Hi dear members of the community: we just released some features. Please share some love: https://www.linkedin.com/posts/mergenthaler_machinelearning-timeseries-forecasting-ac[…]837886816256-X4jI?utm_source=share&utm_medium=member_desktop

Special mention to @Valeriy, that has been a promoter of conformal prediction.

Christmas 🎅 comes early, amazing release team

Nixtla tweet about new features including CP is 20K views Ilon Musk has competition now 👍 🌟

Hi, is it possible for us to submit some PRs regarding ExponentialMovingAverage and RollingMovingAverage models for statsforecast ? Like these are quick and easy changes and would help out I guess. You could get these models by using Statsforecast objects and benefit from the fast compute, instead of relying on other libs.

Nixtla tweet on fire 30K 🔥🔥🔥🔥🔥

Hey, I am running through your stastsforecast documentation and struggling to understand if there is an option to save the models? Do I just pickle it? Example utilisation of the package:

df = process_df()

    models = [
            AutoARIMA(max_D=12, season_length=12),
    ]

    sf = StatsForecast(
            df=df,
            sort_df=True,
            models=models,
            freq='M',
            n_jobs=-1,
    )

Hi! Does anybody have any idea why AutoARIMA seems to suddenly forecast large quantities where it should not ? Like it`s weird, because a month ago, it seemed to be working fine and predicting at a high enough rate and we did not change anything.

Another additional question: how come, if you add season_length=[x], (x > 1) for AutoCES/ AutoARIMA, the time needed for it to run is 2x than when we keep it default (meaning season_length=1). It`s weird and I have seen this behaviour consistently.

Hey Guys! Hope everyone doing well? I came across problem in statsforecast,can anyone tell me how to check weights of models?

Hello! Any idea why AutoCES(complex exp smoothing) would give me a No model able to be fitted exception ? Like I tried different season lengths for my data and it`s weird that we get this issue. I have seen that this issue arises because the best_ic is still infinite. which shouldn`t happen . .. .

Hi everyone! I was using AutoARIMA of statsforecast in order to retrieve the order of the best model that fits the data but I don't know how to do it. Using pmdarima auto_arima the order is retrieved by calling auto_arima().order. Is there a similar feature in statsforecast? Thank you very much in advance!

Hey I am guessing the

environment.yml

was created using conda due to the inability to normally install prophet with pip?

Hi guys, quick question if you don't mind. Based on this notebook: https://nixtla.github.io/mlforecast/docs/prediction_intervals.html, I am unsure of how to literally predict the next step that is not in the test set, i.e., t+1. So, for instance, a stock price that will be next minute, that we don't know yet. Are there any examples to do that? Because the horizon in predict seems to predict the series as if we started 48 hours ago, not as if we started now. Could you please clarify?

Hi everyone, I was wondering if there are plans to implement constant in-sample time series cross validation (sliding window rather than expanding window) as below:

image.png

Image from: https://openforecast.org/adam/rollingOrigin.html

This would be useful in energy demand/price forecasting where capturing short-term dynamics is key.

Hi dear community: Today we are comparing Nixtla and Fugue with Google. Please show some love.

Hi I am new to this Framework. Thank you for the cool work! I couldn't find anything on multivariant scenarios, is it possible to use multiple features?

Hi everyone, I went through the paper and code on GitHub, amazing library. I am working on time series forecasting for different product sales. I currently have about 10 products with different names and sales data for last 48 months. Looking forward to any suggestions for this.

No regressors provided

getting error for forecast.predict

Hi Team, I am using hierarchicalforecast I am getting this error when I run HierarchicalReconciliation : hrec = HierarchicalReconciliation(reconcilers=reconcilers) reconciled_forecasts = hrec.reconcile(Y_hat_df=forecasts_df, S=S, tags=tags)

ValueError                                Traceback (most recent call last)
Cell In[185], line 3
      1 hrec = HierarchicalReconciliation(reconcilers=reconcilers)
----> 3 reconciled_forecasts = hrec.reconcile(Y_hat_df=forecasts_df, S=S, tags=tags2)

File ~\Anaconda3\envs\pycaret\lib\site-packages\hierarchicalforecast\core.py:265, in HierarchicalReconciliation.reconcile(self, Y_hat_df, S, tags, Y_df, level, intervals_method, num_samples, seed, sort_df)
    261     fcsts_model = reconciler.predict(S=reconciler_args['S'], 
    262                                      y_hat=reconciler_args['y_hat'], level=level)
    263 else:
    264     # Memory efficient reconciler's fit_predict
--> 265     fcsts_model = reconcile_fn(**kwargs, level=level)
    267 # Parse final outputs
    268 Y_tilde_df[recmodel_name] = fcsts_model['mean'].flatten()

File ~\Anaconda3\envs\pycaret\lib\site-packages\hierarchicalforecast\methods.py:468, in MiddleOut.fit_predict(self, S, y_hat, tags, y_insample, level, intervals_method)
    456 """Middle Out Reconciliation Method.
    457 
    458 **Parameters:**<br>
   (...)
    465 `y_tilde`: Reconciliated y_hat using the Middle Out approach.
    466 """
    467 if not is_strictly_hierarchical(S, tags):
--> 468     raise ValueError('Middle out reconciliation requires strictly hierarchical structures.')
    469 if self.middle_level not in tags.keys():
    470     raise ValueError('You have to provide a `middle_level` in `tags`.')

ValueError: Middle out reconciliation requires strictly hierarchical structures.

This is how my tags looks like :

{'Total': array(['Total'], dtype=object),
 'Total/level0': array(['DRAM_DRAM_M', 'NAND_SSD', 'DRAM_DRAM', 'NAND_FLASH'], dtype=object),
 'Total/level0/Level1': array(['DRAM_DRAM_M_DDR3', 'DRAM_DRAM_M_DDR4', 'NAND_SSD_PCI-e',
        'NAND_SSD_SATA', 'DRAM_DRAM_DDR4', 'DRAM_DRAM_LPDDR4X',
        'NAND_FLASH_COMPONENT', 'DRAM_DRAM_HBM', 'DRAM_DRAM_M_DDR5'],
       dtype=object),
 'Total/level0/Level1/Level2': array(['DRAM_DRAM_M_DDR3_DDR3', 'DRAM_DRAM_M_DDR4_DDR4',
        'NAND_SSD_PCI-e_PCI-e', 'NAND_SSD_SATA_SATA',
        'DRAM_DRAM_DDR4_DDR4', 'DRAM_DRAM_LPDDR4X_LPDDR4X',
        'NAND_FLASH_COMPONENT_COMPONENT', 'DRAM_DRAM_HBM_HBM',
        'DRAM_DRAM_M_DDR5_DDR5'], dtype=object)}

can anyone pls help ?

Hey! I was wondering if StatsForecast currently supports something like the "simulate" method in statsmodels? (https://www.statsmodels.org/dev/generated/statsmodels.tsa.statespace.sarimax.SARIMAX.simulate.html). I could not find anything like it in the docs, so I assume not. I am currently using this method to generate scenarios as input for an optimization problem, but would prefer to switch over to StatsForecast for the speed gains 🙂 .

Hi, everytime I run the following

mlf = MLForecast(models=models, differences=[1], date_features=['month'],
                 lags=None, num_threads=mp.cpu_count()-1,
                 lag_transforms=None, freq='W-FRI')

Where it just transforms the price series via

differences=[1]

, I always end up with

ValueError: Input contains NaN, infinity or a value too large for dtype('float64').

Well, I guess it's obvious that the moment you apply first-difference there will be nans, why MLForecast does not delete the first row? And to precede your question, no, my original dataset has no nans.

hi all, has anyone got a clear example of using AutoArima.forward ?( to apply existing model to other time series). I've been trying to following with no success: from statsforecast import * ARIMA_new = StatsForecast( models = [AutoARIMA(season_length=12)], freq = 'M', n_jobs = 1 ) ARIMA_new = ARIMA_new.fit(y=new_df)

Hi All. Does anyone know of any tutorials using the Nixtla framework on top of Dask or Ray from within Azure Machine Learning? I'm just getting started with both Nixtla and Azure Machine Learning, and I haven't come across anything linking the two yet.

Hello, when using the cross-validation function in statsforecast or mlforecast (https://github.com/Nixtla/statsforecast/blob/main/statsforecast/core.py#L242). Is the model re-fitted with every prediction? And if so, is there a way to control how often this happens to speed up training?