A pull request that waited for review 12 days, 17 hours has been merged: statsforecast/[FEAT] Add M5 and M4-Daily experiments (Amazon Forecast)

A pull request was merged after being reviewed for 5 days, 19 hours: statsforecast/[FEAT] Add M5 and M4-Daily experiments (Amazon Forecast)

No se puede mostrar este contenido.

A pull request that waited for review 14 hours, 21 minutes has been merged: neuralforecast/[FIX] Errors with

fill_between

function (matplotlib)

A pull request that waited for review 1 day, 17 hours has been merged: neuralforecast/[FIX] Add correct link to hierarchicalforecast

A pull request that waited for review 1 day, 13 hours has been merged: neuralforecast/[DOCS] Add table-based documentation for core nb

A pull request that waited for review 4 days, 17 hours has been merged: mlforecast/[FIX] Improve description preprocessing ERCOT dataset

A pull request was merged after being reviewed for 3 days, 15 hours: mlforecast/[FIX] Improve description preprocessing ERCOT dataset

A pull request was merged after being reviewed for 5 days, 0 hours: mlforecast/[FEAT] Add step size argument to cross validation method

A pull request that waited for review 7 days, 4 hours has been merged: mlforecast/[FEAT] Add step size argument to cross validation method

No se puede mostrar este contenido.

A pull request that waited for review 2 days, 17 hours has been merged: neuralforecast/[FEAT] Optimization improvements and new features

A pull request that waited for review 4 days, 3 hours has been merged: mlforecast/[FEAT] Add

new_data

argument to

predict

method (allow transferability)

A pull request with 48 code changes was merged without review: statsforecast/[FEAT] Test recover M3 performance

A pull request that waited for review 2 days, 17 hours has been merged: neuralforecast/[FEAT] Added

valid_loss

possibilities

A pull request that waited for review 3 hours, 27 minutes has been merged: neuralforecast/[FEAT] Zero Inflated and Categorical Distributions (NBinomial, Tweedie, Bernoulli)

A pull request with 35 code changes was merged without review: mlforecast/automate release

A pull request that waited for review 2 hours, 24 minutes has been merged: neuralforecast/[FEAT] Added predict_rolled method

A pull request that waited for review 20 hours, 57 minutes has been merged: mlforecast/[DOCS] Add transfer learning tutorial

A pull request with 48 code changes was merged without review: statsforecast/Update README.md

A pull request with 35 code changes was merged without review: mlforecast/add cross validation fitted values

A pull request with 46 code changes was merged without review: mlforecast/allow id_col in static_features

A pull request that waited for review 1 day, 9 hours has been merged: mlforecast/allow id_col in static_features

Kraft Heinz is HIRING data scientists! We are looking to fill 3 roles by individuals with experience in time-series forecasting! Are you interested? Apply now, currently interviewing for these roles! Candidates must be US located. Headquartered in Chicago, IL but remote US also available. Principal Data Scientist: https://lnkd.in/gxrXbvt4 Senior Manager of Data Science: https://lnkd.in/gPfTWdKD Senior Data Scientist: https://lnkd.in/gCUNy2QG

Hi there @Max (Nixtla) @Cristian (Nixtla). I have a general question about how

mlforecast

,

statsforecast

and

neuralforecast

can interact (or not?). From what I understand,

mlforecast

allows to train and apply any models from sklearn as well as lightgbm/xgboost. Is it possible to also use models from

statsforecast

(e.g. AutoARIMA) or from

neuralforecast

(e.g. NBEATS) ?

something like

mlf = MLForecast(
    models = [LinearRegression(), lgb.LGBMRegressor(),AutoARIMA(),NBEATS()],
    lags=[1, 12],(
)
mlf.fit(df)
mlf.predict(12)

Or am I missing something here ?

Random, more theoretical question, has anyone ever tried training a sequence model to take in the time series as input but the output sequence is the best performing stat model rather than the actuals? Seems to me like the output sequence could contain signals that aren't in the input sequence (major jumps or changepoints) but if the output sequence is based on a stat output it would. Let me know if I'm crazy but either way a fast 'auto-stat emulator' would be neat.

Out of curiosity, have any of you tried some timeseries data augmentation techniques like the block bootstrap. I was thinking of Hyndman's Bagged ETS model. https://otexts.com/fpp2/bootstrap.html#bagged-ets-forecasts That got me thinking about having a bootstrap timeseries loader for every batch that goes into one of your models.

Hi, is it possible to save a trained MLForecast model object and later load it for prediction ?