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://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.
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.
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.
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.