Thanks Max!

Hi all! I've some experience with sktime but now want to try out your stats/ml/hier-forecasts packages. I have two quick questions: 1. I know statsforecast can use fugue as a backend to work with spark which connects nicely with Databricks which is my cloud env of choice. Do the others (ml and hier) also support fugue/spark backend? 2. What would be the equivalent of the sktime evaluate() fun where you can input a df, a model and a CV strategy and get back some kind of results df? Thanks!

Hi all, is there currently an option to ensemble models for statsforecast?

Hey everyone, quick question about MLForecast: Is there a way to utilize LightGBM's categorical features: https://lightgbm.readthedocs.io/en/latest/Advanced-Topics.html . Specifically for the unique id column? Thanks!

hi all, i need your opinions about cross validation in general do we get a good benefit from that in time series, any advice we be helpful for me?

Hi, I am quite new to time series field and am starting a forecasting project. I would like to handle it in a proper way with good project management practice and communication to other people and also having a kind of template for future needs for colleagues. I have so many questions, but I'll try to keep it simple to start (please note I have been through the documentation for Statsforecast and MLForecast): Context : A previous POC has been made to forecast the monthly company revenue (Univariate TS, Around 230 months history, 10y history, forecast horizon 24 months) using Holtwinter (tuned parameters). Libraries used where pandas and statsmodel. 1. In terms of forecasting generally speaking, what is the good process? Below my understanding, but please correct it, if I am wrong. a. Analysis? (finding trends, saisonnalities, outliers, missing values...) i. Can I do it with Nixtla stack? Or I should use also others libraries? b. Choose metrics i. Can I choose the MAPE for one univariate TS? c. Crossvalidation: i. Only train and test? or train/val/test? ii. n_window : is there a good number of windows? If I use 5 or 30, the best model won't be the same. iii. The mean score I get from a crossval, should it be the score to communicate to management to explain the capability of a model? d. Choose best model e. Forecast the desired horizon 2. I would like to use the Nixtla stack. For my time serie where should I start? a. StatsForecast, then MLForecast and finally neural-forecast? b. StatsForecast questions : i. Should I try it with all models and see the result with a crossval? Is it appropriate to tune these models? (I have not seen how to perform it in the documentation) c. MLForecast questions : i. My understanding here is we have to "transform" target (if needed) and create relevant features) 1. How can I know what transformation should I use for the target? 2. How can I know what are best features to create and how many? ii. I will try to find relevant external data, that can be used to help the model. 1. Should I do it here after a first iteration with only the target. Can it be done with StatsForecast and Neural-Forecast also? 2. Are there ways to evaluate if external data is relevant? d. NeuralForecast, I have not been through the documentation yet, so I will send questions later once I read it. e. Hierarchical forecast: This can be a last step, I will have to explore the data deeply before getting to this, and understand all the above. i. Is it correct, to try this type of forecasting and then compare it to the others above? More questions when I get there... Thank you very much for taking the time to read and answer me. I hope my questions are relevant and in the right place. Best regards Nass

Hi all, I want to use StatsForecast with AutoARIMA model to forecast a timeserie with missing data point in it? Is it possible for now? If not, what is the work around for this? Thank you very much! Best, Hieu

hello could any one explain this in example

level

(list of floats): this optional parameter is used for probabilistic forecasting. Set the level (or confidence percentile) of your prediction interval. For example, level=[90] means that the model expects the real value to be inside that interval 90% of the times

Hi @Max (Nixtla) could i know where is the paper that you talk about in your video .

When using AutoARIMA or other methods in StatsForecast we get confidence intervals (CIs) for each individual forecast. Does anyone know how we can combine these CIs to get one CI for the average of all forecasted series? I am sure we cannot average CIs because calculation of confidence interval is not a linear operation. But I have never seen how bunch of CIs can be combined to represent one CI for the average series. This is more of a math question than related to StatsForecast but I thought someone here might have thought of this before.

IF I am using cross validation for neuralforcast

Y_hat_df = nf.cross_validation(df=Y_df, val_size=val_size,
                               test_size=test_size, n_windows=None)

in the librery said we can choose between the test size,val size or n-widows so we are divide the data to two train and test right not to three train and test and validation

Hey hey, I have a dataset with 1M+ customers across several different products. I'd like to do a hierarchical forecast Cohort -> Product -> Customer using Nixtla, but I'm running into memory issues on my local machine, especially when putting the data into the right format. What would you recommend I do? I've read this tutorial, but that doesn't seem to apply to hierarchical forecasts or neural forecasts.

are we doing historical forcast in neuralforcast ?

Hi all, for mlforecast target transform, how does it work? For example if I want to use standardscaler on target variables and perform cross validation, does it just apply standard scaler to individual time series training data and transform the individual series forecasts back? My main concern is data leakage and whether applying standard scaler can account for the different scale for multiple time series.

How does the hierarchicalforecast package (+ statsforecast) handle multiple TS that don't have the same dates? I notice that

aggregate

adds leading, but do those zeros affect the statsforecast models? For instance if one TS started in January 2019 and other started in April 2019, the 2nd TS would have zeros added after using

aggregate

, but do those zeros affect model performance?

Hi all, got a quick question. Why does ml forecast cross validation take much longer when I fit a random forest regressor compared to xgboost regressor? Xgboost took about 30 seconds for my data but random forest took about 12 mins. Is there a way to speed it up?

Has anyone tried week numbers (sin & cos) as exogenous variables with ARIMA? Does it generally improve the forecast?

can i know about the cross validation in the neuralforcast what is the type you have used ?

Hello, I am new in mlforecast and i have a question about this library. Is it possible that add external features such as weather temperature or trend, seasonality of series that i will forecast ? Thank you for your answer.

Hi all, for AutoARIMA, is there a way to forecast for the same time period but for different xreg all at once? For example, I want to forecast for horizon h=1,but I want to do that for different xreg inputs. The idea was to get some sort of simulation for the different scenarios And for cross validation, does it support exogenous regressors?

if my data random and it has one reguler peak at each day any recommendation for how can i start with it ?

What do you think about this GARCH model

Has anyone tried AutoCES with exogenous features? The model accepts exogenous features but it makes no difference in the prediction. The predictions are exactly the same with or without them. I tried with holidays as exogenous feature. I did the same with AutoARIMA too and it made a difference in the output but it seems with AutoCES the model is not really using the exogenous feature.

Hello to the Nixtla community, I have written this post on how you can work with #AutoArima with #StatsForecast Maybe it can help some https://www.linkedin.com/pulse/autoarima-con-statsforecast-naren-castellon?utm_source=share&utm_medium=member_android&utm_campaign=share_via

Does the NeuralForecast library support GPU acceleration on non-NVIDIA GPUs, such as Intel integrated GPUs or AMD GPUs?

hello everyone, last week I did a presentation at our local (Milan, Italy) Python Meetup about time series forecasting and Nixtla. Nothing particularly sophisticated, just a general intro to Time Series Forecasting and some examples on what you can do with Nixtla (mostly covered statsforecast). Thought of sharing here the repo (you find there link to see a video recording and slides) in case other are interested. Thanks for the excellent work on Nixtla's libraries! https://github.com/pietroppeter/pymi-timeseries-forecasting-nixtla

Has anyone found a good way to create lagged versions of exogenous features within the MLForecast framework?

Hi, I wonder if it is possible to run mlforecast and neuralforecast on Databricks?

Can the Naive method be used in an in-sample way not just for

y_t = y_t-1

but for

y_t = t-t-n

? This is particularly important because we essentially have multiple tasks, forecasting say 1 week, 2 weeks, 3 weeks into the future and we assume that after the following week we will observe the next value. From what I understood from reading the docs in its entirety the

forecast

method in Naive would not achieve this as it only looks at your training data.

i didnt understand what the predict.in-sample mean can any one explain please