Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Prophet: forecasting at scale

research.fb.com

Post reply on HN
Print view View original

31–40 of 114 posts

Re: Prophet: forecasting at scale

#31

Post by fpvracing »

rodionos wrote: post #16

I didn't know wikipedia page view counters are available for public usage. The wikipediatrend R package relies on http://stats.grok.se/ , which in turn relies on https://dumps.wikimedia.org/other/pagecounts-raw/ which has been deprecated. The new dump is located at https://dumps.wikimedia.org/other/pageviews/ Data is available in hourly intervals. * pageviews-20170227-050000 en Peyton_Manning 58 0 [edit] There is a w…

Cool! I wonder what spiked the views for artificial intelligence on 10/11/2016?

https://tools.wmflabs.org/pageviews/?project=en.wikipedia.or...

Re: Prophet: forecasting at scale

#32

Post by nodesocket »

throwaway_374 wrote: post #25
nodesocket wrote: post #24

Earlier quoted context omitted.

Is it possible for example to send you monthly revenue numbers for my startup for the last two years (24 data points) and have yoghurt predict the next two years of monthly revenue?

If the model is autoregressive you can only forecast N steps ahead. Any further forecasting will be based on these generated near-future forecasts. In English, no. See https://www.youtube.com/watch?v=tJ-O3hk1vRw#t=01h16m

Thanks for posting this talk by Jeffrey Yau. I am 9 minutes into it and can't stop watching. He explains things very easily and clearly.

Re: Prophet: forecasting at scale

#33

Post by etjossem »

blazespin wrote: post #3
asafira wrote: post #2

So...How much will this do at forecasting stock prices? =) Very cool though --- I would be interested to dive into the methods they've implemented sometime in the near future!

Probably just help verify that the stock market is a random walk with a meager trend upwards that doesn't beat inflation + trading costs.

> Probably just help verify that the stock market is a random walk with a meager trend upwards that doesn't beat inflation + trading costs.

That doesn't sound right. Let me clear that up for you. Since 1950:

  S&P 500 Annual Price Change: 7.2%
  S&P 500 Annual Div Dist: 3.6%
  S&P 500 Annual Total Return: 11.0%
  Annual Inflation: 3.8%
  Annual Real Price Change: 3.3%
  Annual Real Total Return: 7.0 %
Buying the straight S&P 500 beats inflation by seven percent, on average, every year. You're welcome!

Re: Prophet: forecasting at scale

#35

Post by techno_modus »

It seems that they have developed a model for only univariate forecasts and only numeric regular time series which is a classical use case in statistics. Yet, most data sources have many dimensions (for example, energy consumption, temperature, humidity etc.) as well as categorical data like current state (On, Off). The situation is even more difficult if the data is not a regular time series but is more like asynchronous event stream. It would be interesting to find a good forecasting model for some of these use cases. In particular, it is interesting if this Prophet model can be generalized and applied to multivariate data.

Re: Prophet: forecasting at scale

#36

Post by fgpwd »

nodesocket wrote: post #29
yoghurtio wrote: post #26

Earlier quoted context omitted.

Its very simple to use Yoghurt, just upload the data and rest it does automatically. 24 data points is less to make any accurate prediction. You need more data points. However, Yoghurt currently supports 1 week prediction only and very soon we will be adding prediction upto 1 Month and plus.

Excuse my ignorance, but how does 1 week fit into the equation? Why does the time scale (x-axis) matter? I.E. if I pass 180 points of revenue (y-axis) does it matter if they were sampled each day or each hour in terms of forecasting?

They probably take into account day-specific trends , such as if the data shows sales are usually lower on a Monday than a Tuesday, they would take that into account in the forecast. This is as far as I understand.

So, assuming they are doing this, the time scale does matter. What I am trying to say is that these solutions (like prophet) are opinionated and that is why they can get accurate, as they are taking into account these time-scale specific trends.

But being opinionated means that they are assuming stuff about your data. For example saying that the number of sales you make in a day is a function of or correlated to the day of the week is probably a reasonable statement. However if you move away from sales and marketing, and try to forecast say the number of seismic events in a day, nature doesn't care if it's a Monday or Tuesday or holiday. So any such correlation the program is able to find out and use in forecasting would be incorrect. Like maybe there are more earthquakes on Monday than any other day in a particular dataset, but that would just be incidental and doesn't mean earthquakes are more likely to occur in future on Mondays. It's not a good example but there could be other such cases where such assumptions could be wrong.

Re: Prophet: forecasting at scale

#37

Post by dragandj »

confounded wrote: post #11

Worth noting Prophet is R/Python wrappers to some models with reasonable defaults, written in and fit by Stan , a probabilistic programming language, and Bayesian estimation framework. Stan is amazing in that you can fit pretty much any model you can describe in an equation (given enough time and compute, of course)! More on Stan here: http://mc-stan.org/

... and if you like Clojure, you might try Bayadera, which has its own engine running the analysis on the GPU.

http://github.com/uncomplicate/bayadera

Re: Prophet: forecasting at scale

#38

Post by hnarayanan »

rodionos wrote: post #27
hnarayanan wrote: post #10

Is there a way to extend these models to handle spatial variation (e.g. weather forecasting, property price estimation etc.) as well?

This would be non-trivial. Consider this paper on marijuana usage where the researchers had to group statistics by adjacent counties in Oregon and Washington in order to control the tests. https://papers.ssrn.com/sol3/papers2.cfm?abstract_id=2841267

Thank you for the pointer, will read the article.

All my attempts thus far have pointed me to something called Gaussian Proceeses that I am still working through grokking.

Re: Prophet: forecasting at scale

#39

Post by dandermotj »

matheweis wrote: post #5
blazespin wrote: post #3

Earlier quoted context omitted.

Probably just help verify that the stock market is a random walk with a meager trend upwards that doesn't beat inflation + trading costs.

> the stock market is a random walk with a meager trend upwards that doesn't beat inflation + trading costs. That assumes that the efficient-market hypothesis holds true, but it has yet to be thoroughly proven or disproven... (and funds like Medallion would strongly suggest otherwise for the medium term: https://www.bloomberg.com/news/articles/2016-11-21/how-renai... )

It doesn't assume the Effiecient Market hypothesis - empirical studies of returns support random returns without the imposing a model (non-parametric tests).

That's not to say returns are actually random, but in any given time range, it appears to be.

Re: Prophet: forecasting at scale

#40

Post by fgpwd »

recurser wrote: post #30

Very cool. Could this be re-purposed for detecting anomalies/outliers in time series data?

My guess would be yes. I'm thinking this could be used to find out how effective a particular marketing campaign was. Just compare the forecast with actuals and the difference would be the number of sales/clicks you got from that campaign.
Post reply on HN