If you like Skforecast , please give us a star on GitHub! ⭐️
Reducing the influence of Covid-19 on time series forecasting models using skforecast
More about forecasting: Skforecast examples
Weighted time series forecasting¶
Although historical data is available in many real use cases of forecasting, not all are reliable. Some examples of these scenarios are:
IoT sensors: within the Internet of Things, sensors capture the raw data from the physical world. Often the sensors are deployed or installed in harsh environments. This inevitably means that the sensors are prone to failure, malfunction, and rapid attrition, causing the sensor to produce unusual and erroneous readings.
Factory shutdown: every certain period of operation, factories need to be shut down for repair, overhaul, or maintenance activities. These events cause production to stop, generating a gap in the data.
Pandemic (Covid-19): the Covid 19 pandemic changed population behavior significantly, directly impacting many time series such as production, sales, and transportation.
The presence of unreliable or unrepresentative values in the data history is a major problem, as it hinders model learning. For most forecasting algorithms, removing that part of the data is not an option because they require the time series to be complete. An alternative solution is to reduce the weight of the affected observations during model training. This document shows two examples of how skforecast makes it easy to apply this strategy.
Note
In the following examples, a portion of the time series is excluded from model training by giving it a weight of zero. However, the use of weights is not limited to including or excluding observations, but to balancing the degree of influence of each observation in the forecasting model. For example, an observation with a weight of 10 has 10 times more impact on the model training than an observation with a weight of 1.Warning
In most gradient boosting implementations (LightGBM, XGBoost, CatBoost), samples with zero weight are ignored when calculating the gradients and hessians. However, the values for those samples are still considered when building the feature histograms. Therefore, the resulting model may differ from the model trained without the zero-weighted samples. See more details in this issue.Libraries¶
# Libraries
# ==============================================================================
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
plt.style.use('fivethirtyeight')
plt.rcParams['lines.linewidth'] = 1.5
dark_style = {
'figure.facecolor': '#212946',
'axes.facecolor': '#212946',
'savefig.facecolor':'#212946',
'axes.grid': True,
'axes.grid.which': 'both',
'axes.spines.left': False,
'axes.spines.right': False,
'axes.spines.top': False,
'axes.spines.bottom': False,
'grid.color': '#2A3459',
'grid.linewidth': '1',
'text.color': '0.9',
'axes.labelcolor': '0.9',
'xtick.color': '0.9',
'ytick.color': '0.9',
'font.size': 12
}
plt.rcParams.update(dark_style)
from sklearn.linear_model import Ridge
from sklearn.metrics import mean_absolute_error
from skforecast.ForecasterAutoreg import ForecasterAutoreg
from skforecast.model_selection import backtesting_forecaster
Covid-19 lockdown¶
During the lockdown period imposed as a consequence of the covid-19 pandemic, the behavior of the population was altered. An example of this can be seen in the use of the bicycle rental service in the city of Madrid (Spain).
Data¶
# Data download
# ==============================================================================
url = (
'https://raw.githubusercontent.com/JoaquinAmatRodrigo/'
'Estadistica-machine-learning-python/master/data/usuarios_diarios_bicimad.csv'
)
data = pd.read_csv(url, sep=',')
# Data preprocessing
# ==============================================================================
data['fecha'] = pd.to_datetime(data['fecha'], format='%Y-%m-%d')
data = data[['fecha', 'Usos bicis total día']]
data.columns = ['date', 'users']
data = data.set_index('date')
data = data.asfreq('D')
data = data.sort_index()
# Split data into train-val-test
# ==============================================================================
data = data.loc['2020-01-01': '2021-12-31']
end_train = '2021-06-01'
data_train = data.loc[: end_train, :]
data_test = data.loc[end_train:, :]
print(f"Dates train : {data_train.index.min()} --- {data_train.index.max()} (n={len(data_train)})")
print(f"Dates test : {data_test.index.min()} --- {data_test.index.max()} (n={len(data_test)})")
# Time series plot
# ==============================================================================
fig, ax = plt.subplots(figsize=(12, 4))
data_train.users.plot(ax=ax, label='train', linewidth=1)
data_test.users.plot(ax=ax, label='test', linewidth=1)
ax.axvspan(
pd.to_datetime('2020-03-16'),
pd.to_datetime('2020-04-21'),
label="Covid-19 confinement",
color="red",
alpha=0.3
)
ax.axvspan(
pd.to_datetime('2020-04-21'),
pd.to_datetime('2020-05-31'),
label="Recovery time",
color="white",
alpha=0.3
)
ax.set_title('Number of users BiciMAD')
ax.legend();
Include the whole time series¶
A forecaster is initialized without taking into consideration the lockdown period.
# Create recursive multi-step forecaster (ForecasterAutoreg)
# ==============================================================================
forecaster = ForecasterAutoreg(
regressor = Ridge(),
lags = 21,
)
forecaster
Once the model is created, a backtesting process is run to simulate the behavior of the forecaster if it had predicted the test set in 10-day batches.
# Backtesting: predict next 7 days at a time.
# ==============================================================================
metric, predictions_backtest = backtesting_forecaster(
forecaster = forecaster,
y = data.users,
initial_train_size = len(data.loc[:end_train]),
fixed_train_size = False,
steps = 7,
metric = 'mean_absolute_error',
refit = False,
verbose = False
)
print(f"Backtest error: {metric:.1f}")
Exclude part of the time series¶
To minimize the influence on the model of these dates, a custom function is created weights following the rules:
Weight of 0 if index date is:
Within the lockdown period (2020-03-16 to 2020-04-21).
Within the recovery period (2020-04-21 to 2020-05-31).
21 days after the recovery period to avoid including impacted values as lags (2020-05-31) to 2020-06-21).
Weight of 1 otherwise.
If an observation has a weight of 0, it has no influence during model training.
# Custom function to create weights
# ==============================================================================
def custom_weights(index):
"""
Return 0 if index is one between 2020-03-16 and 2020-06-21.
"""
weights = np.where((index >= '2020-03-16') & (index <= '2020-06-21'), 0, 1)
return weights
Again, a ForecasterAutoreg is initialized but this time including the custom_weights function.
# Create recursive multi-step forecaster (ForecasterAutoreg)
# ==============================================================================
forecaster = ForecasterAutoreg(
regressor = Ridge(random_state=123),
lags = 21,
weight_func = custom_weights
)
# Backtesting: predict next 7 days at a time.
# ==============================================================================
metric, predictions_backtest = backtesting_forecaster(
forecaster = forecaster,
y = data.users,
initial_train_size = len(data.loc[:end_train]),
fixed_train_size = False,
steps = 7,
metric = 'mean_absolute_error',
refit = False,
verbose = False
)
print(f"Backtest error: {metric:.1f}")
Giving a weight of 0 to the lockdown period (excluding it from the model training) slightly improves the forecasting performance.
Power plant shutdown¶
Power plants used to generate energy are very complex installations that require a high level of maintenance. It is common that, every certain period of operation, the plant has to be shut down for repair, overhaul, or maintenance activities.
Data¶
# Data download
# ==============================================================================
url = ('https://raw.githubusercontent.com/JoaquinAmatRodrigo/skforecast/master/'
'data/energy_production_shutdown.csv')
data = pd.read_csv(url, sep=',')
# Data preprocessing
# ==============================================================================
data['date'] = pd.to_datetime(data['date'], format='%Y-%m-%d')
data = data.set_index('date')
data = data.asfreq('D')
data = data.sort_index()
data.head()
# Split data into train-test
# ==============================================================================
data = data.loc['2012-01-01 00:00:00': '2014-12-30 23:00:00']
end_train = '2013-12-31 23:59:00'
data_train = data.loc[: end_train, :]
data_test = data.loc[end_train:, :]
print(f"Dates train : {data_train.index.min()} --- {data_train.index.max()} (n={len(data_train)})")
print(f"Dates test : {data_test.index.min()} --- {data_test.index.max()} (n={len(data_test)})")
# Time series plot
# ==============================================================================
fig, ax = plt.subplots(figsize=(12, 4))
data_train.production.plot(ax=ax, label='train', linewidth=1)
data_test.production.plot(ax=ax, label='test', linewidth=1)
ax.axvspan(
pd.to_datetime('2012-06-01'),
pd.to_datetime('2012-09-30'),
label="Shutdown",
color="red",
alpha=0.1
)
ax.set_title('Energy production')
ax.legend();
Include the whole time series¶
A forecaster is initialized without taking in consideration the shutdown period.
# Create recursive multi-step forecaster (ForecasterAutoreg)
# ==============================================================================
forecaster = ForecasterAutoreg(
regressor = Ridge(random_state=123),
lags = 21,
)
# Backtesting: predict next 10 days at a time.
# ==============================================================================
metric, predictions_backtest = backtesting_forecaster(
forecaster = forecaster,
y = data.production,
initial_train_size = len(data.loc[:end_train]),
fixed_train_size = False,
steps = 10,
metric = 'mean_absolute_error',
refit = False,
verbose = False
)
print(f"Backtest error: {metric:.1f}")
Exclude part of the time series¶
The factory shutdown took place from 2012-06-01 to 2012-09-30. To minimize the influence on the model of these dates, a custom function is created that gives a value of 0 if the index date is within the shutdown period or 21 days later (lags used by the model) and 1 otherwise. If an observation has a weight of 0, it has no influence at all during model training.
# Custom function to create weights
# ==============================================================================
def custom_weights(index):
"""
Return 0 if index is one between 2012-06-01 and 2012-10-21.
"""
weights = np.where((index >= '2012-06-01') & (index <= '2012-10-21'), 0, 1)
return weights
# Create recursive multi-step forecaster (ForecasterAutoreg)
# ==============================================================================
forecaster = ForecasterAutoreg(
regressor = Ridge(random_state=123),
lags = 21,
weight_func = custom_weights
)
# Backtesting: predict next 10 days at a time.
# ==============================================================================
metric, predictions_backtest = backtesting_forecaster(
forecaster = forecaster,
y = data.production,
initial_train_size = len(data.loc[:end_train]),
fixed_train_size = False,
steps = 12,
metric = 'mean_absolute_error',
refit = False,
verbose = False
)
print(f"Backtest error: {metric:.1f}")
As in the previous example, excluding the observations during the shutdown period slightly improves the forecasting performance.
Session information¶
import session_info
session_info.show(html=False)
¿How to cite this document?
Reducing the influence of Covid-19 on time series forecasting models using skforecast by Joaquín Amat Rodrigo and Javier Escobar Ortiz, available under a CC BY-NC-SA 4.0 at https://www.cienciadedatos.net/documentos/py45-weighted-time-series-forecasting.html 
Did you like the article? Your support is important
Website maintenance has high cost, your contribution will help me to continue generating free educational content. Many thanks! 😊
This work by Joaquín Amat Rodrigo and Javier Escobar Ortiz is licensed under a Attribution-NonCommercial-ShareAlike 4.0 International.