Did you ever try extending it out to other methods of probability estimation other than the forms of regression? I have only skimmed your excellent article, but I think you are first calculating the average probabilities from a regression model and then minimizing the loss to calculate Harville corrections for place and show markets? Is that correct or am I missing something here? I guess I am curious if there has been any improvement on using regressions for combining the various initial odds as I don't really follow the literature anymore.
Yes! There have been big improvements since then but they are beyond the scope of the post. I just wanted to reproduce the calculations in the paper using PyTorch.
Bill Benter subsequently replaced the multinomial logit model with a multinomial probit model, which assumes Normally distributed errors rather than errors that follow the Laplace distribution.
Sorry - to be clear this is just re-running the model detailed in Bill Benter’s 1995 paper (he uses the time period 1986-1993) on more recent time periods (1996-2003, 2006-2013, 2016-2023) using PyTorch.
Nice work. Have you made available the CSV of this data, or is it easily obtainable using horse racing information online? I didn't see how that HKJC site offered it in CSV format. Perhaps you scraped the data to generate your CSV? Thank you.
Please read the first paragraph of the post again. The original author of the paper is Bill Benter, and the GP is the author of this excellent writeup.
I'm thinking of improving my Cantonese by reading a 金庸 novel like 倚天屠龍記. Hoping to find an online resource similar to this website that has the traditional Chinese characters, Jyutping romanisation and translation all aligned, along with the spoken audio as well.
I’ve read a lot of people griping (quite validly) that Spotify doesn’t distribute each subscriber’s monthly fee according to the streams they listen to that month. Does anyone why they don’t do that? Is it complex artist deals or difficulty in computing attribution at scale?
I don't have any insider knowledge, but I can't imagine this to be a technical limitation.
I think this is the outcome of contract negotiations, with big labels being shareholders of Spotify the power balance is not in favor of the smaller artists.
Edit:maybe there is some overhead in paying so many smaller artists
It's not too complex for them to do it because they used to do that. The problems is scammers caught on to it and abused the system. They'd create music and get it listed on Spotify, and then create fake listeners that would listen to their music non-stop and then get paid out as an artist with a lot of listens. After getting tired of playing whack-a-mole, Spotify changed the system to what it is now.
The gist is that your listening hours for your subscription are pooled with everybody else’s, then artist payments are made out of ratios from that pool. That means if you subscribe and listen to music 24hrs a day, and somebody else subscribes and listens to 1hr a day, you essentially have more influence over the pool of money and so can pay out to the artist you are listening to more than you pay Spotify.
Agreed, at best they could “make” is the cost of the subscription minus Spotify’s fee in this system. So I put $10 in and get $7 out (making up numbers). For money laundering it could make sense but I agree, nothing else makes much sense.
See my comment as a sibling to yours for a link, but Matt Levine made a post about how the pooling seems to work at Spotify. All the subscription money is pooled together, and all listening hours are pooled together. Artists are paid out at the ratios of everybody’s listening hours pool, not just each individual subscriber.
This means what individuals pay doesn’t matter at all to the artists they listen to. All you need to do is rack up a lot of listening hours, to have an outsized impact on the resultant listening hours pool. Then if enough other subscribers barely listen to any music, you can earn more than the $10 you put in.
I get that a locked phone needs to have everything already in memory, but what technical hurdles are stopping Apple from making a locked phone as secure as a rebooted phone?
In the BFU state, notification previews, contact information for incoming calls, and other user-specific data is locked because it’s not decrypted. These things would also change the user experience dramatically, so that’s why Apple doesn’t do it.
