Easy Forecast

Forecasting methods and algorithms from:

Introduction to Time Series and Forecasting by Peter J. BrockwellRichard A. Davis.

The state-space Arima models and other methods will be implemented in the future.

For now, only ARAR algorithm is implemented from Peter J. BrockwellRichard A. Davis[2016]

• Free software: MIT license
• Install from GitHub "pip install git+https://github.com/Akai01/easyforecast"
• Documentation: https://easyforecast.readthedocs.io/en/latest/
• The usage:

Forecasting using ARAR algorithm:

from easyforecast.arar import ARAR
import pandas as pd
from prophet import Prophet
from easyforecast.utils import accuracy
import numpy as np


df = pd.read_csv("https://raw.githubusercontent.com/Akai01/example-time-series-datasets/main/Data/retail.csv", sep= ",")
df.head()

df_sub = df[['date', 'series_38']]
df_sub.columns = ["ds", "y"]
df_sub.head()

train = df_sub[:-12]
test = df_sub[-12:]
test_series = pd.DataFrame(pd.array(test['y']), index= pd.to_datetime( test["ds"]))
test_series.columns = ["test"]
train.head()

# forecast using ARAR model
arar = ARAR(train, h = 12, freq = "MS")
arar.forecast()
arar_forecast = arar.get_forecast()
arar.accuracy(test_set = test)

test_series.plot()
arar.plot()

Prophet forecast vs ARAR:

m = Prophet()
m.fit(train)
future = m.make_future_dataframe(periods=12, freq = "MS", include_history= False)
prophet_forecast = m.predict(future)
prophet_forecast[['ds', 'yhat', 'yhat_lower', 'yhat_upper']]

accuracy(np.array(test["y"]), np.array(prophet_forecast["yhat"]))

accuracy(np.array(test["y"]), np.array(arar_forecast["mean"]))

MAPE ARAR = 1.9 vs MAPE Prophet = 3.4

Forecast using auto_arima:

from easyforecast.arima import ARIMA
arima = ARIMA(train, freq = "MS")
arima.auto_arima(m = 12, ic = "aicc")
arima.forecast(h = 12)
arima.get_forecast()
arima.accuracy(test_set = test)
test_series.plot()
arima.plot()

Features

• Automatic model selection
• Visualisation of the forecast models
• Risidual diagnostics
• Error diagnostics