Hi, what are the rules regarding data privacy when using TimeGPT? Thank you.

Hello to evereone. I have a li ttle trouble with prediction intervals here. The attached file contains my dataframe. (with 924 unique ids) and some static variables. Follow my code:

train = df_completo_encoded.query("ds < '2023-07-07'")
test = df_completo_encoded.query("ds >= '2023-07-07'")


from mlforecast.utils import PredictionIntervals
from xgboost import XGBRegressor
from lightgbm import LGBMRegressor
from mlforecast.target_transforms import Differences
from mlforecast.utils import PredictionIntervals
from window_ops.ewm import ewm_mean
from window_ops.rolling import rolling_mean, seasonal_rolling_mean,rolling_min, rolling_max, rolling_std


mlf = MLForecast(
    freq = 'D',
    models=[ XGBRegressor(n_jobs = -1),LGBMRegressor(n_jobs = -1)],
     target_transforms=[Differences([1,7])],
     lag_transforms={
                        1: [(rolling_mean, 2),(rolling_mean, 3),(rolling_mean, 4),(rolling_mean, 5),(rolling_mean, 6),(rolling_mean, 7),
                            (rolling_mean, 7), (rolling_mean, 14), (rolling_mean, 28),(ewm_mean, 0.9), expanding_mean,
                            (rolling_min,7), (rolling_min,14),(rolling_min,28),
                            (rolling_max,7), (rolling_max,14),(rolling_max,28),
                            (rolling_std,2),(rolling_std,3),(rolling_std,4),(rolling_std,5),(rolling_std,6),(rolling_std,7), (rolling_std,14),(rolling_std,28),
      
                      },
    lags=[1,7,14,21,28],
   date_features=['month', 'year', 'day_of_week', 'day_of_year','is_month_start','quarter','days_in_month'],
      num_threads=4
                    )
%%time

mlf.fit( train,
    id_col='unique_id',
    #max_horizon = 47,
    prediction_intervals=PredictionIntervals(n_windows=10, window_size=47),
    time_col='ds',
    target_col='y',
    static_features= ['gtin','ADI','CV2','cluster_0','cluster_1','cluster_2','cluster_3','cluster_4'],)

To here, everything ok. But follow de error:

levels = [50, 80, 95]

forecasts = mlf.predict(47, level = levels )

---------------------------------------------------------------------------
NotImplementedError                       Traceback (most recent call last)
Cell In[81], line 3
      1 levels = [50, 80, 95]
----> 3 forecasts = mlf.predict(47, level = levels )
      4 forecasts.head()

File /usr/local/lib/python3.10/site-packages/mlforecast/utils.py:186, in old_kw_to_pos.<locals>.decorator.<locals>.inner(*args, **kwargs)
    184                 new_args.append(kwargs.pop(arg_names[i]))
    185             new_args.append(kwargs.pop(old_name))
--> 186 return f(*new_args, **kwargs)

File /usr/local/lib/python3.10/site-packages/mlforecast/forecast.py:532, in MLForecast.predict(self, h, dynamic_dfs, before_predict_callback, after_predict_callback, new_df, level, X_df, ids, horizon, new_data)
    528         model_names = self.models.keys()
    529         conformal_method = _get_conformal_method(
    530             self.prediction_intervals.method
    531         )
--> 532         forecasts = conformal_method(
    533             forecasts,
    534             self._cs_df,
    535             model_names=list(model_names),
    536             level=level_,
    537             cs_h=self.prediction_intervals.h,
    538             cs_n_windows=self.prediction_intervals.n_windows,
    539             n_series=self.ts.ga.ngroups,
    540             horizon=h,
    541         )
    542 return forecasts

File /usr/local/lib/python3.10/site-packages/mlforecast/forecast.py:55, in _add_conformal_distribution_intervals(fcst_df, cs_df, model_names, level, cs_n_windows, cs_h, n_series, horizon)
     53 scores = scores[:, :, :horizon]
     54 mean = fcst_df[model].values.reshape(1, n_series, -1)
---> 55 scores = np.vstack([mean - scores, mean + scores])
     56 quantiles = np.quantile(
     57     scores,
     58     cuts,
     59     axis=0,
     60 )
     61 quantiles = quantiles.reshape(len(cuts), -1)

File /usr/local/lib/python3.10/site-packages/pandas/core/arrays/masked.py:528, in BaseMaskedArray.__array_ufunc__(self, ufunc, method, *inputs, **kwargs)
    525         return NotImplemented
    527 # for binary ops, use our custom dunder methods
--> 528 result = ops.maybe_dispatch_ufunc_to_dunder_op(
    529     self, ufunc, method, *inputs, **kwargs
    530 )
    531 if result is not NotImplemented:
    532     return result

File /usr/local/lib/python3.10/site-packages/pandas/_libs/ops_dispatch.pyx:113, in pandas._libs.ops_dispatch.maybe_dispatch_ufunc_to_dunder_op()

File /usr/local/lib/python3.10/site-packages/pandas/core/ops/common.py:81, in _unpack_zerodim_and_defer.<locals>.new_method(self, other)
     77             return NotImplemented
     79 other = item_from_zerodim(other)
---> 81 return method(self, other)

File /usr/local/lib/python3.10/site-packages/pandas/core/arraylike.py:198, in OpsMixin.__rsub__(self, other)
    196 @unpack_zerodim_and_defer("__rsub__")
    197 def __rsub__(self, other):
--> 198     return self._arith_method(other, roperator.rsub)

File /usr/local/lib/python3.10/site-packages/pandas/core/arrays/masked.py:659, in BaseMaskedArray._arith_method(self, other, op)
    657         other = np.asarray(other)
    658     if other.ndim > 1:
--> 659         raise NotImplementedError("can only perform ops with 1-d structures")
    661 # We wrap the non-masked arithmetic logic used for numpy dtypes
    662 #  in Series/Index arithmetic ops.
    663 other = ops.maybe_prepare_scalar_for_op(other, (len(self),))

NotImplementedError: can only perform ops with 1-d structures

Is there a StatsForecast method or some way to return samples? I'm forecasting h periods ahead and getting point estimates and confidence intervals. Just wondering if there's a way to easily return some sample h-period sequences.

hi, I pip installed the latest version 0.9.1 but somehow the

target_transformer

is not updated to latest. I’m running into this error when importing

GlobalSklearnTransformer

and have to manually update this file.

ImportError: cannot import name 'GlobalSklearnTransformer' from 'mlforecast.target_transforms' (/home/ec2-user/anaconda3/envs/plaid_python38/lib/python3.8/site-packages/mlforecast/target_transforms.py)

Hi everyone, I use statsforecast for my time series project that is based for demand forecasting. So, i have a question for you. I am trying to predict next 13 periods. In some series, model gave me good results when compare with test values. However, in some series, i encounter a problem. The problem is to predict same values in every 13 periods. Is there any reason for this? And can i fix this problem via using model parameters etc.? Thank you for interest.

Hi everyone, I have 2 questions: (Windows 10, ray 2.6.3, neuralforecast 1.6.2) 1.I try to do Hyperparameter Optimization on LSTM. Does anyone know how to resolve: "ray.tune.error.TuneError: Sync process failed: GetFileInfo() yielded path" error? I tried to add in BaseAuto->Tuner->Run_config storage_path, but the error still persist. 2.It's posible to add Exogenous Variables to AutoLSTM?

Hi everyone! I'm a data scientist working in retail modeling, and I've been learning Nixtla recently. It seems incredibly useful and well made! Definitely going to make my life easier! That said, I did have a few questions, specifically pertaining to the MLForecast library: 1. Can I save just the fitted model somehow? From what I gather the fitted model object has not only the model and transformation parameters, but also the entire data series that were used to train. This is obviously a problem if the data sets are gigabytes and I train multiple models on different sections. Is setting 'keep_last_n' to 0 and then feeding in entire series with 'new_df' in the predict step the right way to solve this? Relatedly, how do I access the fitted booster itself to run LightGBM.plot_importance(). 2. What happens if new_df is entered in the prediction that is discontinuous with the series at training time? Or if the training series itself has discontinuities? Are zeros filled in to the gaps? Or NAs? 3. Can I give a forecast date, e.g. forecast all series from a specific date? It seems like forecasts are built individually from the end of each series, even when different series have different last points. 4. Can I derive features based off columns other than the target? For example, if I have a Price column, I might also want a Price_lag1, Price_std, Price_cumrel etc. Do I need to calculate these manually? 5. Does any feature processing happen under the hood? e.g. removing a feature if all values are the same, normalize a feature etc.? 6. How are the identifier, target and the date column passed to LightGBM? Does target_col in MLForecast.fit overwrite label_col in the LightGBM_Parameters object? If the date and identifier columns aren't passed to LightGBM at all, does that effect column numbering (e.g. if I indicate in LightGBM_Parameters that the weight column is column 4). 7. Whats the difference between X_df and dynamic_dfs in MLForecast.predict?

8. I'm not an expert at Python but I noticed that in some of the examples feature properties are explicitly called out, for example here (https://nixtla.github.io/mlforecast/forecast.html#dynamic-features) the first line of code is: series = generate_daily_series(100, equal_ends=True, n_static_features=2, static_as_categorical=False) I'm assuming this labels the features as not categorical; does this have any effect on forecasting (I have been labeling categorical features through LightGBM_Parameters, but previously in R the dataframe itself carried that information; not sure how Python works).

Hi <!channel> I just joined and hit the wall when trying to import ConformalIntervals

from statsforecast.utils import ConformalIntervals

gives the error:

ImportError: cannot import name 'ConformalIntervals' from 'statsforecast.utils'

Has anyone else experienced this? Python = 3.10.9 & statsforecast = 1.5.0 Thanks a bunch.

I'm still trying to figure out how to add sample weights to MLForecast. According to https://lightgbm.readthedocs.io/en/latest/pythonapi/lightgbm.LGBMRegressor.html#, sample weights can be added in two places; either in the parameters given when LGBMRegressor is called, or in the LGBMRegressor.fit method. However, the parameter only works "in case of loading data directly from text file" (https://lightgbm.readthedocs.io/en/latest/Parameters.html), and the fit call from MLForecast doesn't allow a weight parameter (https://nixtla.github.io/mlforecast/forecast.html#mlforecast.fit). Anyway, my thought was that I could make a work-around by making a MLForecast object, doing the preprocessing, then extracting the model and training it with lightgbm calls, plugging it back in to MLForecast, and then using the predict feature of MLForecast (hence allowing me to take advantage of Nixtla's rolling feature calculations for future features). Code (see below) did not work as I was hoping (ValueError: No fitted models found. You have to call fit or preprocess + fit_models.). Is there a way to make this work?

import pandas as pd
import numpy as np
from mlforecast import MLForecast
import lightgbm as lgb

cols = 4        # col 0 is target, col 1 is weights, cols 2 and 3 are features
rows = 20000

data=np.random.rand(rows,cols)
columns = ['y','weight','feat_1','feat_2']

data_df = pd.DataFrame(data[:,0:4], columns=columns)
data_df['unique_id'] = 1
data_df['ds']= [i for i in range(0, rows)]

model_1 = MLForecast(
    models={
        'LGBM': lgb.LGBMRegressor()
    },
    freq=1,
    lags = [1, 2],
)

prep = model_1.preprocess(data_df, static_features = None, dropna = False)
train_weight = prep['weight']
train_x = prep.iloc[:,[2,3,6,7]]
train_y = prep['y']
model_1.models['LGBM'].fit(train_x,train_y,train_weight)

data_future = pd.DataFrame({'unique_id': [1,1,1],'ds':[rows,rows+1,rows+2], 'feat_1':[1,2,3], 'feat_2':[1,2,3]})
model_1.predict(3,X_df=data_future)

Hi all, I'm new to the project, just got token and experimenting with timegpt, i'm constantly hitting this error , does anyone also encounter this ? I'm just using the sample code . I'm able to get the forecast results. but statsforecast spit some error. I've checked both df, there is no null value.. Appreciate any help .. I'm on python@3.11/3.11.5 , numba 0.57.1. I suspect it is the version issue?

Traceback (most recent call last):
  File "/Users/lilypad/Documents/Code/timegpt/test1.py", line 19, in <module>
    sf.plot(df, fcst_df, level=[80, 90], max_insample_length=24 * 5)
  File "/opt/homebrew/lib/python3.11/site-packages/statsforecast/core.py", line 1789, in plot
    axes[idx, idy].fill_between(
  File "/opt/homebrew/lib/python3.11/site-packages/matplotlib/__init__.py", line 1446, in inner
    return func(ax, *map(sanitize_sequence, args), **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/opt/homebrew/lib/python3.11/site-packages/matplotlib/axes/_axes.py", line 5425, in fill_between
    return self._fill_between_x_or_y(
           ^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/opt/homebrew/lib/python3.11/site-packages/matplotlib/axes/_axes.py", line 5330, in _fill_between_x_or_y
    ind, dep1, dep2 = map(
    ^^^^^^^^^^^^^^^
  File "/opt/homebrew/lib/python3.11/site-packages/numpy/ma/core.py", line 2360, in masked_invalid
    return masked_where(~(np.isfinite(getdata(a))), a, copy=copy)
                          ^^^^^^^^^^^^^^^^^^^^^^^
TypeError: ufunc 'isfinite' not supported for the input types, and the inputs could not be safely coerced to any supported types according to the casting rule ''safe''

python script is :

import pandas as pd
import os
from nixtlats import TimeGPT
from statsforecast import StatsForecast as sf

print(os.environ)


df = pd.read_csv('<https://raw.githubusercontent.com/Nixtla/transfer-learning-time-series/main/datasets/electricity-short.csv>')

timegpt = TimeGPT(token=os.environ['TIMEGPT_TOKEN'])
fcst_df = timegpt.forecast(df, h=24, level=[80, 90])

print("NaN count in df:")
print(df.isna().sum())
print("NaN count in fcst_df:")
print(fcst_df.isna().sum())

sf.plot(df, fcst_df, level=[80, 90], max_insample_length=24 * 5)

Hello! Thanks for the awesome project! I have a question regarding cross_validation function. Is it possible to define an initial "training window" size and let the algorithm perform rolling window until end of timeseries is reached without explicitly defining "training_size" or "n_windows"? I have multiple timeseries of different length and it seems a bit weird to specify a fixed amount of n_windows because I want as many cross validations on each timeseries as possible. My current approach is:

# Find timeseries with least data
min_ts_count = big_df_for_stats.groupby(by="unique_id").count()["y"].min()
# This is the amount of minimum trained wanted for rolling! window
initial_training_size = 365
step_size = 4
h = 1

test_size_parameter = ((min_ts_count - initial_training_size)//step_size)
# This is the workaround to prevent the exception which happens when "(test_size-h)%step_size != 0"
test_size_parameter = test_size_parameter - (test_size_parameter%step_size) + h

res_df = sf.cross_validation(h=h, step_size=step_size, fitted=True, test_size=test_size_parameter, n_windows=None)

Another question about MLForecast: Is it possible to train different models on different subsets features, without having to make separate MLForecast objects? I'd like to be able to do the preprocessing only once, but then train multiple models on different sets of features. I can run the preprocessing and train the models easily enough, but when I go to predict them using the MLForecast.ts.predict() method I get an error like: ValueError: Number of features of the model must match the input. Model n_features_ is 31 and input n_features is 32 Is it possible to manipulate the data in the MLForecast object to only select the features used in training?

Also, I think I found a bug in the utilsforecast.preprocessing.fill_gaps() method. When using weekly frequency, data is lost unless specific weekly bucketing is used:

from utilsforecast import preprocessing as pr
import pandas as pd
import numpy as np

df = pd.DataFrame(
    {
        'unique_id': [0, 0, 0, 1, 1],
        'ds': pd.to_datetime(['2018-11-11', '2018-11-18', '2018-12-02', '2018-11-11', '2018-12-02']),
        'y': np.arange(5),
    }
)
df1 = pr.fill_gaps(
    df,
    freq='W',
)
print(df)
print(df1)

##### OUTPUT ######

   unique_id         ds  y
0          0 2018-11-11  0
1          0 2018-11-18  1
2          0 2018-12-02  2
3          1 2018-11-11  3
4          1 2018-12-02  4
   unique_id         ds   y
0          0 2018-11-08 NaN
1          0 2018-11-15 NaN
2          0 2018-11-22 NaN
3          0 2018-11-29 NaN
4          1 2018-11-08 NaN
5          1 2018-11-15 NaN
6          1 2018-11-22 NaN
7          1 2018-11-29 NaN

This is happening because the numpy definition for week is bucketed starting on Thursdays:

np.datetime64(pd.to_datetime("2018-11-11"),'W')

##### OUTPUT ######

numpy.datetime64('2018-11-08')

If I comment out the last two lines of _determine_bound in the fill_gaps method, the problem is resolved:

def fill_gaps(
    df: pd.DataFrame,
    freq: str = 'W',
    start: str = "per_serie",
    end: str = "global",
    id_col: str = "unique_id",
    time_col: str = "ds",
) -> pd.DataFrame:
    delta = np.timedelta64(1, freq) if isinstance(freq, str) else freq
    times_by_id = df.groupby(id_col)[time_col].agg(["min", "max"])
    starts = _determine_bound(start, freq, times_by_id, "min")
    ends = _determine_bound(end, freq, times_by_id, "max") + delta
    sizes = ((ends - starts) / delta).astype(np.int64)
    times = np.concatenate(
        [np.arange(start, end, delta) for start, end in zip(starts, ends)]
    )
    if isinstance(freq, str):
        times = times.astype("datetime64[ns]", copy=False)
    uids = np.repeat(times_by_id.index, sizes)
    idx = pd.MultiIndex.from_arrays([uids, times], names=[id_col, time_col])
    return df.set_index([id_col, time_col]).reindex(idx).reset_index()
    
def _determine_bound(bound, freq, times_by_id, agg) -> np.ndarray:
    if bound == "per_serie":
        out = times_by_id[agg].values
    else:
        # the following return a scalar
        if bound == "global":
            val = getattr(times_by_id[agg].values, agg)()
            if isinstance(freq, str):
                val = np.datetime64(val)
        else:
            if isinstance(freq, str):
                # this raises a nice error message if it isn't a valid datetime
                val = np.datetime64(bound)
            else:
                val = bound
        out = np.full(times_by_id.shape[0], val)
    #if isinstance(freq, str):
        #out = out.astype(f"datetime64[{freq}]")
    return out

df = pd.DataFrame(
    {
        'unique_id': [0, 0, 0, 1, 1],
        'ds': pd.to_datetime(['2018-11-11', '2018-11-18', '2018-12-02', '2018-11-11', '2018-12-02']),
        'y': np.arange(5),
    }
)
df1 = fill_gaps(
    df,
    freq='W',
)
print(df)
print(df1)

##### OUTPUT #####
   unique_id         ds  y
0          0 2018-11-11  0
1          0 2018-11-18  1
2          0 2018-12-02  2
3          1 2018-11-11  3
4          1 2018-12-02  4
   unique_id         ds    y
0          0 2018-11-11  0.0
1          0 2018-11-18  1.0
2          0 2018-11-25  NaN
3          0 2018-12-02  2.0
4          1 2018-11-11  3.0
5          1 2018-11-18  NaN
6          1 2018-11-25  NaN
7          1 2018-12-02  4.0

However, with this change problems occur when the weekly bucketing of the data is not precise (i.e. if I accidentally put '2018-12-01' in it would error)

Regarding my initial question (https://nixtlacommunity.slack.com/archives/C031EJJMH46/p1693989589210909) I would like to ask if my approach is the recommended way to do an extensive cross_validation in the special case that the forecasting horizon is only 1? I am asking because all examples in the documentation show the usage of 3<n_windows<20 which essentially means only "20 crossvalidated days " in the case of h=1. Also, I would like to ask if the Auto* Models perform hyperparameter tuning and model selection for each window inside the cross validation? Or, do they only perform one initial model selection on the whole time series, find the best model, and use that best model for each window inside the cross validation? Thank you in advance!

Using hierarchicalforecast, with MiddleOut (forecast proportions) reconciliation, it happened that for a certain point the second-to-last level has a forecasted value of 4 and the last (lowest) level, for the same timestep, has a forecasted value of over 7000 (I expect it to be equal or lower than 4). Could it be related to some division by a value close to zero that is happening somewhere?

Welcome!

BWBarber

joined #general.

Hi Everyone, I went through some of the previous messages and everyone seems doing well with the platform, so excuse me if my question seems quite rookie. I just found out about Nixtla a week ago and was trying to go through after getting a token from the platform, however, my requests were denied. I have already applied on the platform to gain access (https://www.nixtla.io/.) but didn't receive any feedback. can anyone please advise if I should do anything else ? or when to expect a reply

Hello everyone. First I would like to thank everyone on the team for implementing the fill_gaps function, my team and I are already using it a lot here! We are testing scale modeling for approximately 40k time series with mlforecast, the results are looking promising. But we would like to test with TimeGPT as well. However, when passing the tocken, it started to run, but soon after the following message appeared:

ApiError: status_code: 429, body: {'data': None, 'message': 'Too many requests', 'details': 'You have reached your request limit, email <mailto:ops@nixtla.io|ops@nixtla.io> to continue using the API', 'code': 'A20', 'requestID': 'CJA2TUG4YJ', 'support': 'If you have questions or need support, please email <mailto:ops@nixtla.io|ops@nixtla.io>'}

Hello, I have been working in R with Hyndman et al.'s fable family of packages. I'm trying to replicate my work using StatsForecast and MLforecast to see if I can improve performance (both accuracy and speed). Right now it appears what I've produced using Fable family of packages and Nixtla packages are getting somewhat similar results in terms of MASE on data trained using CV. However, in some cases Fable does better and others Nixtla. I believe there are two reasons: 1) I'm able to use Prophet with Fable (I know Nixtla's research suggestes ensembles using StatsForcast can produce on par or better, but I'm looking at ensembles and prophet is still getting better results); and 2) in Fable I'm able to use the "common exogenous" (https://fable.tidyverts.org/reference/common_xregs.html) variables trend and season. I've searchd for answers to these questions in previous posts, but didn't find answers that directly addressed them. My questions are: • Is there a way to replicate the common exogenous regressors using Nixtla packages? I'm thinking I might need to decompose the series and then merge that data into the dataframe. If that's right, how do I deal with needing those external regressors in the future frame for forecasts? Or is there something else I might have missed? • Is there a way to incorporate prophet into MLforecast? I think only sklearn, xgboost, and lightgbm work, but wanted to confirm. I have been trying to also forcast using the prophet API and them merging the data, but it'd be better to have it all in one set from the get-go. Appreciate any guidance.

Hi! Timegpt is a amazing work. Can you share what is the self-supervised pre-training task for TimeGPT? Contrastive learning or mask reconstruction

I'm interested in ensemble/combination modeling. The R Fable takes care of the math in the background for both the point forecast and the prediction intervals (both are important for my current problem). The link below is all I've been able to find. It references the

nixtla

package, which I can't find. I haven't had luck with google/stackoverflow searches. I did find one ticket/issue on the GitHub page about FFORMA (which would be great), but right now I'm just looking to do combo models of simple averages--with statsforecast and mlforecast models. Is this a feature I've just not been able to find or something perhaps in the works? https://medium.com/@gosshhh9/nixtla-using-ensembling-method-for-time-series-e1d23964ac30

Hi everyone! Just waiting for the API KEY activation and start my journey with TimeGPT with OpenBB!

morning folks, I'd like to know your tooling stack if you have built end to end ml ops pipeline for the nixtla libraries - esp ML Forecast. Our current approach is 1. to use cloud pickle to export the model binary 2. track experiments using our own custom built scripts. We are specifically looking into either mlflow or sagemaker.

If feeding the same data to TimeGPT, should I expect the same (stable) predictions each time or no? Does it have a temperature parameter?

Do the Auto* models from neuralforecast also include the exogenous covariates into the hyperparameter search? I am asking because there is no "futr_exog_list" parameter for the Auto* models.

Hey! Y'all think that any of the libraries present can be used for anomaly detection in audio files? If you've tried it

Welcome!

GoostValley

joined #general.

Hi everyone! Thanks to Max and all of you for admitting me as a TimeGPT Beta Test user! I will be cool If you can help me with some credits to start, my registered email is hernandogaitandiaz@gmail.com

Hello everybody !