Have you thought about adding an ensemble feature? I see improved performance by averaging a few of the models, but it'd be great to see an automated version.

Is it possible to extract forecast simulation draws from model fits, rather than just intervals and central estimates?

I'm sorry as I'm a begginer. But can I find a completed example to train LSTM and GRU models on multivariate time series ?

My case study about crude oil prices forecasting with multivariate dataset

When using

AutoARIMA

, I get a lot of warnings like

UserWarning: Having 3 or more differencing operations is not recommended. Please consider reducing the total number of differences.

and

RuntimeWarning: divide by zero encountered in long_scalars fit["aicc"] = fit["aic"] + 2 * npar * (npar + 1) / (nstar - npar - 1)

. Shouldn't these warnings be suppressed by using the automated version of the ARIMA?

Hi guys, when I try to use AutoArima, I have this error ZeroDivisionError: division by zero, can you help me with this problem Follow my code and the data attached.

Hi guys! Can anyone help me with the question above please? I'm afraid that this problem turn my tool impratctible.👆👆👆👆👆

Hey everyone! When using statsforecast models on my Mac M1, I reveive the error message

Assertion failed: (isInt<33>(Addend) && "Invalid page reloc value."), function encodeAddend, (...)

This seems to be a known (and open) issue of numba/LLVM on M1: https://github.com/numba/numba/issues/8567 Do you have any experiences with this problem? And maybe have an idea/roadmap for a possible fix? Any help & hints are highly appreciated 🙂 Thanks!

Hi, I have tried the intermittent models (Croston...) and it seems they forecast straight lines. I'va also seen it in your tutorial. Is it the normal behaviour? Thank you

Hey all, I've been working to switch some code over to statsforecast for baseline forecasting. I've hit a wall with integrating past covariates (date/uniuqe_id dependent features that are only known up to the prediction time). Future covariates seem to work effortless - you simply cut them at prediction time and merge the past to training data and feed the future part as df_X. Are past covariates simply not supported or am I just not seeing something obvious🙂? Thanks for your help! 🙌

Hi, I am trying to use crossval to evaluate multiple models on multiple time series. I have intermittent business daily data taking into account holidays (0 sales on holidays days). When I run the crossval, it forecasts values on holidays days... How can I make sure to avoid forecasting these days? Or at least get a 0 on these days?

can someone help me why it's doing it like this

Hi all, the Original MSTL paper suggest the following set up for trend estimation

Next, the trend component of the time series is computed using the last iteration of STL. On the other hand, if the time series is non-seasonal, MSTL uses the Friedman's Super Smoother function, supsmu, available in R (R Core Team, 2020), to directly estimate the trend.

In Nixtla it seems AutoArima is used and I wonder if this is the preferred set up, or perhaps can I set up the original MSTL as presented in the paper somehow? Without the AutoARIMA?

One more question : ), I am trying to follow the MSTL tutorial: https://nixtla.github.io/statsforecast/docs/tutorials/multipleseasonalities.html This section throws an exception for me and I am unable to get around it:

models = [MSTL(
    season_length=[24, 24 * 7], # seasonalities of the time series 
    trend_forecaster=AutoARIMA() # model used to forecast trend
)]

forecasts = sf.predict(h=24, level=[90])
forecasts.head()

I get the following exception when I include the level parameter in the predict method:

Exception: You have to instantiate either the trend forecaster class or MSTL class with `prediction_intervals` to calculate them

(Without the level=[90] parameter everything seem to work fine) I have also tried adding the conformal intervals:

MSTL(
        season_length=[24, 24*7],
        trend_forecaster=AutoARIMA(),
        prediction_intervals=ConformalIntervals(
            h = 3, # fcst horizon
            n_windows = 10 # number of CV windows, (conformal scores)
            )
    )

This also causes the predict method to crash with the following error:

ValueError: operands could not be broadcast together with shapes (1,24) (10,3)

I have statsforecasts 1.5.0 btw. Thanks in advance for patience.

Can statsforecast fit on a / "for a" specific prediction window ? https://nixtla.github.io/statsforecast/src/core/core.html#statsforecast.fit

sf.fit(df = Y_df)

For instance, currently the MSTL with AutoArima is not picking up any exogenous regressors. If the prediction window is only the next step clearly the AR MA lags dominate but my objective is e.g. 48 hours ahead and there the covariates might prove beneficial. How can I do this using statsforecast package ? Also is there some summary of fitted model similar to statsmodels? : ) Thanks so much

Hello everyone! I'm trying to forecast multiple time series of a combination of product-store sales. I would appreciate your help and knowledge in the following: 1. I have a df with 4,5 million rows, Running an AutoARIMA takes too long. Can it be my pc specs? Do you have any guidance with this or more volume of data, on how long it should take? (There are like 15k independent time series) 2. Additionally, I've attempted to run the same script on an M1 Max, but I encountered a compilation issue. I did some research and came across a related GitHub issue. Have any of you experienced something similar? 3. With some datasets I'm getting ZeroDivisionError. Could it be my data isn't "clean" enough? 4. And last one, should I homologate all my time series to the same min/max data range of my dataframe? Meaning that should I fill with zeroes and gaps of dates for all unique_ids of my df (using de global min/max of my df)? Thank you in advance for your time. This would be super helpful for me.

Dear all, I've been following this tutorial on anomaly detection https://nixtla.github.io/statsforecast/docs/tutorials/anomalydetection.html and everything is good, looks fine until the very last plot:

StatsForecast.plot(insample_forecasts, plot_random = False, plot_anomalies = True)

where only y is shown. No signs of the MSTL forecast or of the red circular markers pointing to where the anomalies were found. I recently bumped from statsforecast = 1.5.0 to 1.6.0. Perhaps there's a bug in the plot function? Thanks, Diego.

Can anyone tell me how to or direct me to documentation on how to do cross-validation while using spark for distributed processing? The page linked below says it's possible, but I can't find any documentation or tutorials on how to. Also couldn't find any previous posts with this question. https://nixtla.github.io/statsforecast/docs/getting-started/getting_started_complete.html#train-multiple-models-for-many-series

When I run the forecast command withh a single level I have no issues. However, when I run it with multiple levels, which for this particular task I need, I get the warning below. This happens whether I use the standard model prediction intervals or conformal PIs. Code;

forecasts_df = sf.forecast(h=3, level=[80,90,95,99], fitted = True)

Warning: `PerformanceWarning: DataFrame is highly fragmented. This is usually the result of calling frame.insert many times, which has poor performance. Consider joining all columns at once using pd.concat(axis=1) instead. To get a de-fragmented frame, use `newframe = frame.copy()`` I've tested the same code limiting to just one level and repeated with the other leves as separate runs. I could then merge them. But, despite the warning, the code does seem to work correctly--I get the same output when I run the single levels on their own and compare. Is this a known issue? Is it, as I am thinking, a warning I can ignore?

hello, I am new to forecasting... I have a simple question about the library. can anyone help me ?

Hi Nixtla team, first off let me say this is the best timeseries library Ive ever used. I have a question that has probably been asked already, so thanks for patience. In this tutorial: https://nixtla.github.io/statsforecast/docs/tutorials/statisticalneuralmethods.html This section:

from time import time

# Get the current time before forecasting starts, this will be used to measure the execution time
init = time()

# Call the forecast method of the StatsForecast instance to predict the next 28 days (h=28) 
# Level is set to [90], which means that it will compute the 90% prediction interval
fcst_df = sf.forecast(df=Y_df, h=28, level=[90])

# Get the current time after the forecasting ends
end = time()

# Calculate and print the total time taken for the forecasting in minutes
print(f'Forecast Minutes: {(end - init) / 60}')

throws:

RemoteTraceback 
Exception: You must pass `prediction_intervals` to compute them.
"""

Q: How should I modify the models for this to work? Should I add the Conformal Intervals somehow. (Many statistical models have conf intervals out of box?) Thank you and have a great day.

Hi again : ), I wonder if statistical models from statsforecast are suited for parallelization ? • I set up the StatsForecast object as follows:

models_stats = [
    AutoCES(season_length=96), 
    AutoETS(season_length=96)
    ]

# Instantiate StatsForecast class with the models
sf = StatsForecast(
    models=models_stats,
    freq='15T',
    n_jobs = 7 # NOTE: n_jobs instead of num_threads
)

sf.fit(df_train)

• yet I do not see more CPUs being utilized and the fitting takes very long time. • num_threads in mlforecast works amazingly well, so I wonder maybe this is perhaps due to the nature of statsforecast algos, that they cant be parallelized as well as e.g. the lgbm ? • (I do run the analysis in jupyter notebook but I doubt that is the culprit) Thanks and have a great weekend.

I noticed that one small difference between Statsforecast and MLforecast is that in cross validation, the MLforecast allows for refit param to take integer, whereas in statsforecast it is simply true / false, which means I either refit after every single step or I do not refit at all during the CV. • Is this by design ? • Is it possible to set up a daily refit in statsforecast same as in MLforecast, where I can simply set: refit = 24 ?

cv_df = sf.cross_validation(
    df = df,
    h = h,
    level = [90],
    step_size = step_size,
    test_size = test_size_adjusted,
    # input_size = 90*96,
    n_windows = None,
    refit = False,
    fitted = True
  )

hello, can someone explain to me cross validation ?

Hey folks, I am new here and recently started using the

statsforecast

package. I have a qq about

AutoARIMA()

. If I fit

AutoARIMA

to a dataframe with many unique_ids, is the best model across all unique_ids what gets returned? It seems like the answer is 'yes', but wanted to confirm. Thanks 🙂

Please could someone assist with 2 queries. 1. I am getting an error when directly forecast with statsforecast with the following code:

# define model(s)
ts_models = [
        AutoETS(model=['Z','Z','Z'], season_length=season, alias='AE')
    ]


# Instantiate StatsForecast class
model = StatsForecast(df = train,
                    models=ts_models,
                      freq = 'H',
                      n_jobs=-1)


# Forecast
forecast = model.forecast(h=1500, level=[95])

Error: _‘StatsForecast’ object has no attribute ‘fitted_’_ However if I use the fit and predict methods with the code below, the forecast is successful.

# plot data
StatsForecast.plot(train, engine='plotly')

# define model(s)
ts_models = [
        AutoETS(model=['Z','Z','Z'], season_length=season, alias='AE')
    ]


# Instantiate StatsForecast class
model = StatsForecast(
                    models=ts_models,
                      freq = 'H',
                      n_jobs=-1)

model.fit(train)

# Forecast
h = round((valid['ds'].nunique()) * 0.5)
forecast = model.predict(h=h, level=[95])

model.plot(df, forecast, engine='plotly')

Any ideas why the more efficient forecast method is not working in my case? 2. And a secondary issue, is that the StatsForecast.plot() is also not working when I run in a script in PyCharm. If anyone has guidance on this second query it would be appreciated.

Hi again, I am trying to save off the fitted model for re-use with the StatsForecast.save() and StatsForecast.load() methods but am hitting the error in the SS. Code:

from statsforecast import StatsForecast    

def fit():
    model = StatsForecast(df=panel_df,
                         models=[AutoARIMA(), SeasonalNaive(season_length=7)],
                         freq='D',
                         n_jobs=1,
                         verbose=True)
    model.fit(panel_df)
    model.save('/Users/tmb/PycharmProjects/data-science/UFE/output_files/model')

Error:

AttributeError: type object 'StatsForecast' has no attribute 'save'

I have the latest version of the package installed (e.g. 1.6.0)

I'm experiencing an odd thing with statsforecast that I haven't been able to figure out yet. I can use a horizon of 1 to 11 and things look great. When I try to go above H=11 the forecasts become uniform (all the same value) for all of the unique_ids in the dataframe and all of the algos I'm trying. It happens with and without exogenous variables. They look like the following. Any ideas of what might cause this?

Hello Nixtla! Thanks for all you do with this ecosystem I'm getting some transient warnings on

Naive()

fits on statsforecast 1.6.0 (and 1.5.0). Does this make sense to anyone? Details in 🧵

Hello Nixtla community! I have a few questions regarding distributed compute over Spark for statsforecasts models. Indeed, I have 30k series to forecast, and by staying on a single machine (even a big one), it takes too long to calculate, especially for "Auto" methods. So, I launched a large cluster of 320 cores on databricks, and the code below works very well (for example, I go from more than 2 hours to less than 15 minutes for an

AutoETS

run).

prediction_intervals = ConformalIntervals(h=forecast_horizon, n_windows=n_windows_conformal)

df = spark.read.parquet(f"{base_s3_dir_path}/{cutoff}/{input_dfs_dir_name}/df.parquet")
futr_df = spark.read.parquet(f"{base_s3_dir_path}/{cutoff}/{input_dfs_dir_name}/futr_df.parquet")

sf = StatsForecast(
    models=[eval(F"{algo}(season_length=season_length, alias=algo, prediction_intervals=prediction_intervals)")],
    freq=freq,
    fallback_model=SeasonalNaive(season_length=season_length),
    #n_jobs=1,
    #verbose=True
)

predictions = sf.forecast(
    h=forecast_horizon,
    df=df, 
    X_df=futr_df,
    level=[50, 80, 90, 95, 99],
    prediction_intervals=prediction_intervals,
).toPandas()

However, for

AutoARIMA

the computation time was still too high for me, so I dug a little deeper into the logs, and found that only 30% of CPUs were used on average during the run. • I've tried tweaking the input spark dfs repartition, and modifying a few spark confs, but nothing changes. Do you have any ideas? • Is the

n_jobs

parameter used when the detected backend is spark? • Would you advise me to constrain ARIMA's search fields a little to save a bit of computing time? If so, which parameters should be edited first? Thanks in advance!