.. _sec_forecasting_quickstart:

Forecasting Time Series - Quick Start
=====================================


Via a simple ``fit()`` call, AutoGluon can train and tune

-  simple forecasting models (e.g., ARIMA, ETS, Theta),
-  powerful deep learning models (e.g., DeepAR, Temporal Fusion
   Transformer),
-  tree-based models (e.g., XGBoost, CatBoost, LightGBM),
-  an ensemble that combines predictions of other models

to produce multi-step ahead *probabilistic* forecasts for univariate
time series data.

This tutorial demonstrates how to quickly start using AutoGluon to
generate hourly forecasts for the `M4 forecasting
competition <https://www.sciencedirect.com/science/article/pii/S0169207019301128>`__.
For a short summary of how to train models and make forecasts in a few
lines of code with ``autogluon.timeseries``, scroll to the `bottom of
this page <#summary>`__.

Loading time series data as a ``TimeSeriesDataFrame``
-----------------------------------------------------


First, we import some required modules

.. code:: python

    import pandas as pd
    import matplotlib.pyplot as plt
    
    from autogluon.timeseries import TimeSeriesDataFrame, TimeSeriesPredictor

To use ``autogluon.timeseries``, we will only need the following two
classes:

-  ``TimeSeriesDataFrame`` stores a dataset consisting of multiple time
   series.
-  ``TimeSeriesPredictor`` takes care of fitting, tuning and selecting
   the best forecasting models, as well as generating new forecasts.

We start by downloading the M4 Hourly dataset from the official website
(click on the arrow to show the preprocessing code).

.. raw:: html

.. raw:: html

   <details>

Loader for the M4 Hourly dataset

.. code:: python

    pd.set_option('display.max_rows', 6)  # Save space when printing
    
    M4_INFO_URL = "https://github.com/Mcompetitions/M4-methods/raw/master/Dataset/M4-info.csv"
    M4_HOURLY_URL = "https://github.com/Mcompetitions/M4-methods/raw/master/Dataset/Train/Hourly-train.csv"
    
    def download_m4_hourly_dataset(save_path):
        metadata = pd.read_csv(M4_INFO_URL)
        metadata = metadata[metadata["SP"] == "Hourly"].set_index("M4id")
        data = pd.read_csv(M4_HOURLY_URL, index_col="V1")
        results = []
        for item_id in metadata.index:
            time_series = data.loc[item_id].dropna().values
            start_time = pd.Timestamp(metadata.loc[item_id]["StartingDate"])
            timestamps = pd.date_range(start_time, freq="H", periods=len(time_series))
            results.append(pd.DataFrame({"M4id": [item_id] * len(time_series), "Date": timestamps, "Value": time_series}))
        result = pd.concat(results, ignore_index=True)
        result.to_csv(save_path, index=False)
    
    download_m4_hourly_dataset("m4_hourly.csv")

.. raw:: html

.. raw:: html

   </details>

The M4 dataset contains time series from various domains like finance,
demography and economics. Our goal is to forecast the future values of
each time series in the dataset given the past observations.

We load the dataset as a ``pandas.DataFrame``

.. code:: python

    df = pd.read_csv(
        "m4_hourly.csv",
        parse_dates=["Date"],  # make sure that pandas parses the dates
    )
    df




.. raw:: html

    <div>
    <style scoped>
        .dataframe tbody tr th:only-of-type {
            vertical-align: middle;
        }
    
        .dataframe tbody tr th {
            vertical-align: top;
        }
    
        .dataframe thead th {
            text-align: right;
        }
    </style>
    <table border="1" class="dataframe">
      <thead>
        <tr style="text-align: right;">
          <th></th>
          <th>M4id</th>
          <th>Date</th>
          <th>Value</th>
        </tr>
      </thead>
      <tbody>
        <tr>
          <th>0</th>
          <td>H1</td>
          <td>2015-01-07 12:00:00</td>
          <td>605.0</td>
        </tr>
        <tr>
          <th>1</th>
          <td>H1</td>
          <td>2015-01-07 13:00:00</td>
          <td>586.0</td>
        </tr>
        <tr>
          <th>2</th>
          <td>H1</td>
          <td>2015-01-07 14:00:00</td>
          <td>586.0</td>
        </tr>
        <tr>
          <th>...</th>
          <td>...</td>
          <td>...</td>
          <td>...</td>
        </tr>
        <tr>
          <th>353497</th>
          <td>H414</td>
          <td>2017-06-06 09:00:00</td>
          <td>35.0</td>
        </tr>
        <tr>
          <th>353498</th>
          <td>H414</td>
          <td>2017-06-06 10:00:00</td>
          <td>26.0</td>
        </tr>
        <tr>
          <th>353499</th>
          <td>H414</td>
          <td>2017-06-06 11:00:00</td>
          <td>17.0</td>
        </tr>
      </tbody>
    </table>
    <p>353500 rows × 3 columns</p>
    </div>



Each row of the data frame contains a single observation (timestep) of a
single time series represented by

-  unique ID of the time series (``"M4id"``)
-  timestamp of the observation (``"Date"``) as a ``pandas.Timestamp``
-  numeric value of the time series (``"Value"``)

The raw dataset should always follow this format with at least three
columns for unique ID, timestamp, and target value, but the names of
these columns can be arbitrary. It is important, however, that we
provide the names of the columns when constructing a
``TimeSeriesDataFrame`` that is used by AutoGluon. AutoGluon will raise
an exception if the data doesn’t match the expected format.

.. code:: python

    ts_dataframe = TimeSeriesDataFrame.from_data_frame(
        df,
        id_column="M4id",  # column that contains unique ID of each time series
        timestamp_column="Date",  # column that contains timestamps of each observation
    )
    ts_dataframe




.. raw:: html

    <div>
    <style scoped>
        .dataframe tbody tr th:only-of-type {
            vertical-align: middle;
        }
    
        .dataframe tbody tr th {
            vertical-align: top;
        }
    
        .dataframe thead th {
            text-align: right;
        }
    </style>
    <table border="1" class="dataframe">
      <thead>
        <tr style="text-align: right;">
          <th></th>
          <th></th>
          <th>Value</th>
        </tr>
        <tr>
          <th>item_id</th>
          <th>timestamp</th>
          <th></th>
        </tr>
      </thead>
      <tbody>
        <tr>
          <th rowspan="3" valign="top">H1</th>
          <th>2015-01-07 12:00:00</th>
          <td>605.0</td>
        </tr>
        <tr>
          <th>2015-01-07 13:00:00</th>
          <td>586.0</td>
        </tr>
        <tr>
          <th>2015-01-07 14:00:00</th>
          <td>586.0</td>
        </tr>
        <tr>
          <th>...</th>
          <th>...</th>
          <td>...</td>
        </tr>
        <tr>
          <th rowspan="3" valign="top">H414</th>
          <th>2017-06-06 09:00:00</th>
          <td>35.0</td>
        </tr>
        <tr>
          <th>2017-06-06 10:00:00</th>
          <td>26.0</td>
        </tr>
        <tr>
          <th>2017-06-06 11:00:00</th>
          <td>17.0</td>
        </tr>
      </tbody>
    </table>
    <p>353500 rows × 1 columns</p>
    </div>



We refer to each individual time series stored in a
``TimeSeriesDataFrame`` as an *item*. For example, items might
correspond to different products in demand forecasting, or to different
stocks in financial datasets. This setting is also referred to as a
*panel* of time series. Note that this is *not* the same as multivariate
forecasting — AutoGluon generates forecasts for each time series
individually, without modeling interactions between different items
(time series).

``TimeSeriesDataFrame`` inherits from
`pandas.DataFrame <https://pandas.pydata.org/pandas-docs/stable/reference/api/pandas.DataFrame.html>`__,
so all attributes and methods of ``pandas.DataFrame`` are also available
in a ``TimeSeriesDataFrame``. Note how ``TimeSeriesDataFrame`` organizes
the data with a ``pandas.MultiIndex``: the first *level* of the index
corresponds to the item ID and the second level contains the timestamp
when each observation was made. For example, we can use the ``loc``
accessor to access each individual time series.

.. code:: python

    ts_dataframe.loc["H2"].head()




.. raw:: html

    <div>
    <style scoped>
        .dataframe tbody tr th:only-of-type {
            vertical-align: middle;
        }
    
        .dataframe tbody tr th {
            vertical-align: top;
        }
    
        .dataframe thead th {
            text-align: right;
        }
    </style>
    <table border="1" class="dataframe">
      <thead>
        <tr style="text-align: right;">
          <th></th>
          <th>Value</th>
        </tr>
        <tr>
          <th>timestamp</th>
          <th></th>
        </tr>
      </thead>
      <tbody>
        <tr>
          <th>2015-01-07 12:00:00</th>
          <td>3124.0</td>
        </tr>
        <tr>
          <th>2015-01-07 13:00:00</th>
          <td>2990.0</td>
        </tr>
        <tr>
          <th>2015-01-07 14:00:00</th>
          <td>2862.0</td>
        </tr>
        <tr>
          <th>2015-01-07 15:00:00</th>
          <td>2809.0</td>
        </tr>
        <tr>
          <th>2015-01-07 16:00:00</th>
          <td>2544.0</td>
        </tr>
      </tbody>
    </table>
    </div>



We can also plot some of the time series in the dataset

.. code:: python

    plt.figure(figsize=(20, 3))
    for item_id in ["H1", "H2"]:
        plt.plot(ts_dataframe.loc[item_id], label=item_id)
    plt.legend();



.. figure:: output_forecasting-quickstart_a33e23_13_0.png


Forecasting problem formulation
-------------------------------

Models in ``autogluon.timeseries`` provide *probabilistic* forecasts of
time series *multiple steps* into the future. We choose the number of
these steps—the *prediction length* (also known as the *forecast
horizon*) depending on our task. For example, our dataset contains time
series measured at hourly *frequency*, so we set
``prediction_length = 48`` to train models that forecast 48 hours into
the future. Moreover, forecasts are probabilistic: in addition to
predicting the *mean* (expected value) of the time series in the future,
models also provide the *quantiles* of the forecast distribution.

In order to report realistic results for how AutoGluon will perform on
unseen data; we will split our dataset into a training set, used to
train & tune models, and a test set used to evaluate the final
performance. In forecasting, this is usually done by hiding the last
``prediction_length`` steps of each time series during training, and
only using these last steps to evaluate the forecast quality (also known
as “out of time validation”). We perform this split using the
``slice_by_timestep`` method of ``TimeSeriesDataFrame``.

.. code:: python

    prediction_length = 48
    
    test_data = ts_dataframe  # the full data set
    
    # last prediction_length timesteps of each time series are excluded, akin to `x[:-48]`
    train_data = ts_dataframe.slice_by_timestep(None, -prediction_length)

Below, we plot the training and test parts of the time series for a
single country, and mark the test forecast horizon. We will compute the
test scores by measuring how well the forecast generated by a model
matches the actually observed values in the forecast horizon.




.. figure:: output_forecasting-quickstart_a33e23_17_0.png


Training time series models with ``TimeSeriesPredictor.fit``
------------------------------------------------------------

Below we instantiate a ``TimeSeriesPredictor`` object and instruct
AutoGluon to fit models that can forecast up to 48 timesteps into the
future (``prediction_length``) and save them in the folder
``./autogluon-m4-hourly``. We also specify that AutoGluon should rank
models according to mean absolute percentage error (MAPE), and that data
that we want to forecast is stored in the column ``"Value"`` of the
``TimeSeriesDataFrame``.

.. code:: python

    predictor = TimeSeriesPredictor(
        path="autogluon-m4-hourly",
        target="Value",
        prediction_length=prediction_length,
        eval_metric="MAPE",
    )
    
    predictor.fit(
        train_data,
        presets="medium_quality",
        time_limit=600,
    )


.. parsed-literal::
    :class: output

    ================ TimeSeriesPredictor ================
    TimeSeriesPredictor.fit() called
    Setting presets to: medium_quality
    Fitting with arguments:
    {'enable_ensemble': True,
     'evaluation_metric': 'MAPE',
     'hyperparameter_tune_kwargs': None,
     'hyperparameters': 'medium_quality',
     'prediction_length': 48,
     'random_seed': None,
     'target': 'Value',
     'time_limit': 600}
    Provided training data set with 333628 rows, 414 items (item = single time series). Average time series length is 805.9.
    Training artifacts will be saved to: /home/ci/autogluon/docs/_build/eval/tutorials/timeseries/autogluon-m4-hourly
    =====================================================
    AutoGluon will save models to autogluon-m4-hourly/
    AutoGluon will gauge predictive performance using evaluation metric: 'MAPE'
    	This metric's sign has been flipped to adhere to being 'higher is better'. The reported score can be multiplied by -1 to get the metric value.
    tuning_data is None. Will use the last prediction_length = 48 time steps of each time series as a hold-out validation set.
    
    Starting training. Start time is 2022-11-17 03:17:42
    No path specified. Models will be saved in: "AutogluonModels/ag-20221117_031742/"
    Models that will be trained: ['Naive', 'SeasonalNaive', 'ETS', 'Theta', 'ARIMA', 'AutoGluonTabular', 'DeepAR']
    Training timeseries model Naive. Training for up to 599.91s of the 599.91s of remaining time.
    	-0.3718       = Validation score (-MAPE)
    	0.00    s     = Training runtime
    	6.15    s     = Validation (prediction) runtime
    Training timeseries model SeasonalNaive. Training for up to 593.73s of the 593.73s of remaining time.
    	-0.1922       = Validation score (-MAPE)
    	0.00    s     = Training runtime
    	1.14    s     = Validation (prediction) runtime
    Training timeseries model ETS. Training for up to 592.57s of the 592.57s of remaining time.
    	-0.3554       = Validation score (-MAPE)
    	0.00    s     = Training runtime
    	108.32  s     = Validation (prediction) runtime
    Training timeseries model Theta. Training for up to 484.23s of the 484.23s of remaining time.
    	-0.2136       = Validation score (-MAPE)
    	0.00    s     = Training runtime
    	35.40   s     = Validation (prediction) runtime
    Training timeseries model ARIMA. Training for up to 448.80s of the 448.80s of remaining time.
    	-0.5144       = Validation score (-MAPE)
    	0.00    s     = Training runtime
    	43.48   s     = Validation (prediction) runtime
    Training timeseries model AutoGluonTabular. Training for up to 405.30s of the 405.30s of remaining time.
    	-0.0973       = Validation score (-MAPE)
    	70.49   s     = Training runtime
    	3.54    s     = Validation (prediction) runtime
    Training timeseries model DeepAR. Training for up to 331.26s of the 331.26s of remaining time.
    	-0.1321       = Validation score (-MAPE)
    	135.96  s     = Training runtime
    	3.64    s     = Validation (prediction) runtime
    Fitting simple weighted ensemble.
    	-0.0973       = Validation score (-MAPE)
    	161.38  s     = Training runtime
    	7.18    s     = Validation (prediction) runtime
    Training complete. Models trained: ['Naive', 'SeasonalNaive', 'ETS', 'Theta', 'ARIMA', 'AutoGluonTabular', 'DeepAR', 'WeightedEnsemble']
    Total runtime: 611.45 s
    Best model: WeightedEnsemble
    Best model score: -0.0973




.. parsed-literal::
    :class: output

    <autogluon.timeseries.predictor.TimeSeriesPredictor at 0x7efed78886a0>



Here we used the ``"medium_quality"`` presets and limited the training
time to 10 minutes (600 seconds). The presets define which models
AutoGluon will try to fit. For ``medium_quality`` presets, these are
simple baselines (``Naive``, ``SeasonalNaive``), statistical models
(``ARIMA``, ``ETS``, ``Theta``), tree-based models XGBoost, LightGBM and
CatBoost wrapped by ``AutoGluonTabular``, a deep learning model
``DeepAR``, and a weighted ensemble combining these. Other available
presets for ``TimeSeriesPredictor`` are ``"fast_training"``
``"high_quality"`` and ``"best_quality"``. Higher quality presets will
usually produce more accurate forecasts but take longer to train and may
produce less computationally efficient models.

Inside ``fit()``, AutoGluon will train as many models as possible within
the given time limit. Trained models are then ranked based on their
performance on an internal validation set. By default, this validation
set is constructed by holding out the last ``prediction_length``
timesteps of each time series in ``train_data``.

Evaluating the performance of different models
----------------------------------------------

We can view the test performance of each model AutoGluon has trained via
the ``leaderboard()`` method. We provide the test data set to the
leaderboard function to see how well our fitted models are doing on the
held out test data. The leaderboard also includes the validation scores
computed on the internal validation dataset.

In AutoGluon leaderboards, higher scores always correspond to better
predictive performance. Therefore our MAPE scores are multiplied by
``-1``, such that higher “negative MAPE”s correspond to more accurate
forecasts.

.. code:: python

    predictor.leaderboard(test_data, silent=True)


.. parsed-literal::
    :class: output

    Additional data provided, testing on additional data. Resulting leaderboard will be sorted according to test score (`score_test`).




.. raw:: html

    <div>
    <style scoped>
        .dataframe tbody tr th:only-of-type {
            vertical-align: middle;
        }
    
        .dataframe tbody tr th {
            vertical-align: top;
        }
    
        .dataframe thead th {
            text-align: right;
        }
    </style>
    <table border="1" class="dataframe">
      <thead>
        <tr style="text-align: right;">
          <th></th>
          <th>model</th>
          <th>score_test</th>
          <th>score_val</th>
          <th>pred_time_test</th>
          <th>pred_time_val</th>
          <th>fit_time_marginal</th>
          <th>fit_order</th>
        </tr>
      </thead>
      <tbody>
        <tr>
          <th>0</th>
          <td>WeightedEnsemble</td>
          <td>-0.134751</td>
          <td>-0.097290</td>
          <td>38.683255</td>
          <td>7.183047</td>
          <td>161.380724</td>
          <td>8</td>
        </tr>
        <tr>
          <th>1</th>
          <td>AutoGluonTabular</td>
          <td>-0.134804</td>
          <td>-0.097343</td>
          <td>3.553736</td>
          <td>3.540360</td>
          <td>70.487470</td>
          <td>6</td>
        </tr>
        <tr>
          <th>2</th>
          <td>DeepAR</td>
          <td>-0.165094</td>
          <td>-0.132091</td>
          <td>35.730400</td>
          <td>3.642687</td>
          <td>135.958061</td>
          <td>7</td>
        </tr>
        <tr>
          <th>...</th>
          <td>...</td>
          <td>...</td>
          <td>...</td>
          <td>...</td>
          <td>...</td>
          <td>...</td>
          <td>...</td>
        </tr>
        <tr>
          <th>5</th>
          <td>Naive</td>
          <td>-0.376335</td>
          <td>-0.371842</td>
          <td>6.070503</td>
          <td>6.153785</td>
          <td>0.003158</td>
          <td>1</td>
        </tr>
        <tr>
          <th>6</th>
          <td>ETS</td>
          <td>-0.514677</td>
          <td>-0.355394</td>
          <td>107.134129</td>
          <td>108.316061</td>
          <td>0.001543</td>
          <td>3</td>
        </tr>
        <tr>
          <th>7</th>
          <td>ARIMA</td>
          <td>-0.570275</td>
          <td>-0.514399</td>
          <td>44.550157</td>
          <td>43.481608</td>
          <td>0.001504</td>
          <td>5</td>
        </tr>
      </tbody>
    </table>
    <p>8 rows × 7 columns</p>
    </div>



Generating forecasts with ``TimeSeriesPredictor.predict``
---------------------------------------------------------

We can now use the fitted ``TimeSeriesPredictor`` to make predictions.
By default, AutoGluon will make forecasts using the model that had the
best validation score (as shown in the leaderboard). Let’s use the
predictor to generate forecasts starting from the end of the time series
in ``train_data``

.. code:: python

    predictions = predictor.predict(train_data)
    predictions.head()


.. parsed-literal::
    :class: output

    Model not specified in predict, will default to the model with the best validation score: WeightedEnsemble




.. raw:: html

    <div>
    <style scoped>
        .dataframe tbody tr th:only-of-type {
            vertical-align: middle;
        }
    
        .dataframe tbody tr th {
            vertical-align: top;
        }
    
        .dataframe thead th {
            text-align: right;
        }
    </style>
    <table border="1" class="dataframe">
      <thead>
        <tr style="text-align: right;">
          <th></th>
          <th></th>
          <th>mean</th>
          <th>0.1</th>
          <th>0.2</th>
          <th>0.3</th>
          <th>0.4</th>
          <th>0.5</th>
          <th>0.6</th>
          <th>0.7</th>
          <th>0.8</th>
          <th>0.9</th>
        </tr>
        <tr>
          <th>item_id</th>
          <th>timestamp</th>
          <th></th>
          <th></th>
          <th></th>
          <th></th>
          <th></th>
          <th></th>
          <th></th>
          <th></th>
          <th></th>
          <th></th>
        </tr>
      </thead>
      <tbody>
        <tr>
          <th rowspan="5" valign="top">H1</th>
          <th>2015-02-03 16:00:00</th>
          <td>661.648553</td>
          <td>527.424048</td>
          <td>573.553312</td>
          <td>606.749377</td>
          <td>635.119121</td>
          <td>661.653156</td>
          <td>688.196726</td>
          <td>716.543494</td>
          <td>749.736627</td>
          <td>795.803720</td>
        </tr>
        <tr>
          <th>2015-02-03 17:00:00</th>
          <td>564.042775</td>
          <td>429.804605</td>
          <td>475.938705</td>
          <td>509.165693</td>
          <td>537.551593</td>
          <td>564.098123</td>
          <td>590.574221</td>
          <td>618.866071</td>
          <td>652.086809</td>
          <td>698.224921</td>
        </tr>
        <tr>
          <th>2015-02-03 18:00:00</th>
          <td>522.598897</td>
          <td>388.328571</td>
          <td>434.461357</td>
          <td>467.700832</td>
          <td>496.044361</td>
          <td>522.624422</td>
          <td>549.157935</td>
          <td>577.513033</td>
          <td>610.770607</td>
          <td>656.798562</td>
        </tr>
        <tr>
          <th>2015-02-03 19:00:00</th>
          <td>480.429825</td>
          <td>346.242903</td>
          <td>392.311366</td>
          <td>425.501970</td>
          <td>453.917915</td>
          <td>480.434913</td>
          <td>506.927372</td>
          <td>535.295228</td>
          <td>568.505382</td>
          <td>614.689521</td>
        </tr>
        <tr>
          <th>2015-02-03 20:00:00</th>
          <td>459.061518</td>
          <td>324.892940</td>
          <td>370.896964</td>
          <td>404.155884</td>
          <td>432.562379</td>
          <td>459.073569</td>
          <td>485.619210</td>
          <td>513.987359</td>
          <td>547.199333</td>
          <td>593.239758</td>
        </tr>
      </tbody>
    </table>
    </div>



Predictions are also stored as a ``TimeSeriesDataFrame``. However, now
the columns contain the mean and quantile predictions of each model. The
quantile forecasts give us an idea about the range of possible outcomes.
For example, if the ``"0.1"`` quantile is equal to ``501.3``, it means
that the model predicts a 10% chance that the target value will be below
``501.3``.

We will now visualize the forecast and the actually observed values for
one of the time series in the dataset. We plot the mean forecast, as
well as the 10% and 90% quantiles to show the range of potential
outcomes.

.. code:: python

    plt.figure(figsize=(20, 3))
    
    item_id = "H3"
    y_past = train_data.loc[item_id]["Value"]
    y_pred = predictions.loc[item_id]
    y_true = test_data.loc[item_id]["Value"][-prediction_length:]
    
    # prepend the last value of true range to predicted range for plotting continuity
    y_pred.loc[y_past.index[-1]] = [y_past[-1]] * 10
    y_pred = y_pred.sort_index()
    
    plt.plot(y_past[-200:], label="Training data")
    plt.plot(y_pred["mean"], label="Mean forecast")
    plt.plot(y_true, label="Observed")
    
    plt.fill_between(
        y_pred.index, y_pred["0.1"], y_pred["0.9"], color="red", alpha=0.1, label=f"10%-90% confidence interval"
    )
    plt.title("Forecasted time series values vs. the real observations")
    plt.legend();



.. figure:: output_forecasting-quickstart_a33e23_25_0.png


Summary
-------

We used ``autogluon.timeseries`` to make probabilistic multi-step
forecasts on the M4 Hourly dataset. Here is a short summary of the main
steps for applying AutoGluon to make forecasts for the entire dataset
using a few lines of code.

.. code:: python

   import pandas as pd

   from autogluon.timeseries import TimeSeriesPredictor, TimeSeriesDataFrame

   # Load the data into a TimeSeriesDataFrame
   df = pd.read_csv(
       "m4_hourly.csv",
       parse_dates=["Date"],
   )
   ts_dataframe = TimeSeriesDataFrame.from_data_frame(
       df,
       id_column="M4id",  # name of the column with unique ID of each time series
       timestamp_column="Date",  # name of the column with timestamps of observations
   )

   # Create & fit the predictor
   predictor = TimeSeriesPredictor(
       path="autogluon-m4-hourly",  # models will be saved in this folder
       target="Value",  # name of the column with time series values
       prediction_length=48,  # number of steps into the future to predict
       eval_metric="MAPE",  # other options: "MASE", "sMAPE", "mean_wQuantileLoss", "MSE", "RMSE"
   ).fit(
       train_data=ts_dataframe,
       presets="medium_quality",  # other options: "fast_training", "high_quality", "best_quality"
       time_limit=600,  # training time in seconds
   )

   # Generate the forecasts
   predictions = predictor.predict(ts_dataframe)

Check out :ref:`sec_forecasting_indepth` to learn about the advanced
capabilities of AutoGluon for time series forecasting.