve found an example of HackerNoon article engagement prompt that works like a predictor:

I don’t really think that it works like a real prediction, it builds a correlation between a set of words in an article title and value in the brackets and that’s it.

So if I will construct a prompt like:

Month (total sales)
may (12312)
june(21412)
july(12312)
august (1241)
September (21431)

It won’t predict for October, right? It will try to follow the sequence, right?

Anyways, any comment would be appreciated.

2 Likes

Exactly! I’m not sure I’d use a language model to predict quantities, but it’s theoretically possible (in the same way that a pencil doubles as a chopstick).

4 Likes

But it’s not actually a prediction in the sense of being based on a causal model, is it? It’s a report of a correlation, in the sense of “I have observed that when numeric examples are presented like X, the next values are often Y.” Like a kid’s game almost, where you shout out the response in a rhythm. “Duck! Duck! Goose!”

If you give it enough context in the prompt, it can likely do some deduction. Consider all the thousands of articles and papers GPT-3 was trained on that discussed data and extracted insights from. This behavior is clearly embedded in the model. The only question is how sophisticated are those behaviors? Either way, this is not the kind of thing that I would rely on a single prompt to do. You should try extracting explanations from it - the Davinci engine is pretty good at explaining its reasoning.

Thanks, guys, for explanations. My plan is to run the experiment - we have performance statistics per item, and we will just run some test to see the results.

I also think that I will experiment with the context explaining to A.I. how the ranking was built, the product description and etc. And without any context.

This is really exciting!

Okay, I did some tests, results are not really good. Here is the data set:
https://raw.githubusercontent.com/jbrownlee/Datasets/master/shampoo.csv

It is some sample shampoo sales for 2001-2003 years. I’ve tried the predictor, but it does not go well (the shampoo sales were growing, A.I., from the looks of it, repeats the sequence of previous years.

This is the prompt:

Continue the table by predicting the sales of the product per month:
“Month”,“Sales”
“1-01”,266.0
“1-02”,145.9
“1-03”,183.1
“1-04”,119.3
“1-05”,180.3
“1-06”,168.5
“1-07”,231.8
“1-08”,224.5
“1-09”,192.8
“1-10”,122.9
“1-11”,336.5
“1-12”,185.9
“2-01”,194.3
“2-02”,149.5
“2-03”,210.1
“2-04”,273.3
“2-05”,191.4
“2-06”,287.0
“2-07”,226.0
“2-08”,303.6
“2-09”,289.9
“2-10”,421.6
“2-11”,264.5
“2-12”,342.3

Maybe I phrased it bad.

It works a bit better if I predict per month:

Continue the table by predicting the sales of the product per month:

The table with sales per month:
“Month”,“Sales”
“1-01”,266.0
“1-02”,145.9
“1-03”,183.1
“1-04”,119.3
“1-05”,180.3
“1-06”,168.5
“1-07”,231.8
“1-08”,224.5
“1-09”,192.8
“1-10”,122.9
“1-11”,336.5
“1-12”,185.9
“2-01”,194.3
“2-02”,149.5
“2-03”,210.1
“2-04”,273.3
“2-05”,191.4
“2-06”,287.0
“2-07”,226.0
“2-08”,303.6
“2-09”,289.9
“2-10”,421.6
“2-11”,264.5
“2-12”,342.3
“3-01”,337.9
“3-02”,346.6

My prediction is:
“3-03”,

Any comment would be highly appreciated.

1 Like

i found that phrasing is really important, similar how i use chess notation, that and something as low as babage could also work

1 Like