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

Prophet: Automatic Forecasting Procedure

github.com

Post reply on HN
Print view View original

51–60 of 90 posts

Re: Prophet: Automatic Forecasting Procedure

#51

Post by cjalmeida »

ImaCake wrote: post #47
rdli wrote: post #30

As others have pointed out, Prophet is not a particularly good model for forecasting, and has been superseded by a multitude of other models. If you want to do time series forecasting, I'd recommend using Darts: https://github.com/unit8co/darts . Darts implements a wide range of models and is fairly easy to use. The problem with time series forecasting in general is that they make a lot of assumptions on the shape of…

> they expect that your data comes at a very regular interval Does prophet rely on this assumption? For health timeseries data the tool of choice is survival analysis - typically using Cox proportional hazards regression or similar regression tools that are able to handle irregular or censored data. I've seen some moves towards using fancy bayesian or fancier machine learning stuff for clinical trials but a big issue…

I find Bayesian regression models are actually simpler to explain as the assumptions you make are explicit and part of the model specification.

(Thought the actual sampling mechanics and tooling can be much more complex)

Re: Prophet: Automatic Forecasting Procedure

#52

Post by chubs »

naijaboiler wrote: post #48
chubs wrote: post #45

On this topic, does anyone know of a suitable time-series forecaster for multivariate analysis? Eg 8 independent/input variables, and one output variable? I've been using multiple linear regression (which works impressively!) but it doesn't take into account the time series, only the single prior day of inputs. Thanks :)

i use GLM with my own hand-crafted features based on my knowledge of the business and the things that influence it. Works very very well

Thanks! I take it that this means a Generalised Linear Model? Could i ask for a link to a relevant article to get me started on the flavour of GLM that you recommend?

Re: Prophet: Automatic Forecasting Procedure

#53

Post by rdli »

ImaCake wrote: post #47
rdli wrote: post #30

As others have pointed out, Prophet is not a particularly good model for forecasting, and has been superseded by a multitude of other models. If you want to do time series forecasting, I'd recommend using Darts: https://github.com/unit8co/darts . Darts implements a wide range of models and is fairly easy to use. The problem with time series forecasting in general is that they make a lot of assumptions on the shape of…

> they expect that your data comes at a very regular interval Does prophet rely on this assumption? For health timeseries data the tool of choice is survival analysis - typically using Cox proportional hazards regression or similar regression tools that are able to handle irregular or censored data. I've seen some moves towards using fancy bayesian or fancier machine learning stuff for clinical trials but a big issue…

I tried Prophet via Darts, and all the models in Darts assume a regular time series.

Re: "fancier machine learning" -- I've seen different flavors of RNNs & LSTMs have some success in analyzing time series data. I've struggled to get them to work on real-world (i.e., messy) data, but have had some encouraging results with a transformer encoder-only NN.

Re: Prophet: Automatic Forecasting Procedure

#54

Post by scrlk »

beckingz wrote: post #42
simonhughes22 wrote: post #38

Wondering how many people are now downloading this and other libs like Dart and trying to do stock market prediction or crypto price forecasting. Most of the devs i know, myself included, have dabbled in coding up trading algorithms at some point in time.

It's the classic data nerd trap. "I'm pretty good at statistics and can predict things using software... I bet I could make money in the stock market" And then they realize just how hard it is.

I just hope they come across the 90/90/90 rule first: 90% of new traders lose 90% of their money within 90 days.

VTSAX and chill? :^)

Re: Prophet: Automatic Forecasting Procedure

#55

Post by xs83 »

Im a Data Engineer in a large consulting company and I have been incredibly impressed with AutoGluon for forecasting. You can build and train a model in around 10 lines of code and it frequently gets into the top 3 or 4% of competitions on Kaggle without much data pre-processing

Re: Prophet: Automatic Forecasting Procedure

#56

Post by XTXinverseXTY »

3abiton wrote: post #44
mochomocha wrote: post #41

Earlier quoted context omitted.

Fun fact: if you don't care about the auto-regressive aspect of NeuralProphet (it's turned off by default), you can implement the core of NeuralProphet/Prophet (piecewise linear trend + Fourier on weekly/daily seasonality) in about 60 LOCs with no other dependency than either torch or numpy+scipy.optimize, and without having to deal with Stan or the very poorly chosen heuristics of neuralprophet. Another thing that b…

Why is STAN viewed negatively in this light? I am curious why bayesian libraries are the black sheep.

I think he just means that it can be an incredible pain to install.

Re: Prophet: Automatic Forecasting Procedure

#57

Post by kimukasetsu »

As others have pointed out, it is a good idea to encode domain knowledge in your time series model through specification and priors. Prophet rarely beats a well specified GLM or SARIMA in real world applications, especially when uncertainty estimates are needed. Professionally, I have succesfully applied Gaussian Processes to many such cases.

A GP is an intuitive and expressive way to code time covariance in a model. A famous example is the relative birthdays model, discussed by Gelman et al in Bayesian Data Analysis and here [1].

[1] https://avehtari.github.io/casestudies/Birthdays/birthdays.h...

Re: Prophet: Automatic Forecasting Procedure

#59

Post by mochomocha »

Donald wrote: post #43
mochomocha wrote: post #41

Earlier quoted context omitted.

Fun fact: if you don't care about the auto-regressive aspect of NeuralProphet (it's turned off by default), you can implement the core of NeuralProphet/Prophet (piecewise linear trend + Fourier on weekly/daily seasonality) in about 60 LOCs with no other dependency than either torch or numpy+scipy.optimize, and without having to deal with Stan or the very poorly chosen heuristics of neuralprophet. Another thing that b…

Do you have an example implementation of reimplementing the core of these?

It's literally what I did at work last week, which is why I found this submission timely. I'd have to check with my employer if it can be made public. I don't see any reason why not, there's not much to it.

Re: Prophet: Automatic Forecasting Procedure

#60

Post by carlossouza »

cjalmeida wrote: post #51
ImaCake wrote: post #47

Earlier quoted context omitted.

> they expect that your data comes at a very regular interval Does prophet rely on this assumption? For health timeseries data the tool of choice is survival analysis - typically using Cox proportional hazards regression or similar regression tools that are able to handle irregular or censored data. I've seen some moves towards using fancy bayesian or fancier machine learning stuff for clinical trials but a big issue…

I find Bayesian regression models are actually simpler to explain as the assumptions you make are explicit and part of the model specification. (Thought the actual sampling mechanics and tooling can be much more complex)

PyMC or Pyro/NumPyro make the implementation of Bayesian regression dead simple
Post reply on HN