I am trying to understand performance of the example from https://www.anyscale.com/blog/how-nixtla-uses-ray-to-accurately-predict-more-than-a-million-time-series. It is using a m5.2xlarge instance with 8 cores and the min worker in the yaml file is set to 249. It then says that the deployed cluster is using 2000 CPUs. I don't understand this. Does it mean it spins up 249 EC2 instances of m5.2xlarge with each of them having 8 cores? That seems unlikely to me but I don't get it how it is using 2000 CPUs. I need to understand this because I am currently predicting only 10 time series on a c6a.8xlarge with 32 cores and it's taking 5 minutes! If a million series take 30 minutes, then I think 10 series should take a second. I am confused how that performance was achieved. Any insight please?

Hi, I just wanted to say that in https://nixtla.github.io/statsforecast/arima.html we can see that

imagen.png

but in https://nixtla.github.io/statsforecast/models.html#autoarima we can see that

imagen.png

but in https://github.com/Nixtla/statsforecast/blob/main/statsforecast/models.py#L63 we can see that

imagen.png

although the final truth is that

When using AutoTheta should we remove seasonality beforehand and add it to the results afterwards or does AutoTheta handle series with seasonality?

StatsForecast has an option called fallback_model where we specify what model to use if the main one fails (e.g. SeasonalNaive). How do we know after the forecast is done where it used the main model and where it switched to the fallback model?

When we apply a model to let's say 100 time series using StatsForecast, does it come up with one fit for all of them or 100 different fits? So for example in case of ARIMA, is it trying to find the best p,d,q that is optimal overall for all the series or does it fit to individual series and calculate 100 p,d,q pairs?

I have observed that with AutoCES if I provide a ray_address, it fails to fit but when I remove ray_address it works fine. With AutoARIMA and AutoTheta, this is not an issue. They both work fine with or without using ray_address. Not sure if this is specific to my data or not. I can't share the data to replicate the error but I decided to share the anecdote.

What is the best way to find out in which version models were added (specifically interested in AutoETS, AutoCES and AutoTheta)? And also, from which version these models could produce prediction intervals?

@fede (nixtla) (they/them) any plans on migrating the tsfeatures repo to nbdev, would be great imho. mainly looking at docs and describing how these features come to be

Hola yo tengo una consulta, las librerias statsForecast y tsFeatures son compatibles con pyspark?

I fit AutoARIMA to my data and get the (p, d, q) (P, D, Q) out of it and pass that to ARIMA from the statsmodels package and get a different output with lower errors! Has anyone experienced this?

Hey there - I’ve been toying around with the different forecasting packages Nixtla has made available and I’m hopefully that we can start using it in production to forecast sales of our items. I have a question that I’m hoping I could get some insight on: how do you control for inconsistencies the amount of data available within a hierarchy? Over the course of four years, we have launched and discontinued both items and categories of items. In a category, we might have 30 items today, but only 20 two years ago. Similarly, we have a category that was only available for about six months until it was discontinued. I’ve padded the data with 0's so that every item has an equal number of records, but I expect this isn’t an ideal solution. Any tips?

Hello guys, I fitted a model with exogerous variables, and I can see the coeficients by doing model.model_, but I cant find the coeficients p-values for each model component, can someone help me?

Does StatsForecast change the default options of AutoARIMA? When I use AutoARIMA directly, I get better errors than when I use it with StatsForecast class. The only difference I see is that StatsForecast takes a 'freq' argument that doesn't exist in AutoARIMA. I expected to get the same results.

Amazon Fortuna launches conformal prediction forecasting https://aws-fortuna.readthedocs.io/en/latest/examples/enbpi_ts_regression.html

Are there any plans to add a TBATS model or to add Fourier coefficient option to ARIMA (like R has) in near future?

If you feel like it, please show some love.

Hello everyone, I finished fully annotating and updating my PyData NYC tutorial on Large Scale Timeseries Forecasting. It covers the following topics: 1. Using the statsforecast library to run statistical models on top of Spark, Dask, and Ray 2. Preprocessing for large scale data using Fugue on top of Spark, Dask, and Ray 3. A section on Hierarchical Forecasting with hierchicalforecast, contributed by @fede (nixtla) (they/them) and @Max (Nixtla) from Nixtla. 4. Running the training pipeline on top of a Dask cluster managed by Coiled, though the same setup will work on Spark, Dask, and Ray clusters. Happy to present it at any Meetup/event if you know of any!

is there any gpu support for statsforecast?

Hello Guys,iam facing problem with missing hours in time series, what is the best way to impute ? and is there any way add exogerous variables to auto arima?

Is it possible to interpolate missing values on a MSTL fitting? I have hourly sensor data over several years that has missing observations (missing at random due to sensor failure. Observations are missing in contiguous chunks up to 1 month duration). The data has a very strong daily and weekly seasonality. Fitting with the missing timepoints removed gives an ok fit for the existing data, but I'm not sure how to get the fitted model to interpolate the missing values. Any advice? Edit: Im trying to avoid the flat interpolation of croston etc, and have something that incorporates the seasonality

sorry if this has already been asked but is there a way to extract the entire test set feature matrix, including features added (such as rolling/lagged) when instantiating an MLForecast object? the preprocess method does this for the training set but I want to recover my test set with the aforementioned features added without having to do the calculations for my test from scratch. When I do

test_sample = model.preprocess(test, id_col='my_id', time_col='ds', target_col='y', static_features=[])

, I only recover the last sample of the test set feature matrix. TIA.