Forecasting at Uber: An Introduction | Hacker News Reader

Forecasting at Uber: An Introduction

eng.uber.com

123 pointskiyanwang10 comments

Very interesting post, especially with regards to the granularity Uber needs to forecast. This granularity on the time scale alone means that Uber is kind of forced to compress a pretty big part of there Sales and Operations Planning into a very short period of time.

I'm really looking forward to the follow up posts, espiacially the process is intriguing. Being an Ops guy myself I aleeady see now a lot of challenges and also fun in this.

Also nice note: Traditional taxis don't have that problem (in a positive sense). Having a learning and intelligent network of drivers reacting tobthese paterns and acting them allows them to achieve availability, in less efficient and less customer friendly way. Another advantage taxis is fixed pick-up locations (at least they exist in Germany). This allows traditional taxis to shape demand on the location level. By provding cuatomers a more convenient solution Uber now is forced to actively plan on both axis.

OT: I don't want to make fun of the authors, but I kinda had a laugh when I saw this code next to the authors pictures:

  <div id="sexy-author-bio" style="margin:10px 0;" class="slawek-smyl">

Maybe just not my type ;-)

Some frontend developers should really care more about the identifiers they use...

It's a WordPress plugin.

> Maybe just not my type

Says the guy inspecting Slawek's markup.

Does anyone know if uber or another on demand company has a similar type blog post, but about predictive matching? Algorithms like stable marriage are fascinating from a theoretical standpoint but I would love to hear about them from people who implement or deploy them!

I'm a little amused that their forecast visualization seems to repeatedly show substantial ride demand in the middle of the bay. Kayakers getting tired and hailing a ride?

I wonder if there are people who request a Uber from the commuter ferries right before they're about to dock.

Business meetings?

Interesting article but rather shallow. Still funny to see Uber also uses classical algorithms and some ML here and there, I mean, what did I expect otherwise? Probably some super secret algos no one would know about ... nope, just classical econometrics time series analysis and some ML sprinkled in there.

Gives me hope I could work there one day as an econometrics graduate :)

You absolutely could. I don't work at uber but work at another Big Tech company on a similar team. We have plenty of econometrics grads on our team. Forecasting is pretty resilient to hype methods, since extrapolating into the future is so hard, there is rarely some treasure trove of data to throw into an ML model to predict the future. The biggest drivers of uncertainty in these problems tend to be unforecastable, so data-hungry models don't offer much (depends on the horizon and specific problem, but this is true as a general rule).

I'd say my only observation has been economists who also meet the bar as software engineers are a force to be reckoned with.

Reply on news.ycombinator.com