SentiTrac

Last edition talked about the Sentitrac Beta Launch, and this edition will continue the discussion of Sentitrac, but this time mainly focused on looking forward to what's next. If you missed the last edition, Sentitrac is my current side project, which is the culmination and commercialization of several years of research I did for my master's thesis, completed in 2022 (you can see the full thesis document on my profile). Sentiment analysis is still relatively new and is brand new to the sports world, and that is what Sentitrac is all about, bringing sentiment data to the sports world.

The beta was primarily focused on establishing the foundations of a business on the web:

• Set up authentication.
• Set up a payment portal.
• Provide the most basic sentiment data to have "something" up on the site.

That's what I showed in the previous edition, simple point-in-time sentiment scores for teams and players across the four major US sports leagues (NFL, NBA, NHL, MLB). Well, the beta's over! While I learned a lot during this period, we're just getting started.

Sentiment-infused fantasy projections are just the first product we've launched, and they are still quite basic. Sentitrac's core strategy is to continuously enhance our projections to make them more accurate and easily leveraged by fantasy players. Today, we compute projections by modeling together a set of existing projections and a sentiment grade for that particular player. All aspects of that last sentence are elements we want to experiment with further:

• The model itself (as in, the type of model) can be played with! So far, we've used an elementary neural network.
• The existing projections can be expanded, and we can start computing the accuracy of various projection sources to weigh their projections accordingly in our model.
• The sentiment data can be improved by retrieving more data about players from the internet: whether it's more tweets or data from more sources (IE, Reddit, blogs, etc.)?

In addition to projections themselves getting better as Sentitrac matures, many other features will make life as a fantasy player easier and hopefully more profitable! For example, if you are a daily fantasy player, you'll likely want lineup optimizers that leverage our projections. That's high on our feature list, including mechanisms to create multiple, ideally uncorrelated, lineups that maximize your chances of winning. For year-long leagues, features that analyze the free agent market in your league to provide recommendations about potential roster moves you could make are also high on our list to build in the next few weeks and months.?

To return a moment to how sentiment data can be improved, let's quickly talk about how sentiment grades are computed today:

1. We retrieve news data about a player, pulling the latest stories from ESPN and CBS Sports.
2. We search Twitter for the player's name and collect the first handful of tweets ranked the highest by Twitter. We take each of these pieces of text and run them through sentiment analysis, generating a sentiment grade between 0 and 100 (50 being neutral).
3. We aggregate all these scores into a single sentiment grade for that entity by first averaging all the sentiments just fetched together and then updating our sentiment model for that player with this updated point-in-time sentiment score.

We do this regularly and update our projections accordingly.?

In the future, we may also allow individual users to adjust projections based on their sentiment. I've often been frustrated when putting together lineups when I want to drop the projection for a particular player because I have a bad feeling about their upcoming matchup. Yet, my only options are not to include them in my search or live with the projection I feel is inaccurate. Soon, Sentitrac will allow you to adjust a projection based on your thoughts. Over time, I hope this becomes its own data source for our overall sentiment models, as this is the most direct wisdom of the crowd mechanism I've thought of so far!

As is hopefully clear, this is just the beginning. Perhaps I'm writing this more for myself and as a reminder that we're still early because there's just so much to do, and it can get quite overwhelming at times, especially since I'm doing all of this part-time outside of my regular day job at LinkedIn. That said, I know deep down that there's an appetite for what I'm building, so at the very least, I'll be my own number-one user.