Hierarchical Time Series Forecasting using Prophet

Last update: Dec 02, 2022

Related tags

Machine Learning htsprophet

Overview

htsprophet

Hierarchical Time Series Forecasting using Prophet

Credit to Rob J. Hyndman and research partners as much of the code was developed with the help of their work.

https://www.otexts.org/fpp

https://robjhyndman.com/publications/

Credit to Facebook and their fbprophet package.

https://facebookincubator.github.io/prophet/

It was my intention to make some of the code look similar to certain sections in the Prophet and (Hyndman's) hts packages.

Downloading

pip install htsprophet

If you'd like to just skip to coding with the package, runHTS.py should help you with that, but if you like reading, the following should help you understand how I built htsprophet and how it works.

Part I: The Data

I originally used Redfin traffic data to build this package.

I pulled the data so that date was in the first column, my layers were the middle columns, and the number I wanted to forecast was in the last column.

I made a function called makeWeekly() , that rolls up your data into the weekly level. It’s not a necessary function, it was mostly just convenient for me.

So the data looked like this:

Date	Platform	Medium	BusinessMarket	Sessions
1100 B.C.	Stone Tablet	Land	Birmingham	23234
...	Car Phone	Air	Auburn	2342
	...	Sea	Evanston	233
		...	Seattle	445
			...	46362

I then ran my orderHier() function with just this dataframe as its input.

NOTE: you cannot run this function if you have more than 4 columns in the middle (in between Date and Sessions for ex.)

To run this function, you specify the data, and how you want your middle columns to be ordered.

So orderHier(data, 2, 1, 3) means you want the second column after date to be the first level of the hierarchy.

Our example would look like this:

Date	Total	Land	Air	Sea	Land_Stone tablet	Land_Car Phone	Air_Stone tablet
1100 B.C.	24578	23135	555	888	23000	135	550
1099 B.C.	86753	86654	44	55	2342	84312	22
...	...	...	...	...	...	...	...

*All numbers represent the number of sessions for each node in the Hierarchy

If you have more than 4 categorical columns, then you must get the data in this format on your own while also producing the list of lists called nodes

Nodes – describes the structure of the hierarchy.

Here it would equal [[3],[2,2,2],[4,4,4,4,4,4]]

There are 3 nodes in the first level: Land, Air, Sea.

There are 2 children for each of those nodes: Stone tablet, Car phone.

There are 4 business markets for each of those nodes: Tokyo, Hamburg etc.

If you use the orderHier function, nodes will be the second output of the function.

Part II: Prophet Inputs

Anything that you would specify in Prophet you can specify in hts().

It’s flexible and will allow you to input a dataframe of values for inputs like cap, capF, and changepoints.

All of these inputs are specified when you call hts, and after that you just let it run.

The following is the description of inputs and outputs for hts as well as the specified defaults:

Parameters
----------------
 y - dataframe of time-series data
           Layout:
               0th Col - Time instances
               1st Col - Total of TS
               2nd Col - One of the children of the Total TS
               3rd Col - The other child of the Total TS
               ...
               ... Rest of the 1st layer
               ...
               Xth Col - First Child of the 2nd Col
               ...
               ... All of the 2nd Col's Children
               ...
               X+Yth Col - First Child of the 3rd Col
               ...
               ..
               .   And so on...

 h - number of step ahead forecasts to make (int)

 nodes - a list or list of lists of the number of child nodes at each level
 Ex. if the hierarchy is one total with two child nodes that comprise it, the nodes input would be [2]
 
 method – (String)  the type of hierarchical forecasting method that the user wants to use. 
            Options:
            "OLS" - optimal combination using ordinary least squares (Default), 
            "WLSS" - optimal combination using structurally weighted least squares, 
            "WLSV" - optimal combination using variance weighted least squares, 
            "FP" - forcasted proportions (top-down)
            "PHA" - proportions of historical averages (top-down)
            "AHP" - average historical proportions (top-down)
            "BU" - bottom-up (simple addition)
            "CVselect" - select which method is best for you based on 3-fold Cross validation (longer run time)
 
 freq - (Time Frequency) input for the forecasting function of Prophet 
 
 include_history - (Boolean) input for the forecasting function of Prophet
 
 transform - (None or "BoxCox") Do you want to transform your data before fitting the prophet function? If yes, type "BoxCox"
            
 cap - (Dataframe or Constant) carrying capacity of the input time series.  If it is a dataframe, then
                               the number of columns must equal len(y.columns) - 1
                               
 capF - (Dataframe or Constant) carrying capacity of the future time series.  If it is a dataframe, then
                                the number of columns must equal len(y.columns) - 1
 
 changepoints - (DataFrame or List) changepoints for the model to consider fitting. If it is a dataframe, then
                                    the number of columns must equal len(y.columns) - 1
 
 n_changepoints - (constant or list) changepoints for the model to consider fitting. If it is a list, then
                                     the number of items must equal len(y.columns) - 1
 skipFitting - (Boolean) if y is already a dictionary of dataframes, set this to True, and DO NOT run with method = "cvSelect" or transform = "BoxCox"
 
 numThreads - (int) number of threads you want to use when running cvSelect. Note: 14 has shown to decrease runtime by 10 percent 
 
 All other inputs - see Prophet
 
Returns
-----------------
 ynew - a dictionary of DataFrames with predictions, seasonalities and trends that can all be plotted

Don’t forget to specify the frequency if you’re not using daily data.

All other functions should be self-explanatory.

Part III: Room For Improvement

Prediction intervals

Hierarchical Time Series Forecasting using Prophet

Related tags

Overview

htsprophet

Downloading

Part I: The Data

Part II: Prophet Inputs

Part III: Room For Improvement

Owner

Collin Rooney

Apache Spark & Python (pySpark) tutorials for Big Data Analysis and Machine Learning as IPython / Jupyter notebooks

Karate Club: An API Oriented Open-source Python Framework for Unsupervised Learning on Graphs (CIKM 2020)

A Tools that help Data Scientists and ML engineers train and deploy ML models.

MosaicML Composer contains a library of methods, and ways to compose them together for more efficient ML training

Automatically build ARIMA, SARIMAX, VAR, FB Prophet and XGBoost Models on Time Series data sets with a Single Line of Code. Now updated with Dask to handle millions of rows.

Interactive Parallel Computing in Python

Python ML pipeline that showcases mltrace functionality.

Distributed Tensorflow, Keras and PyTorch on Apache Spark/Flink & Ray

An MLOps framework to package, deploy, monitor and manage thousands of production machine learning models

🌊 River is a Python library for online machine learning.

Production Grade Machine Learning Service

Reproducibility and Replicability of Web Measurement Studies

Home repository for the Regularized Greedy Forest (RGF) library. It includes original implementation from the paper and multithreaded one written in C++, along with various language-specific wrappers.

Markov bot - A Writing bot based on Markov Chain for Data Structure Lab

A Python Automated Machine Learning tool that optimizes machine learning pipelines using genetic programming.

LightGBM + Optuna: no brainer

JMP is a Mixed Precision library for JAX.

ArviZ is a Python package for exploratory analysis of Bayesian models

jaxfg - Factor graph-based nonlinear optimization library for JAX.

Databricks Certified Associate Spark Developer preparation toolkit to setup single node Standalone Spark Cluster along with material in the form of Jupyter Notebooks.