Geographical Aggregation (Prison Population)

In many applications, a set of time series is hierarchically organized. Examples include the presence of geographic levels, products, or categories that define different types of aggregations. In such scenarios, forecasters are often required to provide predictions for all disaggregate and aggregate series. A natural desire is for those predictions to be “coherent”, that is, for the bottom series to add up precisely to the forecasts of the aggregated series.

In this notebook we present an example on how to use HierarchicalForecast to produce coherent forecasts between geographical levels. We will use the Australian Prison Population dataset.

We will first load the dataset and produce base forecasts using an ETS model from StatsForecast, and then reconciliate the forecasts with several reconciliation algorithms from HierarchicalForecast. Finally, we show the performance is comparable with the results reported by the Forecasting: Principles and Practice which uses the R package fable.

You can run these experiments using CPU or GPU with Google Colab.

!pip install hierarchicalforecast statsforecast

1. Load and Process Data

The dataset only contains the time series at the lowest level, so we need to create the time series for all hierarchies.

import numpy as np
import pandas as pd

Y_df = pd.read_csv('https://OTexts.com/fpp3/extrafiles/prison_population.csv')
Y_df = Y_df.rename({'Count': 'y', 'Date': 'ds'}, axis=1)
Y_df.insert(0, 'Country', 'Australia')
Y_df = Y_df[['Country', 'State', 'Gender', 'Legal', 'Indigenous', 'ds', 'y']]
Y_df['ds'] = pd.to_datetime(Y_df['ds']) + pd.DateOffset(months=1)
Y_df.head()

	Country	State	Gender	Legal	Indigenous	ds	y
0	Australia	ACT	Female	Remanded	ATSI	2005-04-01	0
1	Australia	ACT	Female	Remanded	Non-ATSI	2005-04-01	2
2	Australia	ACT	Female	Sentenced	ATSI	2005-04-01	0
3	Australia	ACT	Female	Sentenced	Non-ATSI	2005-04-01	5
4	Australia	ACT	Male	Remanded	ATSI	2005-04-01	7

The dataset can be grouped in the following grouped structure.

hiers = [
    ['Country'],
    ['Country', 'State'], 
    ['Country', 'Gender'], 
    ['Country', 'Legal'], 
    ['Country', 'State', 'Gender', 'Legal']
]

Using the aggregate function from HierarchicalForecast we can get the full set of time series.

from hierarchicalforecast.utils import aggregate

Y_df, S_df, tags = aggregate(Y_df, hiers)
Y_df['y'] = Y_df['y']/1e3

Y_df.head()

	unique_id	ds	y
0	Australia	2005-04-01	24.296
1	Australia	2005-07-01	24.643
2	Australia	2005-10-01	24.511
3	Australia	2006-01-01	24.393
4	Australia	2006-04-01	24.524

S_df.iloc[:5, :5]

	unique_id	Australia/ACT/Female/Remanded	Australia/ACT/Female/Sentenced	Australia/ACT/Male/Remanded	Australia/ACT/Male/Sentenced
0	Australia	1.0	1.0	1.0	1.0
1	Australia/ACT	1.0	1.0	1.0	1.0
2	Australia/NSW	0.0	0.0	0.0	0.0
3	Australia/NT	0.0	0.0	0.0	0.0
4	Australia/QLD	0.0	0.0	0.0	0.0

tags

{'Country': array(['Australia'], dtype=object),
 'Country/State': array(['Australia/ACT', 'Australia/NSW', 'Australia/NT', 'Australia/QLD',
        'Australia/SA', 'Australia/TAS', 'Australia/VIC', 'Australia/WA'],
       dtype=object),
 'Country/Gender': array(['Australia/Female', 'Australia/Male'], dtype=object),
 'Country/Legal': array(['Australia/Remanded', 'Australia/Sentenced'], dtype=object),
 'Country/State/Gender/Legal': array(['Australia/ACT/Female/Remanded', 'Australia/ACT/Female/Sentenced',
        'Australia/ACT/Male/Remanded', 'Australia/ACT/Male/Sentenced',
        'Australia/NSW/Female/Remanded', 'Australia/NSW/Female/Sentenced',
        'Australia/NSW/Male/Remanded', 'Australia/NSW/Male/Sentenced',
        'Australia/NT/Female/Remanded', 'Australia/NT/Female/Sentenced',
        'Australia/NT/Male/Remanded', 'Australia/NT/Male/Sentenced',
        'Australia/QLD/Female/Remanded', 'Australia/QLD/Female/Sentenced',
        'Australia/QLD/Male/Remanded', 'Australia/QLD/Male/Sentenced',
        'Australia/SA/Female/Remanded', 'Australia/SA/Female/Sentenced',
        'Australia/SA/Male/Remanded', 'Australia/SA/Male/Sentenced',
        'Australia/TAS/Female/Remanded', 'Australia/TAS/Female/Sentenced',
        'Australia/TAS/Male/Remanded', 'Australia/TAS/Male/Sentenced',
        'Australia/VIC/Female/Remanded', 'Australia/VIC/Female/Sentenced',
        'Australia/VIC/Male/Remanded', 'Australia/VIC/Male/Sentenced',
        'Australia/WA/Female/Remanded', 'Australia/WA/Female/Sentenced',
        'Australia/WA/Male/Remanded', 'Australia/WA/Male/Sentenced'],
       dtype=object)}

Split Train/Test sets

We use the final two years (8 quarters) as test set.

Y_test_df = Y_df.groupby('unique_id', as_index=False).tail(8)
Y_train_df = Y_df.drop(Y_test_df.index)

2. Computing base forecasts

The following cell computes the base forecasts for each time series in Y_df using the ETS model. Observe that Y_hat_df contains the forecasts but they are not coherent.

from statsforecast.models import AutoETS
from statsforecast.core import StatsForecast

fcst = StatsForecast(models=[AutoETS(season_length=4, model='ZMZ')], 
                     freq='QS', n_jobs=-1)
Y_hat_df = fcst.forecast(df=Y_train_df, h=8, fitted=True)
Y_fitted_df = fcst.forecast_fitted_values()

Y_test_df

	unique_id	ds	y
40	Australia	2015-04-01	35.271
41	Australia	2015-07-01	35.921
42	Australia	2015-10-01	36.067
43	Australia	2016-01-01	36.983
44	Australia	2016-04-01	37.830
…	…	…	…
2155	Australia/WA/Male/Sentenced	2016-01-01	3.894
2156	Australia/WA/Male/Sentenced	2016-04-01	3.876
2157	Australia/WA/Male/Sentenced	2016-07-01	3.969
2158	Australia/WA/Male/Sentenced	2016-10-01	4.076
2159	Australia/WA/Male/Sentenced	2017-01-01	4.088

Y_train_df

	unique_id	ds	y
0	Australia	2005-04-01	24.296
1	Australia	2005-07-01	24.643
2	Australia	2005-10-01	24.511
3	Australia	2006-01-01	24.393
4	Australia	2006-04-01	24.524
…	…	…	…
2147	Australia/WA/Male/Sentenced	2014-01-01	3.614
2148	Australia/WA/Male/Sentenced	2014-04-01	3.635
2149	Australia/WA/Male/Sentenced	2014-07-01	3.692
2150	Australia/WA/Male/Sentenced	2014-10-01	3.726
2151	Australia/WA/Male/Sentenced	2015-01-01	3.780

3. Reconcile forecasts

The following cell makes the previous forecasts coherent using the HierarchicalReconciliation class. Since the hierarchy structure is not strict, we can’t use methods such as TopDown or MiddleOut. In this example we use BottomUp and MinTrace.

from hierarchicalforecast.methods import BottomUp, MinTrace
from hierarchicalforecast.core import HierarchicalReconciliation

reconcilers = [
    BottomUp(),
    MinTrace(method='mint_shrink')
]
hrec = HierarchicalReconciliation(reconcilers=reconcilers)
Y_rec_df = hrec.reconcile(Y_hat_df=Y_hat_df, Y_df=Y_fitted_df, S=S_df, tags=tags)

The dataframe Y_rec_df contains the reconciled forecasts.

Y_rec_df.head()

	unique_id	ds	AutoETS	AutoETS/BottomUp	AutoETS/MinTrace_method-mint_shrink
0	Australia	2015-04-01	34.799497	34.946476	34.923548
1	Australia	2015-07-01	35.192638	35.410342	35.432421
2	Australia	2015-10-01	35.188216	35.580849	35.473386
3	Australia	2016-01-01	35.888628	35.951878	35.939526
4	Australia	2016-04-01	36.045437	36.416829	36.245158

4. Evaluation

The HierarchicalForecast package includes the HierarchicalEvaluation class to evaluate the different hierarchies and also is capable of compute scaled metrics compared to a benchmark model.

from hierarchicalforecast.evaluation import evaluate
from utilsforecast.losses import mase
from functools import partial

eval_tags = {}
eval_tags['Total'] = tags['Country']
eval_tags['State'] = tags['Country/State']
eval_tags['Legal status'] = tags['Country/Legal']
eval_tags['Gender'] = tags['Country/Gender']
eval_tags['Bottom'] = tags['Country/State/Gender/Legal']

df = Y_rec_df.merge(Y_test_df, on=['unique_id', 'ds'])
evaluation = evaluate(df = df,
                      tags = eval_tags,
                      train_df = Y_train_df,
                      metrics = [partial(mase, seasonality=4)])

numeric_cols = evaluation.select_dtypes(include="number").columns
evaluation[numeric_cols] = evaluation[numeric_cols].map('{:.2f}'.format).astype(np.float64)
evaluation.rename(columns={'AutoETS': 'Base'}, inplace=True)

evaluation

	level	metric	Base	AutoETS/BottomUp	AutoETS/MinTrace_method-mint_shrink
0	Total	mase	1.36	1.07	1.17
1	State	mase	1.53	1.55	1.59
2	Legal status	mase	2.40	2.48	2.38
3	Gender	mase	1.08	0.82	0.93
4	Bottom	mase	2.16	2.16	2.14
5	Overall	mase	1.99	1.98	1.98

Fable Comparison

Observe that we can recover the results reported by the Forecasting: Principles and Practice book. The original results were calculated using the R package fable.

Getting Started

Tutorials

API Reference

Geographical Aggregation (Prison Population)

1. Load and Process Data

Split Train/Test sets

2. Computing base forecasts

3. Reconcile forecasts

4. Evaluation

Fable Comparison

References

Getting Started

Tutorials

API Reference

​1. Load and Process Data

​Split Train/Test sets

​2. Computing base forecasts

​3. Reconcile forecasts

​4. Evaluation

​Fable Comparison

​References

1. Load and Process Data

Split Train/Test sets

2. Computing base forecasts

3. Reconcile forecasts

4. Evaluation

Fable Comparison

References