chronos-forecast 0.1.0

Time series prediction using probabilistic programming

Chronos

Simple time series prediction model. Implemented using Pyro and greatly inspired by Prophet.

Requirements:

• python >= 3.7
• pandas >= 1.1
• numpy >= 1.19
• matplotlib >= 3.2
• torch >= 1.5
• pyro-ppl >= 1.3

Installation

To install chronos, run the following command from your terminal:

pip install chronos-forecast

Simple Use Case

With the files included, you can load the Divvy bike daily data (the data has been aggregated since the original file is 2GB) as follows:

import pandas as pd
import numpy as np

divvy_data = pd.read_csv('data/divvy_daily_rides.csv')
divvy_data['ds'] = pd.to_datetime(divvy_data['ds'])
print(divvy_data.head())

          ds          y
0 2014-01-01  105421324
1 2014-01-02  123221770
2 2014-01-03    6662107
3 2014-01-04  201035389
4 2014-01-05   35549270

You can call Chronos as follows:

>>> from chronos import Chronos
>>> import chronos_plotting
>>>
>>> my_chronos = Chronos(seasonality_mode="mul", distribution="Gamma")
>>> my_chronos.fit(divvy_data)
Employing Maximum A Posteriori
100.0% - ELBO loss: -1.5903 | Mean Absolute Error: 11152849920.0000

>>> predictions = my_chronos.predict(period=365)
Prediction no: 1000
>>> chronos_plotting.plot_components(predictions, my_chronos)

Notice we can specify the distribution of the ride shares to be a gamma distribution to ensure they are never negative. Additionally, we made the seasonality multiplicative to make sure that its affect increases as the absolute number of rides increases.