Example: Univariate forecast¶

---

This example shows how to use ATOM to work with a univariate time series dataset.

Import the airline dataset from sktime.datasets. This is a small and easy to train dataset that measures monthly totals of international airline passengers from 1949 to 1960.

Load the data¶

In [1]:

# Import packages
import numpy as np
from sktime.datasets import load_airline
from atom import ATOMForecaster

# Import packages import numpy as np from sktime.datasets import load_airline from atom import ATOMForecaster

In [2]:

# Load the data
y = load_airline()

print(y)

# Load the data y = load_airline() print(y)

Period
1949-01    112.0
1949-02    118.0
1949-03    132.0
1949-04    129.0
1949-05    121.0
           ...  
1960-08    606.0
1960-09    508.0
1960-10    461.0
1960-11    390.0
1960-12    432.0
Freq: M, Name: Number of airline passengers, Length: 144, dtype: float64

Analyze the data¶

In [3]:

atom = ATOMForecaster(y, verbose=2, random_state=1)

atom = ATOMForecaster(y, verbose=2, random_state=1)

<< ================== ATOM ================== >>

Configuration ==================== >>
Algorithm task: Univariate forecast.

Dataset stats ==================== >>
Shape: (144, 1)
Train set size: 116
 --> From: 1949-01  To: 1958-08
Test set size: 28
 --> From: 1958-09  To: 1960-12
-------------------------------------
Memory: 6.47 kB
Duplicates: 26 (18.1%)

In [4]:

# Since the dataset contains only the target column, atom.X is empty
atom.X

# Since the dataset contains only the target column, atom.X is empty atom.X

Out[4]:

Period
1949-01
1949-02
1949-03
1949-04
1949-05
...
1960-08
1960-09
1960-10
1960-11
1960-12

144 rows × 0 columns

In [5]:

# Examine the dataset
atom.plot_series()

# Examine the dataset atom.plot_series()

In [6]:

atom.plot_qq()

atom.plot_qq()

Seasonality¶

In [7]:

# ATOM has a number of plots to understand the seasonality of the data
with atom.canvas(rows=2, cols=1, sharex=True, vspace=0.01, legend=None):
    atom.plot_acf()
    atom.plot_pacf()

# ATOM has a number of plots to understand the seasonality of the data with atom.canvas(rows=2, cols=1, sharex=True, vspace=0.01, legend=None): atom.plot_acf() atom.plot_pacf()

In [8]:

with atom.canvas(rows=2, cols=1, sharex=True, vspace=0.01, legend=None):
    atom.plot_periodogram()
    atom.plot_fft()

with atom.canvas(rows=2, cols=1, sharex=True, vspace=0.01, legend=None): atom.plot_periodogram() atom.plot_fft()

In [9]:

# It's also possible to compute the seasonality
sp = atom.get_seasonal_period()
sp

# It's also possible to compute the seasonality sp = atom.get_seasonal_period() sp

Out[9]:

[12, 24, 36, 11, 48]

In [10]:

# And use that seasonality
atom.sp = sp

# And use that seasonality atom.sp = sp

Run the pipeline¶

In [11]:

# Use the regular data cleaning methods to transform the target column
atom.scale(columns=-1)

# Use the regular data cleaning methods to transform the target column atom.scale(columns=-1)

Fitting Scaler...
Scaling features...

In [12]:

atom.y

atom.y

Out[12]:

Period
1949-01   -1.388211
1949-02   -1.324254
1949-03   -1.175021
1949-04   -1.207000
1949-05   -1.292275
             ...   
1960-08    3.877562
1960-09    2.832935
1960-10    2.331940
1960-11    1.575119
1960-12    2.022816
Freq: M, Name: Number of airline passengers, Length: 144, dtype: float64

In [13]:

atom.run(["AutoARIMA", "TBATS", "Prophet"])

atom.run(["AutoARIMA", "TBATS", "Prophet"])

Training ========================= >>
Models: AutoARIMA, TBATS, Prophet
Metric: mape


Results for AutoARIMA:
Fit ---------------------------------------------
Train evaluation --> mape: -1.2991
Test evaluation --> mape: -0.1214
Time elapsed: 3.857s
-------------------------------------------------
Time: 3.857s


Results for TBATS:
Fit ---------------------------------------------
Train evaluation --> mape: -1.1261
Test evaluation --> mape: -0.3558
Time elapsed: 37.379s
-------------------------------------------------
Time: 37.379s


Results for Prophet:
Fit ---------------------------------------------
Train evaluation --> mape: -1.0705
Test evaluation --> mape: -0.2015
Time elapsed: 0.328s
-------------------------------------------------
Time: 0.328s


Final results ==================== >>
Total time: 41.570s
-------------------------------------
AutoARIMA --> mape: -0.1214 !
TBATS     --> mape: -0.3558
Prophet   --> mape: -0.2015

In [14]:

# It's also possible to use regression models for forecast tasks
atom.run("RF")

# It's also possible to use regression models for forecast tasks atom.run("RF")

Training ========================= >>
Models: RF
Metric: mape


Results for RandomForest:
Fit ---------------------------------------------
Train evaluation --> mape: nan
Test evaluation --> mape: -0.4553
Time elapsed: 0.220s
-------------------------------------------------
Time: 0.220s


Final results ==================== >>
Total time: 0.222s
-------------------------------------
RandomForest --> mape: -0.4553

In [15]:

atom.rf.estimator

atom.rf.estimator

Out[15]:

RecursiveTabularRegressionForecaster(estimator=RandomForestRegressor())

Please rerun this cell to show the HTML repr or trust the notebook.

Analyze the results¶

In [16]:

atom.evaluate()

atom.evaluate()

Out[16]:

	mae	mape	mse	r2	rmse
AutoARIMA	-0.209400	-0.121400	-0.062900	0.910300	-0.250700
TBATS	-0.555500	-0.355800	-0.392200	0.440300	-0.626300
Prophet	-0.350100	-0.201500	-0.181200	0.741500	-0.425600
RF	-0.723300	-0.455300	-0.756200	-0.079100	-0.869600

In [17]:

atom.winner.plot_forecast()

atom.winner.plot_forecast()

In [18]:

atom.plot_errors()

atom.plot_errors()