If you found skforecast useful, you can support us with a donation. Your contribution will help to continue developing and improving this project. Many thanks! PayPal or Github Sponsors
Skforecast: time series forecasting with Python and Scikit-learn
More about forecasting
- ARIMA and SARIMAX models with python
- Forecasting time series with gradient boosting: Skforecast, XGBoost, LightGBM and CatBoost
- Forecasting electricity demand with Python
- Prediction intervals in forecasting models
- Intermittent demand forecasting with skforecast
- Forecasting web traffic with machine learning and Python
- Bitcoin price prediction with Python, when the past does not repeat itself
Introduction¶
A time series is a succession of chronologically ordered data spaced at equal or unequal intervals. The forecasting process consists of predicting the future value of a time series, either by modeling the series solely based on its past behavior (autoregressive) or by using other external variables.
This document describes how to use Scikit-learn regression models to perform forecasting on time series. Specifically, it introduces Skforecast, a simple library that contains the classes and functions necessary to adapt any Scikit-learn regression model to forecasting problems.
More examples in skforecast-examples.
Machine learning for forecasting¶
In order to apply machine learning models to forecasting problems, the time series has to be transformed into a matrix in which each value is related to the time window (lags) that precedes it.
In a time series context, a lag with respect to a time step $t$ is defined as the values of the series at previous time steps. For example, lag 1 is the value at time step $t − 1$ and lag $m$ is the value at time step $t − m$.
This type of transformation also allows to include additional variables.
Once data have been rearranged into the new shape, any regression model can be trained to predict the next value (step) of the series. During model training, every row is considered a separate data instance, where values at lags 1, 2, ... $p$ are considered predictors for the target quantity of the time series at time step $p + 1$.
Multi-Step Time Series Forecasting¶
When working with time series, it is seldom needed to predict only the next element in the series ($t_{+1}$). Instead, the most common goal is to predict a whole future interval (($t_{+1}$), ..., ($t_{+n}$)) or a far point in time ($t_{+n}$). Several strategies allow generating this type of prediction.
Recursive multi-step forecasting¶
Since the value $t_{n-1}$ is required to predict $t_{n}$, and $t_{n-1}$ is unknown, a recursive process is applied in which, each new prediction, is based on the previous one. This process is known as recursive forecasting or recursive multi-step forecasting and can be easily generated with the ForecasterAutoreg and ForecasterAutoregCustom classes.
Direct multi-step forecasting¶
Direct multi-step forecasting consists of training a different model for each step of the forecast horizon. For example, to predict the next 5 values of a time series, 5 different models are trained, one for each step. As a result, the predictions are independent of each other.
The main complexity of this approach is to generate the correct training matrices for each model. The ForecasterAutoregDirect class of the skforecast library automates this process. It is also important to bear in mind that this strategy has a higher computational cost since it requires the train of multiple models. The following diagram shows the process for a case in which the response variable and two exogenous variables are available.
Multiple output forecasting¶
Some machine learning models, such as long short-term memory (LSTM) neural network, can predict simultaneously several values of a sequence (one-shot). This strategy is not currently implemented in skforecast library.
Libraries¶
The libraries used in this document are:
# Data manipulation
# ==============================================================================
import numpy as np
import pandas as pd
# Plots
# ==============================================================================
import matplotlib.pyplot as plt
plt.style.use('fivethirtyeight')
plt.rcParams['lines.linewidth'] = 1.5
plt.rcParams['font.size'] = 10
# Modeling and Forecasting
# ==============================================================================
from sklearn.linear_model import LinearRegression
from sklearn.linear_model import Lasso
from sklearn.ensemble import RandomForestRegressor
from sklearn.metrics import mean_squared_error
from sklearn.metrics import mean_absolute_error
from sklearn.preprocessing import StandardScaler
from skforecast.ForecasterAutoreg import ForecasterAutoreg
from skforecast.ForecasterAutoregCustom import ForecasterAutoregCustom
from skforecast.ForecasterAutoregDirect import ForecasterAutoregDirect
from skforecast.model_selection import grid_search_forecaster
from skforecast.model_selection import backtesting_forecaster
from skforecast.utils import save_forecaster
from skforecast.utils import load_forecaster
# Warnings configuration
# ==============================================================================
import warnings
# warnings.filterwarnings('ignore')
Data¶
A time series is available with the monthly expenditure (millions of dollars) on corticosteroid drugs that the Australian health system had between 1991 and 2008. It is intended to create an autoregressive model capable of predicting future monthly expenditures.
The data used in the examples of this document have been obtained from the magnificent book Forecasting: Principles and Practice by Rob J Hyndman and George Athanasopoulos.
# Data download
# ==============================================================================
url = 'https://raw.githubusercontent.com/JoaquinAmatRodrigo/skforecast/master/data/h2o_exog.csv'
data = pd.read_csv(url, sep=',')
The column date has been stored as a string. To convert it to datetime the pd.to_datetime() function can be use. Once in datetime format, and to make use of Pandas functionalities, it is set as an index. Also, since the data is monthly, the frequency is set as Monthly Started 'MS'.
# Data preparation
# ==============================================================================
data = data.rename(columns={'fecha': 'date'})
data['date'] = pd.to_datetime(data['date'], format='%Y-%m-%d')
data = data.set_index('date')
data = data.rename(columns={'x': 'y'})
data = data.asfreq('MS')
data = data.sort_index()
data.head()
When setting a frequency with the asfreq() method, Pandas fills the gaps that may exist in the time series with the value of Null to ensure the indicated frequency. Therefore, it should be checked if missing values have appeared after this transformation.
print(f'Number of rows with missing values: {data.isnull().any(axis=1).mean()}')
Although it is unnecessary, since a frequency has been established, it is possible to verify that the time series is complete.
# Verify that a temporary index is complete
# ==============================================================================
(data.index == pd.date_range(start=data.index.min(),
end=data.index.max(),
freq=data.index.freq)).all()
# Fill gaps in a temporary index
# ==============================================================================
# data.asfreq(freq='30min', fill_value=np.nan)
The last 36 months are used as the test set to evaluate the predictive capacity of the model.
# Split data into train-test
# ==============================================================================
steps = 36
data_train = data[:-steps]
data_test = data[-steps:]
print(f"Train dates : {data_train.index.min()} --- {data_train.index.max()} (n={len(data_train)})")
print(f"Test dates : {data_test.index.min()} --- {data_test.index.max()} (n={len(data_test)})")
fig, ax = plt.subplots(figsize=(6, 2.5))
data_train['y'].plot(ax=ax, label='train')
data_test['y'].plot(ax=ax, label='test')
ax.legend();
Recursive autoregressive forecasting¶
ForecasterAutoreg¶
With the ForecasterAutoreg class, a model is created and trained from a RandomForestRegressor regressor with a time window of 6 lags. This means that the model uses the previous 6 months as predictors.
# Create and train forecaster
# ==============================================================================
forecaster = ForecasterAutoreg(
regressor = RandomForestRegressor(random_state=123),
lags = 6
)
forecaster.fit(y=data_train['y'])
forecaster
Predictions¶
Once the model is trained, the test data is predicted (36 months into the future).
# Predictions
# ==============================================================================
steps = 36
predictions = forecaster.predict(steps=steps)
predictions.head(5)
# Plot
# ==============================================================================
fig, ax = plt.subplots(figsize=(6, 2.5))
data_train['y'].plot(ax=ax, label='train')
data_test['y'].plot(ax=ax, label='test')
predictions.plot(ax=ax, label='predictions')
ax.legend();
Prediction error in the test set¶
The error that the model makes in its predictions is quantified. In this case, the metric used is the mean squared error (mse).
# Test error
# ==============================================================================
error_mse = mean_squared_error(
y_true = data_test['y'],
y_pred = predictions
)
print(f"Test error (mse): {error_mse}")
Hyperparameter tuning¶
The trained ForecasterAutoreg uses a 6 lag time window and a Random Forest model with the default hyperparameters. However, there is no reason why these values are the most suitable. Skforecast library provides the grid_search_forecaster function. It compares the results obtained with multiple combinations of hyperparameters and lags, and identify the best one.
🖉 Note
The computational cost of hyperparameter tuning depends heavily on the backtesting approach chosen to evaluate each hyperparameter combination. In general, the duration of the tuning process increases with the number of re-trains involved in the backtesting. To effectively speed up the prototyping phase, it is highly recommended to adopt a two-step strategy. First, userefit=False during the initial search to narrow down the range of values. Then, focus on the identified region of interest and apply a tailored backtesting strategy that meets the specific requirements of the use case. For additional tips on backtesting strategies, refer to Hyperparameter tuning and lags selection
# Hyperparameter grid search
# ==============================================================================
steps = 36
forecaster = ForecasterAutoreg(
regressor = RandomForestRegressor(random_state=123),
lags = 12 # This value will be replaced in the grid search
)
# Lags used as predictors
lags_grid = [10, 20]
# Regressor's hyperparameters
param_grid = {'n_estimators': [100, 500],
'max_depth': [3, 5, 10]}
results_grid = grid_search_forecaster(
forecaster = forecaster,
y = data_train['y'],
param_grid = param_grid,
lags_grid = lags_grid,
steps = steps,
refit = False,
metric = 'mean_squared_error',
initial_train_size = int(len(data_train)*0.5),
fixed_train_size = False,
return_best = True,
n_jobs = 'auto',
verbose = False
)
# Grid Search results
# ==============================================================================
results_grid
The best results are obtained using a time window of 20 lags and a Random Forest set up of {'max_depth': 3, 'n_estimators': 500}.
Final model¶
Finally, a ForecasterAutoreg is trained with the optimal configuration found by validation. This step is not necessary if return_best = True is specified in the grid_search_forecaster function.
# Create and train forecaster with the best hyperparameters
# ==============================================================================
regressor = RandomForestRegressor(max_depth=3, n_estimators=500, random_state=123)
forecaster = ForecasterAutoreg(
regressor = regressor,
lags = 20
)
forecaster.fit(y=data_train['y'])
# Predictions
# ==============================================================================
predictions = forecaster.predict(steps=steps)
# Plot
# ==============================================================================
fig, ax = plt.subplots(figsize=(6, 2.5))
data_train['y'].plot(ax=ax, label='train')
data_test['y'].plot(ax=ax, label='test')
predictions.plot(ax=ax, label='predictions')
ax.legend();
# Test error
# ==============================================================================
error_mse = mean_squared_error(
y_true = data_test['y'],
y_pred = predictions
)
print(f"Test error (mse): {error_mse}")
The optimal combination of hyperparameters significantly reduces test error.
Backtesting¶
The process of backtesting consists of evaluating the performance of a predictive model by applying it retrospectively to historical data. Therefore, it is a special type of cross-validation applied to the previous period(s).
🖉 Note
To ensure an accurate evaluation of your model and gain confidence in its predictive performance on new data, it is critical to employ an appropriate backtesting strategy. Factors such as use case characteristics, available computing resources and time intervals between predictions need to be considered to determine which strategy to use. In general, the more closely the backtesting process resembles the actual scenario in which the model is used, the more reliable the estimated metric will be. For more information about backtesting, visit Which strategy should I use?.Backtesting with refit and increasing training size (fixed origin)
The model is trained each time before making predictions. With this configuration, the model uses all the data available so far. It is a variation of the standard cross-validation but, instead of making a random distribution of the observations, the training set increases sequentially, maintaining the temporal order of the data.
Backtesting with refit and fixed training size (rolling origin)
A technique similar to the previous one but, in this case, the forecast origin rolls forward, therefore, the size of training remains constant. This is also known as time series cross-validation or walk-forward validation.
Backtesting with intermittent refit
The model is retrained every $n$ iterations of predictions.
🖉 Note
This strategy usually achieves a good balance between the computational cost of retraining and avoiding model degradation.
Backtesting without refit
After an initial train, the model is used sequentially without updating it and following the temporal order of the data. This strategy has the advantage of being much faster since the model is trained only once. However, the model does not incorporate the latest data available, so it may lose predictive capacity over time.
skforecast library has multiple backtesting strategies implemented. Regardless of which one is used, it is important not to include test data in the search process to avoid overfitting problems.
For this example, a backtesting with refit strategy is followed. Internally, the process that the function applies is:
In the first iteration, the model is trained with the observations selected for the initial training (in this case, 87). Then, the next 36 observations are used to validate the predictions of this first model.
In the second iteration, the model is retrained by adding, to the initial training set, the previous 36 validation observations (87 + 36). In the same way, the next 36 observations are established as the new validation set.
This process is repeated until all available observations are used. Following this strategy, the training set increases in each iteration with as many observations as steps are being predicted.
# Backtesting
# ==============================================================================
steps = 36
n_backtesting = 36*3 # The last 9 years are separated for the backtest
metric, predictions_backtest = backtesting_forecaster(
forecaster = forecaster,
y = data['y'],
initial_train_size = len(data) - n_backtesting,
fixed_train_size = False,
steps = steps,
metric = 'mean_squared_error',
refit = True,
verbose = True,
show_progress = True
)
print(f"Backtest error: {metric}")
# Plot backtest predictions vs real values
# ==============================================================================
fig, ax = plt.subplots(figsize=(6, 2.5))
data.loc[predictions_backtest.index, 'y'].plot(ax=ax, label='test')
predictions_backtest.plot(ax=ax, label='predictions')
ax.legend();
Predictors importance¶
Since the ForecasterAutoreg object uses Scikit-learn models, the importance of predictors can be accessed once trained. When the regressor used is a LinearRegression(), Lasso() or Ridge(), the coefficients of the model reflect their importance. In GradientBoostingRegressor() or RandomForestRegressor() regressors, the importance of predictors is based on impurity.
🖉 Note
get_feature_importances() method only returns values if the regressor used within the forecaster has the attribute coef_ or feature_importances_.
# Predictors importances
# ==============================================================================
forecaster.get_feature_importances()
Recursive autoregressive forecasting with exogenous variables¶
In the previous example, only lags of the predicted variable itself have been used as predictors. In certain scenarios, it is possible to have information about other variables, whose future value is known, so could serve as additional predictors in the model.
Continuing with the previous example, a new variable whose behavior is correlated with the modeled time series and it is wanted to incorporate as a predictor is simulated. The same applies to multiple exogenous variables.
Data¶
# Data download
# ==============================================================================
url = 'https://raw.githubusercontent.com/JoaquinAmatRodrigo/skforecast/master/data/h2o_exog.csv'
data = pd.read_csv(url, sep=',')
# Data preparation
# ==============================================================================
data = data.rename(columns={'fecha': 'date'})
data['date'] = pd.to_datetime(data['date'], format='%Y-%m-%d')
data = data.set_index('date')
data = data.asfreq('MS')
data = data.sort_index()
fig, ax = plt.subplots(figsize=(6, 2.5))
data['y'].plot(ax=ax, label='y')
data['exog_1'].plot(ax=ax, label='exogenous variable')
ax.legend();
# Split data into train-test
# ==============================================================================
steps = 36
data_train = data[:-steps]
data_test = data[-steps:]
print(f"Train dates : {data_train.index.min()} --- {data_train.index.max()} (n={len(data_train)})")
print(f"Test dates : {data_test.index.min()} --- {data_test.index.max()} (n={len(data_test)})")
ForecasterAutoreg¶
# Create and train forecaster
# ==============================================================================
forecaster = ForecasterAutoreg(
regressor = RandomForestRegressor(random_state=123),
lags = 8
)
forecaster.fit(y=data_train['y'], exog=data_train['exog_1'])
forecaster
Predictions¶
If the ForecasterAutoreg is trained with an exogenous variable, the value of this variable must be passed to predict(). It is only applicable to scenarios in which future information on the exogenous variable is available.
# Predictions
# ==============================================================================
predictions = forecaster.predict(steps=steps, exog=data_test['exog_1'])
# Plot
# ==============================================================================
fig, ax = plt.subplots(figsize=(6, 2.5))
data_train['y'].plot(ax=ax, label='train')
data_test['y'].plot(ax=ax, label='test')
predictions.plot(ax=ax, label='predictions')
ax.legend();
Prediction error in the test set¶
# Test error
# ==============================================================================
error_mse = mean_squared_error(
y_true = data_test['y'],
y_pred = predictions
)
print(f"Test error (mse): {error_mse}")
Hyperparameter tuning¶
# Hyperparameter Grid search
# ==============================================================================
steps = 36
forecaster = ForecasterAutoreg(
regressor = RandomForestRegressor(random_state=123),
lags = 12 # This value will be replaced in the grid search
)
lags_grid = [5, 12, 20]
param_grid = {'n_estimators': [50, 100, 500],
'max_depth': [3, 5, 10]}
results_grid = grid_search_forecaster(
forecaster = forecaster,
y = data_train['y'],
exog = data_train['exog_1'],
param_grid = param_grid,
lags_grid = lags_grid,
steps = steps,
refit = False,
metric = 'mean_squared_error',
initial_train_size = int(len(data_train)*0.5),
return_best = True,
n_jobs = 'auto',
verbose = False
)
# Grid Search results
# ==============================================================================
results_grid.head()
The best results are obtained using a time window of 12 lags and a Random Forest set up of {'max_depth': 10, 'n_estimators': 50}.
Final model¶
Setting return_best = True in grid_search_forecaster, after the search, the ForecasterAutoreg object has been modified and trained with the best configuration found.
# Predictions
# ==============================================================================
predictions = forecaster.predict(steps=steps, exog=data_test['exog_1'])
# Plot
# ==============================================================================
fig, ax = plt.subplots(figsize=(6, 2.5))
data_train['y'].plot(ax=ax, label='train')
data_test['y'].plot(ax=ax, label='test')
predictions.plot(ax=ax, label='predictions')
ax.legend();
# Test error
# ==============================================================================
error_mse = mean_squared_error(y_true = data_test['y'], y_pred = predictions)
print(f"Test error (mse) {error_mse}")
Recursive autoregressive forecasting with custom predictors¶
In addition to the lags, it may be interesting to incorporate other characteristics of the time series in some scenarios. For example, the moving average of the last n values could be used to capture the series's trend.
The ForecasterAutoregCustom class behaves very similar to the ForecasterAutoreg class seen in the previous sections, but with the difference that it is the user who defines the function used to create the predictors.
The first example of the document about predicting the last 36 months of the time series is repeated. In this case, the predictors are the first 10 lags and the values' moving average of the last 20 months.
Data¶
# Data download
# ==============================================================================
url = 'https://raw.githubusercontent.com/JoaquinAmatRodrigo/skforecast/master/data/h2o_exog.csv'
data = pd.read_csv(url, sep=',')
# Data preparation
# ==============================================================================
data = data.rename(columns={'fecha': 'date'})
data['date'] = pd.to_datetime(data['date'], format='%Y-%m-%d')
data = data.set_index('date')
data = data.rename(columns={'x': 'y'})
data = data.asfreq('MS')
data = data.sort_index()
# Split data into train-test
# ==============================================================================
steps = 36
data_train = data[:-steps]
data_test = data[-steps:]
print(f"Train dates : {data_train.index.min()} --- {data_train.index.max()} (n={len(data_train)})")
print(f"Test dates : {data_test.index.min()} --- {data_test.index.max()} (n={len(data_test)})")
ForecasterAutoregCustom¶
A ForecasterAutoregCustom is created and trained from a RandomForestRegressor regressor. The create_predictor() function, which calculates the first 10 lags and the moving average of the last 20 values, is used to create the predictors.
# Function to calculate predictors from time series
# ==============================================================================
def custom_predictors(y):
'''
Create first 10 lags of a time series.
Calculate moving average with window 20.
'''
lags = y[-1:-11:-1]
mean = np.mean(y[-20:])
predictors = np.hstack([lags, mean])
return predictors
⚠ Warning
When creating the forecaster, thewindow_size argument must be equal to or greater than the window used by the function that creates the predictors. This value, in this case, is 20.
# Create and train forecaster
# ==============================================================================
forecaster = ForecasterAutoregCustom(
regressor = RandomForestRegressor(random_state=123),
fun_predictors = custom_predictors,
window_size = 20
)
forecaster.fit(y=data_train['y'])
forecaster
It is possible to access the custom function used to create the predictors.
print(forecaster.source_code_fun_predictors)
Using the method create_train_X_y, is it posible to acces the matrices that are created internally in the training process.
X, y = forecaster.create_train_X_y(y=data_train['y'])
X.head(4)
y.head(4)
Predictions¶
# Predictions
# ==============================================================================
steps = 36
predictions = forecaster.predict(steps=steps)
# Plot
# ==============================================================================
fig, ax = plt.subplots(figsize=(6, 2.5))
data_train['y'].plot(ax=ax, label='train')
data_test['y'].plot(ax=ax, label='test')
predictions.plot(ax=ax, label='predictions')
ax.legend();
Prediction error in the test set¶
# Test error
# ==============================================================================
error_mse = mean_squared_error(
y_true = data_test['y'],
y_pred = predictions
)
print(f"Test error (mse): {error_mse}")
Hyperparameter tuning¶
When using the grid_search_forecaster function with a ForecasterAutoregCustom, thelags_grid argument is not specified.
# Hyperparameter Grid search
# ==============================================================================
steps = 36
forecaster = ForecasterAutoregCustom(
regressor = RandomForestRegressor(random_state=123),
fun_predictors = custom_predictors,
window_size = 20
)
# Regressor's hyperparameters
param_grid = {'n_estimators': [100, 500],
'max_depth': [3, 5, 10]}
results_grid = grid_search_forecaster(
forecaster = forecaster,
y = data_train['y'],
param_grid = param_grid,
steps = steps,
refit = False,
metric = 'mean_squared_error',
initial_train_size = int(len(data_train)*0.5),
fixed_train_size = False,
return_best = True,
n_jobs = 'auto',
verbose = False
)
# Grid Search results
# ==============================================================================
results_grid.head(5)
Final model¶
# Predictions
# ==============================================================================
predictions = forecaster.predict(steps=steps)
# Plot
# ==============================================================================
fig, ax = plt.subplots(figsize=(6, 2.5))
data_train['y'].plot(ax=ax, label='train')
data_test['y'].plot(ax=ax, label='test')
predictions.plot(ax=ax, label='predictions')
ax.legend();
# Test error
# ==============================================================================
error_mse = mean_squared_error(y_true = data_test['y'], y_pred = predictions)
print(f"Test error (mse) {error_mse}")
Direct multi-step forecasting¶
The ForecasterAutoreg and ForecasterAutoregCustom models follow a recursive prediction strategy in which each new prediction builds on the previous one. An alternative is to train a model for each of the steps to be predicted. This strategy, commonly known as direct multi-step forecasting, is computationally more expensive than recursive since it requires training several models. However, in some scenarios, it achieves better results. These kinds of models can be obtained with the ForecasterAutoregDirect class and can include one or multiple exogenous variables.
⚠ Warning
`ForecasterAutoregDirect` may require very long training times, as one model is fitted for each step.
ForecasterAutoregDirect¶
Unlike when using ForecasterAutoreg or ForecasterAutoregCustom, the number of steps to be predicted must be indicated in the ForecasterAutoregDirect type models. This means that the number of predictions obtained when executing the predict() method is always the same. It is not possible to predict steps beyond the value defined at their creation.
For this example, a linear model with Lasso penalty is used as a regressor. These models require the predictors to be standardized, so it is combined with a StandardScaler.
For more detailed documentation how to use transformers and pipelines, visit: skforecast with scikit-learn and transformers pipelines.
# Create and train forecaster
# ==============================================================================
forecaster = ForecasterAutoregDirect(
regressor = Lasso(random_state=123),
transformer_y = StandardScaler(),
steps = 36,
lags = 8
)
forecaster
# Hyperparameter Grid search
# ==============================================================================
from skforecast.exceptions import LongTrainingWarning
warnings.simplefilter('ignore', category=LongTrainingWarning)
forecaster = ForecasterAutoregDirect(
regressor = Lasso(random_state=123),
transformer_y = StandardScaler(),
steps = 36,
lags = 8 # This value will be replaced in the grid search
)
param_grid = {'alpha': np.logspace(-5, 5, 10)}
lags_grid = [5, 12, 20]
results_grid = grid_search_forecaster(
forecaster = forecaster,
y = data_train['y'],
param_grid = param_grid,
lags_grid = lags_grid,
steps = 36,
refit = False,
metric = 'mean_squared_error',
initial_train_size = int(len(data_train)*0.5),
fixed_train_size = False,
return_best = True,
n_jobs = 'auto',
verbose = False
)
# Grid Search results
# ==============================================================================
results_grid.head()
The best results are obtained using a time window of 12 lags and a Lasso setting {'alpha': 0.021544}.
# Predictions
# ==============================================================================
predictions = forecaster.predict()
# Gráfico
# ==============================================================================
fig, ax = plt.subplots(figsize=(6, 2.5))
data_train['y'].plot(ax=ax, label='train')
data_test['y'].plot(ax=ax, label='test')
predictions.plot(ax=ax, label='predictions')
ax.legend();
# Test error
# ==============================================================================
error_mse = mean_squared_error(y_true = data_test['y'], y_pred = predictions)
print(f"Test error (mse) {error_mse}")
Prediction intervals¶
A prediction interval defines the interval within which the true value of $y$ is expected to be found with a given probability.
Rob J Hyndman and George Athanasopoulos, list in their book Forecasting: Principles and Practice multiple ways to estimate prediction intervals, most of which require that the residuals (errors) of the model are distributed in a normal way. When this property cannot be assumed, bootstrapping can be resorted to, which only assumes that the residuals are uncorrelated. This is the method used in the Skforecast library. For more information visit skforecast probabilistic forecasting.
# Data download
# ==============================================================================
url = 'https://raw.githubusercontent.com/JoaquinAmatRodrigo/skforecast/master/data/h2o_exog.csv'
data_raw = pd.read_csv(url, sep=',')
# Data preparation
# ==============================================================================
data = data_raw.copy()
data = data.rename(columns={'fecha': 'date'})
data['date'] = pd.to_datetime(data['date'], format='%Y-%m-%d')
data = data.set_index('date')
data = data.rename(columns={'x': 'y'})
data = data.asfreq('MS')
data = data.sort_index()
# Split data into train-test
# ==============================================================================
steps = 36
data_train = data[:-steps]
data_test = data[-steps:]
print(f"Train dates : {data_train.index.min()} --- {data_train.index.max()} (n={len(data_train)})")
print(f"Test dates : {data_test.index.min()} --- {data_test.index.max()} (n={len(data_test)})")
# Create and train forecaster
# ==============================================================================
forecaster = ForecasterAutoreg(
regressor = LinearRegression(),
lags = 15
)
forecaster.fit(y=data_train['y'])
# Prediction intervals
# ==============================================================================
predictions = forecaster.predict_interval(
steps = steps,
interval = [1, 99],
n_boot = 500
)
predictions.head(5)
# Prediction error
# ==============================================================================
error_mse = mean_squared_error(
y_true = data_test['y'],
y_pred = predictions.iloc[:, 0]
)
print(f"Test error (mse): {error_mse}")
# Plot
# ==============================================================================
fig, ax = plt.subplots(figsize=(6, 2.5))
data_test['y'].plot(ax=ax, label='test')
predictions['pred'].plot(ax=ax, label='prediction')
ax.fill_between(
predictions.index,
predictions['lower_bound'],
predictions['upper_bound'],
color = 'red',
alpha = 0.2
)
ax.legend();