I have the feeling the reporter doesn't know what they're talking about:
"For example, if the Raptors were measuring college basketball prospects, Watson could quickly crunch the numbers and display a comparison of their stats on shooting, assists, and rebounds."
One most assuredly does not need Watson to do this.
Moreover,
"Compare that to drafts of past years, in which the Raptors would use whiteboards with player stats printed on magnets, and call up statisticians each time they wanted new information"
Does not seem it could be remotely true, given...
"Even before Watson came along, the Raptors were tech-savvy. They have access to one of the NBA’s leading analytics teams, and have developed a wide range of tools, including a way to use data from the SportVU camera tracking system to model the best moves a player could make."
I suspect the problem is not entirely with the reporter. The reporter is probably paraphrasing a press release describing a really minimal effort to apply Watson to a something that would grab some publicity. IBM and the Raptors' PR people probably put in more work than was done on the implementation. Which is why it sounds like something you could whip up in a spreadsheet. It probably does "display a comparison of their stats on shooting, assists, and rebounds" on a Web page with the Watson logo on it, and the reporter faithfully reported what he saw.
Watson is more of IBM's tool for creating the impression that IBM is a technology company and not just a systems integration shop than it is a serious deep learning project. Eventually it will become another tragic chapter among many in the decline of technology at IBM.
I predict sometime in the near future, when Apple and Google and Microsoft and others build some real businesses on deep learning systems that there will be a mournful article published about how sad and left behind Watson has become.
I'm don't think that is necessarily a conflict. The reason sport teams have these analytics teams is to help decide how limited resources should be utilized. Those limited resources include money and draft picks, but it also applies to man hours. Why spend development time to build a digital whiteboard to present that data when it would take mere minutes to print it out and put it on a whiteboard? Why build a front end for the database when you can "call up statisticians" (probably just a DBA running an SQL query) to give you the answer. It is a game of trade offs. If the high tech and more complicated solution takes time to implement, the improvement over the simpler solution needs to be large enough to justify the investment.
I've spent more time than any sane person should trying to model fantasy sports for the purpose of making money as a psuedo hobby/second job.
My day job is essentially practical applications of simulation and machine learning so this was a natural way to broaden my modelling capabilities.
I'd love to chat with anyone who has any specific insight as to what the raptors are doing. If you're in Toronto I'll buy lunch!! Contact info in my profile.
Specifically a focus on methods applicable to selecting teams for Fantasy Hockey and Football with applications to weekly/daily fantasy is mostly what I'm concerned with.
I've spent alot of time/money figuring out what doesn't work very well so I can offer 3 years of failed experiments as a trade:)
They are using the "Watson Tradeoff Analytics" product as the underpinning [1]
As others have mentioned, "Watson" is a brandname/pillar at IBM, similar to Infosphere (anything information management), Websphere (anything middleware-y), etc. It's a bunch of simple and complex, in-house-built and bought products, some of which play together well some of which don't, some of which were pretty impressive and some of which aren't.
To play with Watson Tradeoff Analytics, you can check out the documentation or get a free account with Bluemix [2]. They used to have a subset-of-functionality demo for Raptors on Bluemix as well. It seemed nifty but not groud-breaking.
My understanding is that for the real-thing, they also used the Watson Tone/Sentiment analyzers to see if players would be culturally a good fit, i.e. get along with their team-mates, not just whether they're good on paper.
[disclosure: I work for IBM... nowhere NEAR Watson, in the plain-ol'-ERP department, but I've been curious myself to figure out what the hoopla is or isn't all about]
> My understanding is that for the real-thing, they also used the Watson Tone/Sentiment analyzers to see if players would be culturally a good fit, i.e. get along with their team-mates, not just whether they're good on paper.
This seems reeeeally unlikely, considering both the inaccuracy of Watson sentiment analysis, and that one would need to collect a rather substantial personal writing sample from each candidate.
The article mentions that they mine social media accounts of the players for the text. That could easily qualify as a substantial personal writing sample :)
That's great for college players, but not necessarily good for pro players, many of which don't control their twitter output (or have to sensor themselves much more than a college player would).
I do some analytical work for a high school football team as well as play a good amount of fantasy football. What would be a good first step to actually making money on fantasy football?
If it's your thing, you could look up Haralabos Voulgaris - he takes on people doing advanced sports stats for betting purposes. He is straight gambling ($1m/day I think he's mentioned) rather than fantasy sports though. From following him for a while, I've had the impression that he has a team of programmers working for him at various points.
He does and he's a very smart person. He would pretty much be the go to guy I'd talk to about anything related to applied analytics and sports betting. It's fun to see him debate sports books that claim they will take "any bets". The biggest problem seems to be getting volume once your model works. You're basically limited to the NFL, NBA and soccer. I believe there are good/winning models for college basketball for example but you can't wager enough money to make it worthwhile.
Personally I think the juice is so high that I won't bother with anything sports betting related other than having a fun exercise/data to play with. Daily sports is way too scummy an industry (used to be a semi-pro online poker player so been there, done that) and the juice is very high as well but I suppose there could be some temporary money in it as there's enough amateurs around. If you want to do that from an ethical POV is another question (one of the reasons I quit poker) but either way the field will get more competitive rather quickly and/or regulated/legislated away.
tl;dr: Don't invest too much time in sports betting other than as a fun hobby with a nice dataset.
I love it, because two things are abundantly clear to even the most casual observer:
1) everyone at IBM has been told to pretend they are this huge, leading "cloud" company
2) 1) is a clear falsehood
And so you see marketing struggle to come up with all these ridiculous "case studies", because it's difficult if you have to invent them out of thin air instead of real customers.
As someone who works for IBM I'd have to disagree with you there. No one has ever told me we are a huge cloud company, I mean obviously I've been told we're investing buckets of money in cloud technologies and that we have made it a key strategy point to focus on growth in the area, but not once has anyone even intimated that we 'are [a] huge cloud company'.
What IBM is, realistically, is a ridiculously large (and I actually mean ridiculous in the sense of 'how are they not under more anti-trust investigations because they seriously own roughly everything") company with an unbelievably large number of very vert wealthy clients who are scared beyond reason of losing their market share if they don't start using <Insert "new" technology here> so internally everything that happens and looks like a publicity stunt is usually met with a response of "Well no shit we could do that. Why the hell didn't our marketing team point out that we had that capability in the 90's when it would have been impressive"
A lot of the Watson hype seems to actually be because a lot of the companies who traditionally would say 'hm, I dunno about this whole analytics thing. Shouldnt we just stick to spreadsheets and man hours' now have a named thing they can buy that really just covers up that they are actually just investing in a platform with some hadoop clustering and some racks full of P8s (obviously theres a little more to it than that).
The fact that watson can do seemingly cool stuff is actually just a nice way of saying 'anyone can do cool stuff if they invest a bit of cash in technologies that have been around for ages but large companies now have a small window where they can start investing in these technologies and they even get to act like they are an early adopter, of course sans-risk'
I find it amusing that our sales pitch to a rugby team in australia to sell them analytics tools and services failed about 4 years ago, so we gave it to them for free then said we would tell them who the next 10 players to get injured would be and how they would be injured. After the predictions were exactly correct at about injury 6 they were suddenly very interested in buying the stuff. Interestingly, if we sold the same thing now it wouldn't be an analytics platform, we would probably use watson (read: not rebuild that thing, just use the existing stuff because its easier) and it would be marketed as a "watson solution".
Because it's so big, there's a lot capable engineers, but also deadwood. I think Watson has attracted many capable devs. There's definitely too many execs flailing around with buzzwords.
It's a great brand, it has potential. The problem is that a) it's being overhyped and b) yet again IBM doesn't understand that if you want it to be used for truly mind-blowing stuff, you have to make it easy for developers to use (or maybe they do, but suck at making it easy to use, see [1]). Then it generates its own, real hype.
For an example of this, I'd say MQ vs MQTT/RabbitMQ, DB2 vs Postgres et al., NoSQL, analytics, Bluemix vs Amazon/GCE/Heroku, Softlayer vs Amazon/Google/Azure. I get that selling to traditional, conservative, and very rich/big firms is a viable strategy - but one that's failing for e.g. Bluemix, Softlayer.
---
[1] The Google search results for "Watson API" suck. The documentation sounds like it was written for managers (" Representational State Transfer (REST) Application Programming Interface (API) "). Stop getting me to use Bluemix, Amazon/GCE/Heroku work better. Just look at this mess: https://www.ibm.com/smarterplanet/us/en/ibmwatson/developerc...
The scary part that is that my company's IT department also talks about "Watson integration" and a lot of execs are completely impressed. I have asked in some meetings what this would exactly do and how it's better than a few SQL statements but got totally brushed off.
I think it's great that there continues to be a stream of "Watson partners with X" or "Watson is used for X" articles, but in actual usage, it feels like Google is winning the ML race. I feel like corporations are using Watson for easy PR wins. Is that just me living in my own development bubble?
My limited understanding with basketball stats is that they are only somewhat useful, as it favours player who have direct impact to the game, those who can score/assist/rebound/etc, and encourages players to game the system.
However it does not reflect how a player defends, how he fits into a particular style/role, how much support his teammates get from him other than assists, etc.
It'll be interesting to see if they use deeper analysis to figure out which player is actually better.
There are actually many advanced statistics available to NBA teams via SportsVu. It can measure who passes to who, how far people travel on a possession, gravity (how close opposing defenders are drawn to the offensive player, how much someone dribbles, and a host of other stats. I don't think the system is in place at many colleges (if any).
The tricky thing with the draft is finding players that have a high ceiling and can perform far in excess of what they were able to do in college. Sabermetrics can help you identify that some of the time but I think attitude, ability to improve specific skills, and innate physical attributes are things a team would have to evaluate in person to uncover undervalued talent.
There are many invented advanced statistics that supposedly fill those intangible gaps. For example, the defensive win shares stat is supposed to measure defensive impact.
As per my other comment, I _believe_ that for the actual application, they did use Watson Tone Analyzer / Sentiment Analyzer / etc, to figure out the "fit" of a player to Raptors culture and with other team-mates.
I've seen a demo where as part of the app, they scrape player's interviews, twitter feeds, etc to get a personality representation.
There are more stats than this, average points participated, win participation, time with ball, all of those are normalized for amount of games played and time on court.
But at the end those stats you've mentioned matter, hoops scored, rebounds caught, penalty shots, and score passes/assists are what "wins" a game.
An important part of a statistical model is to also be simple and cheap enough to implement to gain actual statistics, the chance that there is some magical player that doesn't score high on stats but somehow has a meaningful impact on the game that isn't somehow scored is very very slim and in Basketball specifically probably non-existent since it's only a 5vs5 sport which means that every player handles the ball.
But it's unlikely to increase without increasing the stats of these "magical players", as they need to have an impact of the game to increase the stats of other players and that counts as assists, rebounds caught etc.
Basketball is a small court, 5 people on each team, defense is very important but this hardly is football/rugby or even soccer level of defense everyone handles the ball quite a bit.
Time with ball is also a stat that is counted, and often other stats are normalized to it.
Honest if potentially naive question: why can't you just measure points per minute by the player's team's opponents when the player is on versus off the court?
Because this doesn't guarantee a uniform background.
Let's say we have two teams, the Ayyys and the Bees; each has two defenders, Schlub and Superman play for the Ayys, and Scrub and Slacker play for the Bees. Let's assume that they all suck equally, except for Superman, who is great.
Now we run your statistic: we compare how the Ayyys do with Schlub on/off against how Scrub does on/off.
The problem is, when Schlub is off, Superman is playing, whereas when Scrub is off, Slacker is playing. This means the Ayyys do quite well when Schlub is off, but the Bees do about as well when Scrub is on compared to when Scrub is off. By this comparison, Schlub is a much worse player than Scrub, even though in actual fact they might be equivalent.
ESPN introduced a stat called real plus-minus in 2014, which adjusts for teammates and opponents. It also takes both offensive and defense plus-minus into account. The top two players in RPM this past season were Lebron James and Draymond Green, followed by Chris Paul and Steph Curry [1]. It was received with a lot of skepticism back when it was introduced, but that shouldn't surprise anyone.
Yeah, that occurred to me but I figured it would average out over enough games. It would be super bad luck to always get subbed in and out in sync with your opponents' best players all season.
FYI "+/-":
A metric that looks at how teams perform with a certain player on the court, how they perform with a certain player off the court, and calculates the overall impact that player has on team success.
>My limited understanding with basketball stats is that they are only somewhat useful
One of the major fallacies surrounding sports analytics is dismissing information that isn't all-encompassing. Not picking on you, but "those stats are no good because they don't capture X!" is a common refrain.
> For example, if the Raptors were measuring college basketball prospects, Watson could quickly crunch the numbers and display a comparison of their stats on shooting, assists, and rebounds. Compare that to drafts of past years, in which the Raptors would use whiteboards with player stats printed on magnets, and call up statisticians each time they wanted new information, recalled Lenchner, who visited the Raptors’ headquarters while IBM was developing the software. In the days before Watson, the whole process was much more laborious and time consuming.
I like Watson and it's not their fault for getting publicity...but as for the reporter, c'mon, there's no way you could write that above paragraph without being ignorant of computers pre-Macintosh (or iPod) days. Perhaps the Raptors are old-fashioned but computers have been used for exponentially reducing laborious and time-consuming activities since the dawn of the airlines for solving scheduling problems. And that was much later than the era of computing used for early censuses and cryptoanalysis during the World Wars. But if you've grown up in the "there's an app for that!" age, I guess it's easy to forget how infinite the use cases for computers.
I remember seeing Watson on Jeopardy years ago. He is not quite in the same league as Mr. Butlertron from Clone High, but he's pretty good all the same.
> “[Watson could be used] for the mission to Mars,” he said. After all, that small crew will be crammed together on a spaceship for a few years at least, and getting along will be essential. “You can’t make changes once they’re up there.”
"For example, if the Raptors were measuring college basketball prospects, Watson could quickly crunch the numbers and display a comparison of their stats on shooting, assists, and rebounds."
One most assuredly does not need Watson to do this.
Moreover,
"Compare that to drafts of past years, in which the Raptors would use whiteboards with player stats printed on magnets, and call up statisticians each time they wanted new information"
Does not seem it could be remotely true, given...
"Even before Watson came along, the Raptors were tech-savvy. They have access to one of the NBA’s leading analytics teams, and have developed a wide range of tools, including a way to use data from the SportVU camera tracking system to model the best moves a player could make."