← All videosTHE SHORT VERSION
Explore NVIDIA news sentiment ↗

machine learning / stock market / algotrading

Can News Sentiment Predict NVIDIA Stock? I Tested It

I rebuilt Mariana Antaya’s NVIDIA price regression and checked it against a simple baseline that predicts no price change. Then I added historical news sentiment from NexusTrade and tested how much it helped with forecasts 63 trading sessions out. The video shows how NexusTrade scores headlines, what the sentiment chart picked up when the Arm deal collapsed, and why the flat forecast won in the later holdout.

Transcript

0:00Let's build a machine learning model to predict the price of NVIDIA because it is not looking good Can historical

0:05data and volume actually be able to help us predict what NVIDIA's price target will be in three months from now A

0:12really good feature to also include is actually sentiment

0:14Let's fact check this video Just so there's no ambiguity

0:18you guys know that this doesn't work right Daily price and volume data is some of the easiest data you can find and

0:26every single intro to CS student's first project is trying to build a machine learning model for stock market

0:33prediction I reproduced the findings from her code exactly If you rebuilt the regression model as she describes then

0:41it has an error of around 20 percent.

0:45To put that into perspective

0:47if you simply guess NVIDIA never moves it's flat all the time that error is only 15 percent So what did we do all of

0:57this work for We can just guess the stock will never move With that being said she does mention something that I

1:03think is extremely interesting and it's something that I personally wondered for a long time Can we use historical

1:10sentiment data So here's what I did What I didn't do was launch an army of different web scrapers and scrape the

1:19entire internet This wouldn't work for several reasons Number one we can't go back in time so it's really hard to

1:27find the articles when they were published originally Number two it'd be extremely expensive And number three I

1:35didn't want to have to implement going around bot protection getting

1:38around Cloudflare going across all these websites That seemed like a tremendous data project for very low value So

1:46here's what I did instead I used a historical archive called GDELT and it basically contains an index of news around

1:55the world I trained several different machine learning models to see which one was the best at classifying the

2:04headlines I found that Microsoft's DeBERTa was the best model that I tested

2:09so then I fine tuned it for sentiment analysis That way based purely on the headlines this model could tell us

2:16whether or not the headlines were good or bad As a sanity check

2:21I tested to see what were NVIDIA's worst headlines and I thought the results made sense The most negative week in

2:30our archive was when the deal with Arm fell apart and NVIDIA was forced to spend 1 35 billion to terminate its

2:40agreement This week also shows the dramatic decrease in sentiment and it shows the investigations on NVIDIA and

2:49Jensen Huang Then for the final test I added these new scores to the regression and I expected it to do a little

2:58better and I was completely wrong I tried everything from average 1 week sentiment to average 13 week sentiment and

3:09nothing worked even a little bit Guessing NVIDIA would never move beat every single model This is your friendly

3:18reminder that what sounds like what might work doesn't actually matter It can sound good on paper and not translate

3:26to real world results Follow for more

Join the conversation

Loading conversation…