Take Forecast

A time series forecasting library that uses Kedro pipelines.

For performing the forecast, there are four main steps:

• Tune: Analyse the time series to find seasonality, unit roots, memory lags and stochastic process lags.

• Fit: Fit a SARIMA model with hyperparameters obtained in the last step and input data.

• Predict: Make forecast using the fitted model. Also provide upper and lower boundaries for confidence interval.

• Evaluate: Assess the model quality by error metrics evaluated in train and test sets.

Installation

Use pip to install:

pip install take-forecast

Usage

A simple example using take_forecast.

import numpy as np
import pandas as pd
from take_forecast import ProjectContext
np.random.seed(0)
samples = 60
x = 100 + 80 * np.linspace(0, 1, samples) + 20 * np.random.randn(samples)
y = x[:-3] + x[1:-2] + x[2:-1] + x[3:]
date_end = pd.Timestamp.now()
date_start = date_end - pd.Timedelta(len(y) - 1, 'd')
index = pd.date_range(date_start, date_end, freq='d')
ts = pd.Series(y, index)
context = ProjectContext()
results = context.run_forecast(ts)

The results are informed as a report in a dict with following keys:

• forecast: Prediction
• forecast_lower: Lower limit
• forecast_upper: Upper limit
• alpha: Significance level
• error_metrics_report: Train and test error metrics
• model: Fitted SARIMA model

Author

Take Blip - Data & Analytics - Research Tribe - Squad XD