Forecasting time series with gradient boosting: Skforecast, XGBoost, LightGBM and CatBoost

If you like  Skforecast ,  help us giving a star on   GitHub! ⭐️

Forecasting time series with gradient boosting: Skforecast, XGBoost, LightGBM, Scikit-learn y CatBoost

Joaquín Amat Rodrigo, Javier Escobar Ortiz
February, 2021 (last update July 2023)


Gradient boosting models have gained popularity in the machine learning community due to their ability to achieve excellent results in a wide range of use cases, including both regression and classification. Although these models have traditionally been less common in forecasting, recent research has shown that they can be highly effective in this domain. Some of the key benefits of using gradient boosting models for forecasting include:

  • The ease with which exogenous variables can be included in the model, in addition to autoregressive variables.

  • The ability to capture non-linear relationships between variables.

  • High scalability, allowing models to handle large volumes of data.

  • Some implementations allow the inclusion of categorical variables without the need for one-hot coding.

Despite these benefits, the use of machine learning models for forecasting can present several challenges that can make analysts reluctant to use them, the main ones being

  • Transforming the data so that it can be used as a regression problem.

  • Depending on how many future predictions are needed (prediction horizon), an iterative process may be required where each new prediction is based on previous ones.

  • Model validation requires specific strategies such as backtesting, walk-forward validation or time series cross-validation. Traditional cross-validation cannot be used.

The skforecast library provides automated solutions to these challenges, making it easier to apply and validate machine learning models to forecasting problems. The library supports several advanced gradient boosting models, including XGBoost, LightGBM, Catboost and scikit-learn HistGradientBoostingRegressor. This document shows how to use them to build accurate forecasting models.

🖉 Note

Machine learning models do not always outperform statistical learning models such as AR, ARIMA or Exponential Smoothing. Which one works best depends largely on the characteristics of the use case to which it is applied.

⚠ Warning

The four gradient boosting frameworks – LightGBM, scikit-learn's HistogramGradientBoosting, XGBoost, and CatBoost – are capable of directly handling categorical features within the model. However, it is important to note that each framework has its own configurations, benefits and potential problems. To fully comprehend how to use these frameworks, it is highly recommended to refer to the skforecast user guide for a detailed understanding.

Case of use

Bicycle sharing is a popular shared transport service that provides bicycles to individuals for short-term use. These systems typically provide bike docks where riders can borrow a bike and return it to any dock belonging to the same system. The docks are equipped with special bike racks that secure the bike and only release it via computer control.

One of the major challenges faced by operators of these systems is the need to redistribute bikes to ensure that there are bikes available at all docks, as well as free spaces for returns.

In order to improve the planning and execution of bicycle distribution, it is proposed to create a model capable of forecasting the number of users over the next 36 hours. In this way, at 12:00 every day, the company in charge of managing the system will be able to know the expected demand for the rest of the day (12 hours) and the next day (24 hours).

For illustrative purposes, the current example only models a single station, but the predictive model can be adapted and extended to cover multiple stations, thereby improving the management of bike-sharing systems on a larger scale.


Libraries used in this document.

In [1]:
# Data processing
# ==============================================================================
import numpy as np
import pandas as pd

# Plots
# ==============================================================================
import matplotlib.pyplot as plt
from import plot_acf
from import plot_pacf
import plotly.graph_objects as go
import as px
import as pio
pio.templates.default = "seaborn"'seaborn-v0_8-darkgrid')

# Modelling and Forecasting
# ==============================================================================
from xgboost import XGBRegressor
from lightgbm import LGBMRegressor
from catboost import CatBoostRegressor
from sklearn.ensemble import HistGradientBoostingRegressor

from sklearn.preprocessing import OneHotEncoder
from sklearn.preprocessing import OrdinalEncoder
from sklearn.preprocessing import FunctionTransformer
from sklearn.pipeline import make_pipeline
from sklearn.compose import make_column_transformer
from sklearn.compose import make_column_selector

from skforecast.ForecasterAutoreg import ForecasterAutoreg
from skforecast.model_selection import grid_search_forecaster
from skforecast.model_selection import backtesting_forecaster

# Configuration warnings
# ==============================================================================
import warnings


The data in this document represent the hourly usage of the bike share system in the city of Washington, D.C. during the years 2011 and 2012. In addition to the number of users per hour, information about weather conditions and holidays is available. The original data was obtained from the UCI Machine Learning Repository and has been previously cleaned (code) applying the following modifications:

  • Renamed columns with more descriptive names.

  • Renamed categories of the weather variables. The category of heavy_rain has been combined with that of rain.

  • Denormalized temperature, humidity and wind variables.

  • date_time variable created and set as index.

  • Imputed missing values by forward filling.

The resulting data set contains the columns:

  • date_time: date and time.

  • month: month (1 to 12).

  • hour: hour (0 to 23).

  • holiday: whether or not the day is a holiday (taken from

  • weekday: day of the week (Monday = 0, Sunday = 6).

  • workingday : if it is a working day.

  • weather: the weather of the day (clear, mist, rain).

  • temp: recorded temperature.

  • atemp: thermal sensation.

  • hum: recorded humidity.

  • windspeed: recorded wind speed.

  • users: total number of users of the bike rental service.

In [2]:
# Downloading data
# ==============================================================================
url = (''
data = pd.read_csv(url)
data['date_time'] = pd.to_datetime(data['date_time'], format='%Y-%m-%d %H:%M:%S')
data = data.set_index('date_time')
data = data.asfreq('H')
data = data.sort_index()
data = data.drop(columns=['workingday'])
holiday weather temp atemp hum windspeed users month hour weekday
2011-01-01 00:00:00 0.0 clear 9.84 14.395 81.0 0.0 16.0 1 0 5
2011-01-01 01:00:00 0.0 clear 9.02 13.635 80.0 0.0 40.0 1 1 5
2011-01-01 02:00:00 0.0 clear 9.02 13.635 80.0 0.0 32.0 1 2 5
2011-01-01 03:00:00 0.0 clear 9.84 14.395 75.0 0.0 13.0 1 3 5
2011-01-01 04:00:00 0.0 clear 9.84 14.395 75.0 0.0 1.0 1 4 5

To facilitate the training of the models, the search for optimal hyperparameters and the evaluation of their predictive accuracy, the data are divided into three separate sets: training, validation and test.

In [3]:
# Split train-validation-test
# ==============================================================================
end_train = '2012-03-31 23:59:00'
end_validation = '2012-08-31 23:59:00'
data_train = data.loc[: end_train, :]
data_val   = data.loc[end_train:end_validation, :]
data_test  = data.loc[end_validation:, :]

print(f"Dates train      : {data_train.index.min()} --- {data_train.index.max()}  (n={len(data_train)})")
print(f"Dates validacion : {data_val.index.min()} --- {data_val.index.max()}  (n={len(data_val)})")
print(f"Dates test       : {data_test.index.min()} --- {data_test.index.max()}  (n={len(data_test)})")
Dates train      : 2011-01-01 00:00:00 --- 2012-03-31 23:00:00  (n=10944)
Dates validacion : 2012-04-01 00:00:00 --- 2012-08-31 23:00:00  (n=3672)
Dates test       : 2012-09-01 00:00:00 --- 2012-12-31 23:00:00  (n=2928)

Data exploration

Graphical exploration of time series can be an effective way of identifying trends, patterns, and seasonal variations. This, in turn, helps to guide the selection of potential lags that could serve as strong predictors in the model.

Plot time series

In [4]:
# Plot time series with zoom
# ==============================================================================
zoom = ('2011-08-01 00:00:00','2011-08-15 00:00:00')

fig = plt.figure(figsize=(8, 4))
grid = plt.GridSpec(nrows=8, ncols=1, hspace=0.1, wspace=0)

main_ax = fig.add_subplot(grid[1:3, :])
zoom_ax = fig.add_subplot(grid[5:, :])

data_train['users'].plot(ax=main_ax, label='train', alpha=0.5)
data_val['users'].plot(ax=main_ax, label='validation', alpha=0.5)
data_test['users'].plot(ax=main_ax, label='test', alpha=0.5)
min_y = min(data['users'])
max_y = max(data['users'])
main_ax.fill_between(zoom, min_y, max_y, facecolor='blue', alpha=0.5, zorder=0)
main_ax.legend(loc='lower center', ncol=3, bbox_to_anchor=(0.5, -0.8))
data.loc[zoom[0]: zoom[1]]['users'].plot(ax=zoom_ax, color='blue', linewidth=1)
main_ax.set_title(f'Number of users: {data.index.min()}, {data.index.max()}', fontsize=10)
zoom_ax.set_title(f'Number of users: {zoom}', fontsize=10)
In [5]:
# Interactive plot of time series
# ==============================================================================
data.loc[:end_train, 'partition'] = 'train'
data.loc[end_train:end_validation, 'partition'] = 'validation'
data.loc[end_validation:, 'partition'] = 'test'

fig = px.line(
    data_frame = data.reset_index(),
    x      = 'date_time',
    y      = 'users',
    color  = 'partition',
    title  = 'Number of users',
    width  = 800,
    height = 450
    width  = 800,
    height = 400,
    margin=dict(l=20, r=20, t=35, b=20),