The Quantile Trick

When you're doing a regression you're sometimes not so much interested in predicting the most "likely" value, sometimes you're more interested in predicting a spectrum of likely values. Put differently: you may be interested in predicting the quantiles of a distribution, instead of the median value. In this video we'll explain the quantile trick, which involves a pinball loss, to deal with these situations.

Video Chapters:
00:00 Drawing a dataset
00:58 Predicting Quantiles
03:04 Pinball Loss
07:06 Interactive Demo
09:45 Comparing Models

If you're interested in drawing data yourself, check out this project:

The code for all of our videos can be found on this Github repository:

The code for this specific episode can be found here:

This whiteboard video is part of the open efforts over at probabl. To learn more you can check out website or reach out to us on social media.

We also host a podcast called Sample Space, which you can find on your favourite podcast player. All the links can be found here:

If you're keen to see more videos like this, you can follow us over at @probabl_ai.

#probabl

[Spinnen]

Just wanted to say I love these kind of videos - early days for the channel just wanted to give that positive feedback to continue

[hernanebraga]

Excellent video, I really like how you explain complex concepts and techniques

[patrickleduc]

Great videos! Super helpful and I learn something new all the time. Keep up the good work!

[andreshoyosidrobo]

Great video! I think you nailed the explanation. It is nice to see how to use Jupyter widgets as tools to explain ML-related concepts.

[GarveRagnara]

But what about conformal predictions? :D

[alexmolasmartin]

how do you go from the free parameter of the pinball loss to the quantile you want to predict?

[mikiallen]

thanks sir, great exposition, however am wondering as to how one go about adapting such more dynamic approach to a case where the values are time series values are non stationary (main properties change over time , i.e. mean, variance and covariance ) on top of that such type series exhibit heavier tails i.e. probability of having very large values or very small values tend to be higher relative to normal distribution, so your input is highly appreciated
keep up the good work

[jdt]

How does the w parameter map to tge quantile?

[nicolasr]

It would be interesting to compare to Bayesian linear regression.