hi there 👋 ! first of all thank your very much for your awesome set of libraries! I just have a quick question: Is MLForecast fitting one model for every timeseries or is it one global model for all the timeseries? Sorry if i missed that part in the documentation! Enjoy your weekend!

Hey! I have another question: I am using MLForecast with a sklearn pipeline including some one hot encoding step. I want to do some feature importance plots and I'm having issues getting a dataframe with all the final feature columns (dynamic features from MLForecast + onehot encoded columns). Any idea / hint to get such a dataframe? Any idea is much appreciated! I would go ahead and hack something together with the before_predict_callback printing out the column names. Greetings from Germany and thank you very much in advance 🙂

Hi, is there a way of doing hyperparameter tuning (LGBM or any other model) + feature selection at the same time by launching one crossval? And at the end showing the best hyperparameters and feature importance? Maybe it's not good practice to do it (I don't know), if not what would be the best way and how to do it please? In the current state, is possible to get feature importance after a crossval? Thank you very much. I have checked these 2 tutorials but couldn't find how to do it : https://nixtla.github.io/mlforecast/docs/cross_validation.html https://nixtla.github.io/mlforecast/lgb_cv.html#example

👋 Hello

mlforecast

world 🌎 ! A question for the package maintainers - would you every be interested in adding support for fraction differencing, analogous to ARFIMA models? I'm going to implement something like this for an existing

mlforecast

project, and I'd be happy to create a ticket and/or PR if it's a direction you'd be interested in.

Ina ml forecast cross validation their is only expanding window technique, how to do rolling window?

Hi #mlforecast, I am new to mlforecast and trying to get feature importance for LGBM model in mlforecast but cant find any documentation. The code looks like this. Has anyone done like this before ? fcst = MLForecast( models=[ lgb.LGBMRegressor(**model_params)], #xgb.XGBRegressor(**model_params)], freq='M', # include 6 monthly lags lags=[1,2,3,4,5,6], # create a feature for year and month date_features=['year','month'], lag_transforms={ 1: [expanding_mean, # add 1 period lag for the mean (expanding over the window) (rolling_mean,3), # add 1 period lag for 3 month rolling mean (rolling_mean,6), # add 1 period lag for 6 month rolling mean ] }, target_transforms=[StandardScaler()], # transform to z-score normalized )

Hi Team, I just came across your project - awesome! Would have wished to have seen it before having implemented most of it my self... Two quick questions: 1. I already have my multivariate time series dataset at hand with a variety of self-engineered date, lag and moving average features (including some domain specific impuation). Could I use your pipeline implementation with a LightGBM Regressor on top of my own pre-build extended dataframe? 2. Can I understand your multivarite implementation as a global method? Meaning that your under-hood implementation pools the data from all series (identified by a static time series id) and fits one LightGBM model to all series. It uses that model then to predict each series individually? Thanks for your help! 🙂

Hi all, I was wondering if an equivalent to

StatsForecast.cross_validation_fitted_values

will be implemented in MLForecast? I want to get a measure of how much my model is overfitting when tuning. Cheers, Andrew

Hi everyone, first of all thanks for creating such a cool project. I have been doing forecast with ML by hand until I came across this repo. However, I also notice the StatsForecast project which seems to also cover many of my use cases. Any suggestions as to why choose to do my project with one or the other?

This message was deleted.

<!channel>: Dear community, show some love for MLForecast https://www.linkedin.com/feed/update/urn:li:activity:7087015150888620033?commentUrn=u[…]0560564264960%2Curn%3Ali%3Aactivity%3A7087015150888620033%29

Hello Is there any specific reason why we can't reproduce the same prediction only if we change lag list order ?

Hello, is there a way to pass in categorical features to LightGBM through

fcst.fit(...)

? I could be way off, but it seems like using

date_features

does label encoding but then these are treated as numeric features in the model itself (or maybe that is by design?). Also, trying to specify the cat indexes in the constructor for LGBMRegression provides the following:

UserWarning: categorical_feature keyword has been found in `params` and will be ignored.
Please use categorical_feature argument of the Dataset constructor to pass this parameter.

Thanks!

I'm having somme issues using MLForecast with sklearn models. I'm using monthly data over 3 years with lags of range(1, 7) which causes NA data.

MLForecast.fit

handles dropping the NA data without issue (with the dropna param), but

MLForecast.predict

runs an error with the underlying sklearn models. Am I doing something incorrectly?

Is there documentation for all the possible lag transforms and target transforms we can do?

Is there a plan to add an equivalent of

forecast_fitted_values

to MLForecast?

Hi MLforecast/Nixtla team, great great work with all the Nixtla python packages, thanks in advance for your patience with my question. • I found a really good tutorial for MLforecasts including target transforms that are nicely done in a plug and play manner. https://github.com/Nixtla/mlforecast/blob/main/nbs/docs/end_to_end_walkthrough.ipynb I found the following interface:

class BaseTargetTransform(abc.ABC):
    def set_column_names(self, id_col: str, time_col: str, target_col: str):
        self.id_col = id_col
        self.time_col = time_col
        self.target_col = target_col
        
    @abc.abstractmethod
    def fit_transform(self, df: 'pd.DataFrame') -> 'pd.DataFrame':
        raise NotImplementedError
        
    @abc.abstractmethod
    def inverse_transform(self, df: 'pd.DataFrame') -> 'pd.DataFrame':
        raise NotImplementedError

Q: How would I go about coding my own MSTL decomposition target transform ?

What to do when you have a dataset of frequency of 5 min and you want to use predict with xgb model, but you want predictions for whole year that is approximately 40k samples, so inside the predict function is 40k and that takes too long. How to optimize this?

Thanks for developing such a great project. I have a question around using LightGBM... i'm having issues with the distributed module (completely unrelated) so I'm thinking about just using a large single machine to do the training. My dataset is roughly 800M records and reading it into a pandas dataframe may be an issue (or maybe not, just speculating). does MLforecast accept inputs other than pandas dataframes? would I be able to use the native lgb.Dataset(...)?

I think the new

forecast_fitted_values

function in MLForecast is broken. It's producing the transformed predictions rather than the inverse transformed one. I've written a Github issue.

I am trying to replicate what I've done using the

MLForecast

function using it with cross_validation. No issues with running it normally with a pandas dataframe, but when I try to run it with a spark dataframe (much like I was able to do wtih statsforecast), I get the following error: "RecursionError: maximum recursion depth exceeded in comparison". Haven't been able to find anything super helpful via google/stackoverflow, ChatGPT, nor this slack channel. I've tried all of the algos in the code below and also limiting it to just one (several iterations). Here's the code: @njit def rolling_mean_12(x): return rolling_mean(x, window_size=12) @njit def rolling_mean_24(x): return rolling_mean(x, window_size=24) ****** ML_models = [ lgb.LGBMRegressor(n_jobs=4, random_state=0, verbosity=-1), # xgb.XGBRegressor(n_jobs=4, random_state=0), # MLPRegressor(random_state=0, max_iter=1000, early_stopping=True, n_iter_no_change=10, tol=1e-4), # RandomForestRegressor(random_state=0), # ExtraTreesRegressor(random_state=0), # HistGradientBoostingRegressor(random_state=0), # KNeighborsRegressor() ] mlf = MLForecast( models = ML_models, freq = 'M',# our series have integer timestamps, so we'll just add 1 in every timeste, lags=[1, 12], # lags=range(1,6, 1), lag_transforms={ 1: [expanding_mean], 12: [rolling_mean_12], # 24: [rolling_mean_24], }, # date_features=['year','month','quarter','days_in_month'], target_transforms=[Differences([1, 24])] ) #cross validation of statsforecast models using spark ML_crossvalidation_SDF = mlf.cross_validation( data=sdf, window_size=3, n_windows=5 ).toPandas() ****** Note that the below does work with regular pandas dataframe: ML_crossvalidation_df = mlf.cross_validation( data=df2, window_size=3, n_windows=5, ) Any help appreicated!

I asked a similiar question in the #general channel regarding ensembles about a week ago. Haven't received any responses. Hope it is ok to repost a similar question in this channel. The sklearn

votingregressor

works within #mlforecast--in the sense it produces a result not a warning or failed run. Here are my questions: • Is #mlforecast supposed to be able to use the

votingregressor

? • Even though I'm not specifying any weights, the

votingregressor

does not produce true averages. Is there something I'm missing? • How are the prediction intervals derived? I'm coming from the

R

fable

world where you have to generate sample paths to derive prediction intervals for ensembles. Just wondering if what it is producing for PIs is accurate. • The previous

nixtla

package had a built-in ensemble forecaster. I have not been able to find such a function/method in the current

statsforecast

or

mlforecast

pacakges. If it exists can someone point me to documentation? • If there is an ensembl forecaster, can you use it across SF and MLF? Thanks! #ensemble #combinationmodel

HI, I am not sure why RandomForestRegressor doesnt fit well my data, as it only predicts constant values.

Here is my code:

from sklearn.ensemble import RandomForestRegressor from data_preparation import Preparation import pandas as pd from mlforecast import MLForecast import matplotlib.pyplot as plt from sklearn.metrics import mean_absolute_percentage_error from missing_timestamps import remove_duplicates from sklearn.ensemble import RandomForestRegressor # marvin data = Preparation(r'/home/ieftimska/operato-meteo-1/data/MAS_processed/ELES-MAS-5001.csv.gz', "AMBIENT_TEMPERATURE") #data = Preparation(r'/home/iva/Desktop/operato-meteo-1/data/MAS_processed/ELES-MAS-5001.csv.gz', "AMBIENT_TEMPERATURE") train, test = data.split() train_processed = remove_duplicates(train) #train_processed_ = train_processed["AMBIENT_TEMPERATURE"].copy().squeeze() train_processed_ = train_processed.asfreq("5T").fillna(method="ffill") test_processed = remove_duplicates(test) test_processed_ = test_processed["AMBIENT_TEMPERATURE"].copy().squeeze() dataset_train = pd.DataFrame() dataset_train["ds"] = train_processed.index dataset_train["y"] = train_processed.values dataset_train["unique_id"] = "AMBIENT_TEMPERATURE" models = RandomForestRegressor(random_state=123, n_jobs=-1, max_depth=10, n_estimators=100) model = MLForecast(models=models, freq='5T', lags=[t for t in range(1, 865)], date_features=['dayofweek', 'month'], num_threads=6) prepared_df = model.preprocess(dataset_train, id_col='unique_id', time_col='ds', target_col='y', static_features=[]) X_train, y_train = prepared_df.drop(columns=['unique_id', 'ds', 'y']), prepared_df['y'] model.fit_models(X_train, y_train) predictions = model.predict(horizon=864)

Can someone explain to me, what exactly forecast_fitted_values do? If I put inside fit function max_horizon lets say 864 (that is 3 days) , does that mean that forecast_fitted_values will return "predictions" for the train data, actually the whole train data returned by the fitted Regressor(not the real train data) + max_horizon (the next 3 days). Is it this same with predict function, where you just put the horizon and then get predictions starting from the end of the train data + the horizon? Then why in documentation it is written that you should run first forecast_fitted_values and then predict to compute predictions? Runnin just forecast_fitted_values, with previously fit (with argument max_horizon) is actually the same with fitting the data without the max_horizon and then just use predict to predict the values?

sample weights have come up a few times on github and on here but are there any supported strategies for weighting fits, both at an individual data point level, and a series level, in

MLForecast

? Being able to weight individual date values will enable modeling strategies like this (although this might equally well be handled with an exogenous variable, albeit less confidently). Series level weights would be useful for intentionally biasing the loss toward performance on important series (e.g., revenue weighted global fits in an portfolio of products, with varying sales performances). Are there any plans for this the moment?

When dealing with multiple time-series, is the approach carried out by MLforecast equivalent to multiple independent time series, as described here? Is the one hot encode implicit or does/should it be handled explicitly by the user?

Another one from me - nearly every internal link on https://nixtla.github.io/mlforecast/ is giving me a 404 today. Is it just me, or is something up?

Hi. Thanks for

mlforecast

! When the package runs in default mode using the recursive strategy (predictions at prior horizons become the lag features for predictions at the next horizon), is it possible to see the lag features generated through this process? When i do...

model.fit(X...)
model.predict(h=6)

For QC purposes, I'd like to see how

.predict()

there generated the lag features recursively. Is that possible? Thank you!