A machine learning approach for forecasting hierarchical

time series
Paolo Mancuso, Veronica Piccialli, Antonio M. Sudoso
University of Rome Tor Vergata
arXiv:2006.00630v1 [cs.LG] 31 May 2020

Abstract
In this paper, we propose a machine learning approach for forecasting hierar-
chical time series. Rather than using historical or forecasted proportions, as
in standard top-down approaches, we formulate the disaggregation problem
as a non-linear regression problem. We propose a deep neural network that
automatically learns how to distribute the top-level forecasts to the bottom
level-series of the hierarchy, keeping into account the characteristics of the
aggregate series and the information of the individual series. In order to
evaluate the performance of the proposed method, we analyze hierarchical
sales data and electricity demand data. Besides comparison with the top-
down approaches, the model is compared with the bottom-up method and the
optimal reconciliation method. Results demonstrate that our method does
not only increase the average forecasting accuracy of the hierarchy but also
addresses the need of building an automated procedure generating coherent
forecasts for many time series at the same time.
Keywords: Hierarchical Time Series, Forecast, Machine Learning, Neural
Network

1. Introduction
A hierarchical time series is a collection of time series organized in a
hierarchical structure that can be aggregated at different levels [7]. As an
example, Stock Keeping Unit (SKU) sales aggregate up to product subcate-
gory sales, which further aggregate to product categories. In order to support
decision-making at different levels of the hierarchy, a challenging task is the
generation of coherent forecasts. Forecasts of the individual series must sum
up in a proper way across the levels preserving the hierarchical structure.

Preprint submitted to June 2, 2020

For example, forecasts of regional sales should sum up to give forecasts of
state sales, which should in turn sum up to give a forecast for the national
sales.
In the literature two lines of research among others are pursued: bottom-
up and top-down approaches. Top-down approaches involve forecasting first
the top-level series and then disaggregating by means of historical [5] or fore-
casted proportion [2] to get forecasts for the lower level series. On the other
hand, the bottom-up approach produces first forecasts for the bottom-level
time series and then aggregates them in order to get the forecasts for the
higher level time series. Both classes of methods have their own advantages,
since top-down approaches perform well when the top-level series is easy to
forecast, whereas the bottom-up method presents the advantage that the
pattern of each series is accurately identified and consequently forecasted
without loss of information. However, the bottom-up approach ignores cor-
relations among the series, and this may lead to poor aggregate forecasts
with respect to top-down approaches [17]. In general, a bottom-up approach
should be preferred whenever the forecasts are employed to support decisions
that are mainly related to the bottom rather than the top of the hierarchy,
whereas a top-down approach performs better when the bottom-level series
are too noisy [4]. The objective to reconcile forecasts at all levels of the hi-
erarchy, from the top to the bottom, has lead researchers to investigate the
impact that the association between bottom-level series produce on the ag-
gregation [15]. Analytical approaches to the forecast reconciliation problem
have been proposed by Hyndman et al. (2011, 2014) [9, 6] and by Wickra-
masuriya et al. (2019) [19]. These methods not only ensure that forecasts
are coherent but also lead to improvements in forecast accuracy. However,
a shortcoming of these methods is that they involve two stages, with fore-
casts first produced independently for each series in the hierarchy, and then
optimally combined to satisfy the aggregation constraint.
In this paper, we propose a new top-down approach for forecasting hierar-
chical time series. We formulate the disaggregation problem as a non-linear
regression problem and we choose to solve it with a deep neural network that
jointly learns how to disaggregate and generate coherent forecasts across the
levels of the hierarchy. Our approach is successful at capturing the relations
between the series to disaggregate thanks to the neural network’s ability to
extract meaningful features from the aggregate series and combine them with
the dynamics of the individual series. We test our method on two real-world
datasets with completely different characteristics: the first one comes from

2
sales data and has noisy and intermittent bottom-level series, the second one
comes from electricity demand data and has more regular bottom-level series.
Our numerical experiments show that in both cases our method increases the
average forecasting accuracy of the hierarchy outperforming state-of-the-art
approaches. The rest of the paper is organized as follows. Section 2 dis-
cusses the concept of hierarchical time series and the methods of hierarchical
forecasting. Section 3 contains the detail of the proposed machine learning
algorithm. Section 4 describes the basic forecasting methods employed in
the hierarchical models and the experimental setup. Section 5 discusses the
datasets and the numerical experiments conducted to evaluate the proposed
method. Finally, Section 6 concludes the paper.

2. Hierarchical Time Series

In a general hierarchical structure with K > 0 levels, level 0 is defined
as the completely aggregated series. Each level from 1 to K − 2 denotes a
further disaggregation down to level K −1 containing the most disaggregated
time series. In a hierarchical time series, the observations at higher levels can
be obtained by summing up the series below. Let ytk ∈ <mk be the vector
of all observations at level k = 1, . . . , K −P
1 and t = 1, . . . , T , where mk is
the number of series at level k and M = K−1 k=0 mk is the total number of
series in the hierarchy. Then we define the vector of all observations of the
hierarchy:  0 
yt
 y1 
 t 
yt =  ..  ,
 . 
ytK−1
where yt0 is the observation of the series at the top and the vector ytK−1
contains the observations of the series at the bottom of the hierarchy. The
structure of the hierarchy is determined by the summing matrix S that de-
fines the aggregation constraints:

yt = SytK−1 .

The summing matrix S is a matrix having entries belonging to {0, 1}, of size
M × mK−1 .
Given observations at time t = 1, ..., T and the forecasting horizon h,
the aim is to forecast each series at each level at time t = T + 1, ..., T + h.

3
The current methods of forecasting hierarchical time series are: top-down,
bottom-up, middle-out and optimal reconciliation [9]. The main objective
of such approaches is to ensure that forecasts are coherent across all the
levels of the hierarchy. Regardless of the methods used to forecast the time
series for the different levels of the hierarchy, the individual forecasts must
be reconciled to be useful for any subsequent decision making. Forecast
reconciliation is the process of adjusting forecasts to make them coherent.
By definition, a forecast is coherent if it satisfies the aggregation constraints
defined by the summing matrix.

2.1. Bottom-up Approach

The bottom-up approach focuses on producing the h-step-ahead base fore-
casts for each series at the lowest level ŷhK−1 and aggregating them to the
upper levels of the hierarchy according to the summing matrix. It can be
represented as follows:
ỹh = S ŷhK−1 ,
where ỹh is the vector of coherent h-step-ahead forecasts for all series of the
hierarchy. An advantage of this approach is that we directly forecast the
series at the bottom-level and no information is lost due to aggregation. On
the other hand, bottom-level series can be quite noisy and more challenging
to model and forecast. This approach also has the disadvantage of having
many time series to forecast if there are many series at the lowest level.

2.2. Top-down Approaches

Top-down approaches first involve generating the base forecasts for the
total series and then disaggregating these downwards to get coherent fore-
casts for each series of the hierarchy. The disaggregation of the top-level
forecasts is usually achieved by using the proportions p = (p1 , ..., pmK−1 )T ,
which represent the relative contribution of the bottom-level series to the
top-level aggregate. The two most commonly used top-down approaches are
the Average Historical Proportions (AHP) and the Proportions of the Histor-
ical Averages (PHA). In the case of the AHP, the proportions are determined
as follows:
T K−1
1 X yt,i
pi = , i = 1, . . . , mK−1 .
T t=1 yt0

4
In the PHA approach, the proportions are determined in the following man-
ner: K−1
PT yt,i
t=1 T
pi = PT y0 , i = 1, . . . , mK−1 .
t
t=1 T
For these two methods, once the bottom-level h-step-ahead forecasts have
been generated, these are aggregated to generate coherent forecasts for the
rest of the series of the hierarchy by using the summing matrix. Given the
vector of proportions p, top-down approaches can be represented as:

ỹh = Spŷh0 .

Top-down approaches based on historical proportions usually produce less

accurate forecasts at lower levels of the hierarchy than bottom-up approaches
because they don’t take into account that these proportions may change over
time. To address this issue, instead of using the static proportions as in
AHP and PHA, Athanasopoulos et al. (2009) propose in [2] the Forecasted
Proportion (FP) method in which proportions are based on forecasts rather
than on historical data. It first generates an independent base forecast for
all series in the hierarchy, then for each level, from the top to the bottom,
the proportion of each base forecast to the aggregate of all the base forecasts
at that level are calculated. For a hierarchy with K levels we have:
K−2 k
Y ŷt,i
pi = k+1
, i = 1, ..., mK−1 .
k=0
σ̂t,i
k
where is the base forecast of the series that corresponds to the node
ŷt,i
k+1
which is k levels above node i, and σ̂t,i is the sum of the base forecasts
below the series that is k levels above node i and directly in contact with
that series.

2.3. Middle-out Approach

The middle-out method can be seen as a combination of the top-down
and bottom-up approaches. It combines ideas from both methods by starting
from a middle level where forecasts are reliable. For the series above the
middle level, coherent forecasts are generated using the bottom-up approach
by aggregating these forecasts upwards. For the series below the middle level,
coherent forecasts are generated using a top-down approach by disaggregating
the middle level forecasts downwards.

5
2.4. Optimal Reconciliation
In [9] Hyndman et al. (2011) propose a novel approach that provides op-
timal forecasts that are better than forecasts produced by either a top-down
or a bottom-up approach. Their proposal is based on independently forecast-
ing all series at all levels of the hierarchy and then using a linear regression
model to optimally combine and reconcile these forecasts. Their approach is
based on a generalized least squares estimator that requires an estimate of the
covariance matrix of the errors that arise due to incoherence. In [19] Wick-
ramasuriya et al. (2019) show that this matrix is impossible to estimate in
practice and they propose a state-of-the-art forecast reconciliation approach,
called Minimum Trace (MinT) that incorporates the information from a full
covariance matrix of forecast errors in obtaining a set of coherent forecasts.
MinT minimizes the mean squared error of the coherent forecasts across the
entire hierarchy with the constraint of unbiasedness. The resulting revised
forecasts are coherent, unbiased and have minimum variance amongst all
combination forecasts. An advantage of the optimal reconciliation approach
is that allows for the correlations between the series at each level using all
the available information within the hierarchy. However, it is computation-
ally expensive compared to the other methods introduced so far because it
requires to individually forecast all the time series at all the levels of the
hierarchy.

3. Neural Network Disaggregation

As observed in [7], standard top-down approaches have the disadvantage
of information loss since they are unable to capture the individual time series
characteristics. Departing from the related literature, to the best of our
knowledge, we propose a new top-down approach which first generates a
good forecast for the aggregated time series at some level of the hierarchy
and then disaggregates it downwards, without loss of information, by means
of a machine learning algorithm. In order to explain the proposed algorithm,
we focus on two consecutive levels with the top-level time series being at
node j of level k and the bottom-level series at level k + 1 (see Figure 1).
Let mk+1
j be the number of series at level k + 1 connected to the parent
node j at level k, then we model the disaggregation procedure as a non-linear
regression problem:
k,p k,p k,p
ytk+1,j = f (yt,j , yt−1,j , . . . , yt−l,j , xt,1 , . . . , xt,mk+1 ) + ,
j

6
 ...

... k,p ...

yt,j

k+1,j k+1,j
yt,1 ... yt,m k+1
j

... ... ... ... ... ... ... ... ...

Figure 1: A top-level series at level k and the bottom-level series at level k + 1.

where ytk+1,j is the vector of size mk+1

j
k,p
containing the series at level k +1, yt,j
is the aggregate time series corresponding to the node j at level k connected
to the parent node p at level k − 1, l is the number of the lagged time steps
of the aggregated series, xt,i is a vector of the external regressors for each
series at level k + 1, f is a non-linear function learned, in our case, by a
feed-forward neural network and  is the error term.
Given any aggregate time series yt,jk,p
and the vector of series ytk+1,j , the
algorithm is made up of two steps: in the first one the best forecasting model
for the aggregated time series is chosen, and the neural network is trained
with the real values of the training set of the two levels time series; in the
second step, forecasts for the aggregated time series are fed to the neural
network in order to obtain forecasts for all the lower level time series. The
flow chart of the proposed algorithm is shown in Figure 2. More in detail,
the two steps are the following:
Step 1 In the training phase, the best forecasting model F ∗ for the time series
k,p
yt,j is chosen on the basis of the training set. At the same time, the neu-
k,p
ral network is trained taking as input the training set of yt,j with lagged
time steps and the explanatory variables xt,i relative to the training
set of ytk+1,j . The output are the true values of the disaggregated time
series ytk+1,j . In order to simplify the notation, from now on we refer
to the produced model as NND (Neural Network Disaggregation).

7
 k,p
Step 2 In the disaggregation or test phase, forecasts ŷt,j relative to the time
∗
period of the test set are generated by the model F . Finally, these fore-
casts are fed to the trained NND in order to produce the disaggregated
forecasts ŷtk+1,j for the test set.
In general, the learned function f generates base forecasts that are not
coherent since they do not sum up correctly according to the structure of the
hierarchy. In order to ensure that forecasts are reconciled across the hierar-
chy, we want f to output a set of forecasts that are as close as possible to the
base forecasts, but also meet the requirement that forecasts at upper levels
in the hierarchy are the sum of the associated lower level forecasts. From
an optimization perspective, we want to introduce an equality constraint to
the regression problem in such a way that we can still use backpropagation
to train the network. More in detail, we are looking for the network weights
such that the Mean Squared Error (MSE) between the true values and the
predictions is minimized and, in addition, we want the following constraint
to hold:
k,p k,p
yt,j = 1T ytk+1,j = 1T ŷtk+1,j = ŷt,j ,
where 1 is the vector of all ones of size mk+1
j .
We impose the aggregation constraint by adding a term to the MSE loss
function that penalizes differences between the sum of lower level observa-
tions and the sum of the lower level forecasts:

 T
1 X
L(ytk+1,j , ŷtk+1,j ) = (1 − α) ||ytk+1,j − ŷtk+1,j ||2 +
T t=1
T  (1)
X
T k+1,j T k+1,j 2
+α (1 yt − 1 ŷt ) ,
t=1

where α ∈ (0, 1) is a parameter that controls the relative contribution of

each term in the loss function. Note that the two terms are on the same
scale, and parameter α measures the compromise between minimizing the
MSE and satisfying the aggregation constraint. A too small value of α will
result in the corresponding constraint being ignored, producing, in general,
not coherent forecasts whereas a too large value will cause the MSE being
ignored, producing coherent but possibly inaccurate base forecasts. The

8
 k,p
yt,j ytk+1,j , xt,i
Training Set Training Set

Model Selection NND Training

Step 1

F ∗ Model

k,p
ŷt,j Test Set NND Model

Step 2

ŷtk+1,j Test Set

Figure 2: Decomposition of the aggregated forecast through a neural network: Neural

Network Disaggregation (NND).

9
idea is to balance the contribution of both terms by setting α = 0.5, that
corresponds to giving the two terms the same importance. In principle, the
parameter α may be tuned on each instance. However, if the forecasts were
equal to the true values, the coherence constraint would be satisfied, since
the true values are coherent by construction. This explains why even with
α = 0, the violation of coherence is in practice relatively small and with
α = 0.5 the violation is basically zero in all our experiments. For this reason,
we did not investigate the tuning of parameter α and kept it fixed to 0.5.
Top-down approaches distribute the top-level forecasts down the hier-
archy using historical or forecasted proportions of the data. In our case,
explicit proportions are never calculated since the algorithm automatically
learns how to disaggregate forecasts of the top-level series to the bottom-
level series without loss of information. Furthermore, our method is flexible
enough to be employed in the forecasting process of the whole hierarchy in
two different ways:
1. Standard top-down: a forecasting model F ∗ is developed for the aggre-
gate at level 0, and a single disaggregation model NDD is trained with
the series at level 0 and K −1. Therefore, forecasts for the bottom-level
series are produced by looking only at the aggregated series at level 0.
Then, the bottom-level forecasts are aggregated to generate coherent
forecasts for the rest of the series of the hierarchy.
2. Iterative top-down: the forecasting model F ∗ for an aggregate at level
k is the disaggregation model NDD trained with the series at level
k − 1 and k, for each k = 1, . . . , K − 1. At level 0, instead, F ∗ is
the best model selected among a set of standard forecasting methods.
Forecast for all the levels are then obtained by feeding forecasts to the
disaggregation models at each level.
The difference between the two approaches is that in the standard top-down,
bottom-level forecasts are generated with only one disaggregation model,
whereas in the iterative version, a larger number of disaggregation models
is trained, one for each series to be disaggregated. To be more precise, to
disaggregate the mk series at level k = 0, . . . , K−2, exactly mk disaggregation
models are trained in parallel. In this way, on the one hand we increase the
variance of the approach (and the computational time), but on the other
hand we reduce the bias since we increase flexibility, and keep into account
more the variability at the different levels.
We also notice that this algorithm can be easily plugged into a middle-out

10
strategy: a forecasting model is developed for each aggregate at a convenient
level, and the disaggregation models are trained and tested to distribute these
forecasts to the series below. For the series above the middle level, coherent
forecasts are generated using the bottom-up approach.
Regarding the choice of the neural network architecture, our objective
is to include in the model the relationship between explanatory variables
derived from the lower level series, and the features of the aggregate series
that describe the structure of the hierarchy. In order to better capture the
structure of the hierarchy, we use a Convolutional Neural Network (CNN).
Indeed CNNs are well known for creating useful representations of time series
automatically, being highly noise-resistant models, and being able to extract
very informative, deep features, which are independent from time [13]. Our
model is a deep neural network that is capable of accepting and combining
multiple types of input, including cross-sectional and time series data, in a
single end-to-end model. Our architecture is made up of two branches: the
first branch is a simple Multi-Layer Perceptron (MLP) designed to handle
the explanatory variables xt,i such as price, promotions, day of the week,
or in general, special events affecting the time series of interest; the second
branch is a one-dimensional Convolutional Neural Network that extracts fea-
k,p
ture maps over fixed segments of the aggregate series yt,j . Features extracted
from the two subnetworks are then concatenated together to form the final
input of the multi-output regression model (see Figure 3). The output layer
of the model is a standard regression layer with linear activation function
where the number of units is equal to the number of the series to forecast.

11
 Explanatory Time Series
Variables Data

Multi-Layer Convolutional
Perceptron Neural Network
(MLP) (CNN)

Branch 1 Branch 2

Concatenate

Fully-Connected
(Linear Activation)

Figure 3: Our model has one branch that accepts the numerical data (left) and another
branch that accepts time series data (right).

12
4. Experimental Setup
In this section, we resume first the forecasting models we use to generate
base forecasts for the hierarchical approaches, then we describe our strategy
to select the best forecasting model and the implementation details.

4.1. Forecasting Models

In order to describe the methods, let (y1 , . . . , yT ) be an univariate time
series of length T and (yT +1 , . . . , yT +h ) the forecasting period, where h is the
forecast horizon. We consider the following models:

1. Naive
2. Autoregressive Integrated Moving Average (ARIMA)
3. Exponential Smoothing (ETS)
4. Non-linear autoregression model (NAR)
5. Dynamic regression models: univariate time series models, such as lin-
ear and non-linear autoregressive models, allow for the inclusion of
information from past observations of a series, but not for the inclusion
of other information that may also affect the time series of interest.
Dynamic regression models allow keeping into account the time-lagged
relationship between the output and the lagged observations of both
the time series itself and of the external regressors. More in detail, we
consider two types of dynamic regression models:
(a) ARIMA model with exogeneous variables (ARIMAX)
(b) NAR model with exogenous variables (NARX)

In the literature it has been pointed out that the performance of forecasting
models could be improved by suitably combining forecasts from standard
approaches [18]. An easy way to improve forecast accuracy is to use sev-
eral different models on the same time series, and to average the resulting
forecasts. We consider two ways of combining forecasts:

1. Simple Average: the most natural approach to combine forecasts is

to use the mean.PThe composite forecast in case of simple average is
given by ŷt = m1 mi=1 ŷi,t for t = T + 1, ..., T + h where h is the forecast
horizon, m is the number of combined models and ŷi,t is the forecast
at time t generated by model i.

13
 2. Constrained Least Squares Regression: in this setting, the composed
forecast is not a function of m only as in the simple average but is a
linear function of the individual forecasts whereby the parameters are
determined by solving an optimization problem. The approach pro-
posed by Timmermann in [18] minimizes the sum of squared errors
under some additional constraints. Specifically, the estimated coeffi-
cients βi are constrained to be non-negative and to sum up to one. The
weights obtained are easily interpretable as percentages devoted to each
of the individual forecasts. PGiven the optimal weights, the composed
forecast is obtained as ŷt = mi=1 βi ŷi,t for t = T +1, ..., T +h. From the
mathematical point of view the following optimization problem needs
to be solved:

T
X +h m
X
min (yt − βi ŷi,t )2
t=T +1 i=1

s.t. βi ≥ 0 i = 1, ..., m (2)

Xm
βi = 1
i=1

Differently from the simple average which does not need any training
as the weights are a function of m only, with this method we need to
allocate a reserved portion of forecasts in order to train the meta-model.
In particular, we consider two following composite models:
1. Combination of ARIMAX, NARX and ETS forecasts obtained through
the simple mean.
2. Combination of ARIMAX, NARX and ETS forecasts obtained by solv-
ing the constrained least squares problem.
We choose to combine the two dynamic regression models with the ex-
ponential smoothing in order to take directly into account the effect of the
explanatory variables and the presence of linear and non-linear patterns in
the series.

4.2. Model Selection

Following an approach widely employed in the machine learning litera-
ture, we separate the available data into two sets, training (in-sample) and

14
test (out-of-sample) data. The training data (y1 , . . . , yN ), a time series of
length N , is used to estimate the parameters of a forecasting model and the
test data (yN +1 , . . . , yT ), that comes chronologically after the training set, is
used to evaluate its accuracy.
To achieve a reliable measure of model performance, we implement on the
training set a procedure that applies a cross-validation logic suitable for time
series data. In the expanding window procedure described in [7], the model
is trained on a window that expands over the entire history of the time series
and it is repeatedly tested against a forecasting window without dropping
older data points. This method produces many different train/test splits
and the error on each split is averaged in order to compute a robust estimate
of the model error (see Figure 4). The implementation of the expanding
window procedure requires four parameters:

- Starting window: the number of data points included in the first train-
ing iteration.

- Ending window: the number of data points included in the last training
iteration.

- Forecasting window: number of data points included for forecasting.

- Expanding steps: the number of data points added to the training time
series from one iteration to another.

For each series, the best performing model after the cross-validation phase
is retrained using the in-sample data and forecasts are obtained recursively
over the out-of-sample period. The above procedure requires a forecast error
measure. We consider the Mean Absolute Scaled Error (MASE) proposed by
Hyndman and Koehler in [11]:
1
PT +h
h i=T +1 |yi − ŷi |
M ASE = 1
PT ,
T −m t=m+1 |yt − yt−m |

where the numerator is out-of-sample Mean Absolute Error (MAE) of the

method evaluated across the forecast horizon h, and the denominator is the
in-sample one-step ahead Naive forecast with seasonal period m.

15
 Figure 4: Expanding window procedure.

We also consider the Symmetric Mean Absolute Percentage Error (SMAPE)

defined as follows:
T +h
2 X |yi − ŷi |
SM AP E = .
h i=T +1 |yi | + |ŷi |
The SMAPE is easy to interpret, it has an upper bound of 2 when either
actual or predicted are zero or when actual and predicted are opposite signs.
However, the significant disadvantage of SMAPE is that it produces infinite
or undefined values where the actual values are zero or close to zero. The
MASE and SMAPE can be used to compare forecast methods on a single
series and, because they are scale-free, to compare forecast accuracy across
series. For this reason, we average the MASE and SMAPE values of several
series to obtain a measurement of forecast accuracy for the group of series.

4.3. Implementation
Time series models described above are implemented by using the “fore-
cast” package in R [10]. Hierarchical time series forecasting is performed by
using the “hts” package in R [8]. For the optimal combination approach we
use the MinT(Shrink) algorithm that estimates the covariance matrix of the
base forecast errors using shrinkage. The proposed disaggregation method
is implemented in Python with TensorFlow, a large-scale machine learning
framework [1]. Regarding the training details of the NND, early stopping is
employed as a form of regularization to avoid overfitting since it stops the
training as soon as the error on the validation set starts to grow [3]. The

16
neural network is trained by minimizing the loss function (1) with Adam
optimizer [14], a mini-batch stochastic gradient descent algorithm. Grid
search is used to perform the hyperparameter optimization which is simply
an exhaustive search through a manually specified subset of points in the hy-
perparameter space of the neural network. Each configuration is evaluated
on the validation set and the optimal values are chosen. More in detail, the
hyperparameters we optimize are:

• the size of the mini-batch in {32, 64, 128},

• the number of units of the MLP subnetwork in {64, 128, 256},

• the number of filters in {16, 32, 64} for the CNN subnetwork,

• the kernel size in {7, 14} for the CNN subnetwork.

As for the number of layers, the MLP subnetwork has 3 fully connected layers
whereas the CNN subnetwork has 6 convolutional layers. The training time
of a single disaggregation model requires order of minutes on a commercial
GPU depending on the network dimension and on the granularity of the
dataset.

5. Numerical Experiments
In this section, we aim to evaluate the effectiveness of our approach, by
comparing it with all the hierarchical methods described in Section 2. In
order to be as fair as possible in the comparison, we perform model selection
among the set of forecasting methods described in Section 4 whenever a base
forecast is required. More in detail, this means that different methods may be
used for each time series of the hierarchy we are trying to forecast (bottom-
level series for the bottom-up approach, aggregate series for all the top-down,
all the time series for the optimal reconciliation approach). As for the metrics
used for comparison, we use both the MASE and the SMAPE where possible
(i.e. where no zeros are present). We consider two datasets, coming from two
completely different problems. In both cases, starting from the aggregated
series at some level, we aim to exploit our method to increase the forecasting
accuracy of the hierarchy using the characteristics of the aggregate series and
explanatory variables.

17
5.1. Datasets
1. Sales Data: we analyze sales data gathered from an Italian super-
market. The dataset consists of 118 daily time series, representing the
demand of pasta from 01/01/2014 to 31/12/2018. Besides univariate
time series data, the quantity sold is integrated by information on the
presence or the absence of a promotion (no detail on the type of promo-
tion on the final price is given). Demand time series can be naturally
arranged to follow a hierarchical structure. Here, the idea is to build a
3-level structure: at the top of the hierarchy, there is the total or the
store-level series obtained by aggregating the brand-level series. At the
second level there are the brand-level series (like for instance Barilla)
obtained by aggregating the individual demand at the item-level and at
the third level there are the most disaggregated time series represent-
ing the item-level demand (for example the demand of spaghetti Bar-
illa). The completely aggregated series at level 0 is disaggregated into
4 component series at level 1. Each of these series is further subdivided
into 42, 45, 10 and 21 series at level 2, the completely disaggregated
bottom-level representing the different varieties of pasta for each brand
(see Table 1).
2. Electricity Data: we analyze a public electricity demand dataset that
contains power measurements and meteorological forecasts relative to
a set of 24 power meters installed in low-voltage cabinets of the dis-
tribution network of the city of Rolle in Switzerland [16]. The dataset
contains measurements from 13/01/2018 to 19/01/2019 at the resolu-
tion of 10 minutes and includes mean active and reactive power, voltage
magnitude, maximum total harmonic distortion for each phase, voltage
frequency and the average power over the three phases. We assume that
the grid losses are not significant, so the power at the grid connection
is the algebraic sum of the connected nodes. Based on the historical
measurements, the operator can determine coherent forecasts for all
the grid by generating forecasts for the nodal injections individually.
We build a 2-level hierarchy in which we aggregate the 24 series of the
distribution system at the meter-level to generate the total series at
the grid-level (see Table 2).

18
 Level Number of series Total series per level
Store 1 1
Brand 4 4
Item 42, 45, 10, 21 118
Table 1: Hierarchy for the sales data.

Level Number of series Total series per level

Grid 1 1
Meter 24 24
Table 2: Hierarchy for the electricity data.

Summarizing, we have the first dataset with a three level hierarchy and
the second one with a two level hierarchy. As for the experimental setup, we
have to make some choices for each dataset:
1. Sales Data: for each series, as explanatory variables, we add a binary
variable representing the presence of promotion if the disaggregation
is computed at the item-level or a variable representing the relative
number of items in promotion for each brand if the disaggregation is
computed at the brand-level. In both cases, dummy variables repre-
senting the day of the week and the month are also added to the model.
As for the number of lagged observations of the aggregate demand, we
consider fixed-length time windows of 30 days with a hop size of 1 day.
We consider 4 years from 01/01/2014 to 31/12/2017 for the in-sample
period and the last year of data from 01/01/2018 to 31/12/2018 for the
out-of-sample period. The experimental setup for the cross-validation
procedure is as follows. The starting windows consists of the first three
years of data from 01/01/2014 to 31/12/2016. The training window
expands over the last year of the training data including daily obser-
vations from 01/01/2017 to 31/12/2018. The forecasting window is set
to h = 7, corresponding to a forecasting horizon of one week ahead. At
each iteration, the training window expands by one week to simulate
a production environment in which the model is re-estimated as soon
as new data are available and to better mimic the practical scenario
in which retailing decisions are made every week. In order to evaluate
the forecasting accuracy at each level, for this hierarchy we use the av-
erage MASE, as recommended by Hyndman in [12], since most of the

19
 item-level series are intermittent.
2. Electricity Data: for each series, we use the average power over the
three phases as target variable and the temperature, horizontal irra-
diance, normal irradiance, relative humidity, pressure, wind speed and
wind direction as explanatory variables. Dummy variables representing
the day of the week and the hour of the day are also added to the model.
As for the number of lagged observations of the aggregate power, we
consider fixed-length time windows of 144 observations with hop size of
10 minutes. We consider 6 months from 13/01/2018 to 13/06/2018 for
the training set and the last 6 months from 14/06/2018 to 13/01/2019
for the test set. The configuration of the cross-validation procedure is
as follows. The starting window consists of the first three months of
data from 13/01/2018 to 13/03/2018. At each iteration, the training
window expands by 24 hours over the last 3 months of the training data
including observations from 14/03/2018 to 13/06/2018. The forecast-
ing window is set to h = 144, corresponding to a forecasting horizon of
24 hours ahead. We evaluate the forecasting accuracy at each level by
using the average MASE and the average SMAPE over all the series of
that level, since there are no zero values in these time series.

5.2. Results
In Table 3 we compare the forecasting performance of our method at each
level, in both its versions, standard top-down (NND1) and iterative top-down
(NND2) with the bottom-up, average historical proportions, proportions of
historical averages, forecasted proportions and the optimal reconciliation ap-
proach (MinT). We stress that for all the top down approaches, the perfor-
mance at the most aggregated level are equivalent, and the differences only
emerge at the lower levels of the hierarchy, where we really are interested in
the comparison.
For the NND1, we directly forecast the demand at the item-level using
the aggregate demand at the store-level and then we aggregate the item-
level forecasts to obtain the brand-level forecasts. For the NND2, we train a
disaggregation model that generates the brand-level forecasts starting from
the store-level series and then one NND for each brand-level series to generate
forecasts for each item demand of the brand they belong to. In total, for the
entire hierarchy, we train one NND at the top-level and 4 NND in parallel at
the brand-level.

20
 Average MASE
Level Bottom-up AHP PHA FP NND1 NND2 Optimal
Store 1.103 0.567 0.567 0.567 0.567 0.567 0.559
Brand 1.237 1.413 1.481 1.157 0.862 0.838 1.137
Item 1.057 0.934 0.943 0.891 0.745 0.811 0.893
Table 3: Average MASE for each aggregation level of sales data. In bold the best per-
forming approach.

The bad performance of the bottom-up method can be attributed to the

fact that the demand at the most granular level of the hierarchy is often
challenging to model and forecast effectively because it is too sparse and
erratic. The majority of the item-level time series display sporadic sales
including zeros and the promotion of an item does not always correspond
to an increase in sales. By using traditional or combination of methods to
generate base forecasts for the time series at the lowest level, we end up with
flat line forecasts, representing the average demand, failing to account for
the seasonality that truly exists but is impossible to identify between the
noise. By focusing our attention at the highest or some intermediate level
of the hierarchy, we have enough data to build decent models capturing the
underlying trend and seasonality. In fact, the aggregation tends to regularize
the demand and make it easier to forecast. The only level for which the
optimal reconciliation approach is the best is the top level. As we move
down the hierarchy our approach significantly outperforms all the top-down
approaches, the bottom-up method and the optimal reconciliation with the
NND iterative top-down (NND2) performing best at the brand-level and the
NND standard top-down (NND1) performing best at the item-level. This
result is reasonable since the iterative top-down involves training and testing
multiple disaggregation models and forecasts generated by one model are fed
to the model below. As a side effect, when we go down, the forecasting error
tends to propagate achieving less accurate performance at the bottom of the
hierarchy. In Figure 5 and 6 we show the forecasts generated by the NND
for the all the brand-level series. We perform pairwise t-tests to formally test
whether forecasts produced by the hierarchical methods are different. The
average historical proportions and the proportions of historical averages and
are not significantly different (p-value = 0.765). The remaining methods are
significantly different from each other (p-value < 10−3 for each test).
In Table 4 and 5 we present the MASE and SMAPE at each level of

21
 Average MASE
Level Bottom-up AHP PHA FP NND Optimal
Grid 1.095 1.007 1.007 1.007 1.007 1.087
Meter 1.089 1.371 1.365 1.143 0.956 1.112
Table 4: Average MASE for each aggregation level of the electricity dataset. In bold the
best performing approach.

Average SMAPE (%)

Level Bottom-up AHP PHA FP NND Optimal
Grid 7.347 7.345 7.345 7.345 7.345 7.349
Meter 16.267 19.245 19.219 16.634 12.801 16.323
Table 5: Average SMAPE for each aggregation level of the electricity dataset. In bold the
best performing approach.

the electricity dataset and for each method. Note that here we only have
two levels, so that NND1 and NND2 coincide (which is why we call it only
NND in the tables). We find that all the top-down approaches perform
best at the grid-level. The good performance of the bottom-up method with
respect to the classical top-down approaches can be attributed to the fact
that the series have a strong seasonality, even at the bottom level. Our
NND clearly outperforms all the top-down methods, the bottom-up and the
optimal combination at the meter-level. Note that these conclusions hold
also looking at different metrics like MSE. In Figure 7 we show the forecasts
generated by the NND for some of the meter-level series. In order to identify
the pairs of forecasts which are significantly different from each other, we
perform pairwise t-test. Our method is significantly different from each of
the other methods (p-value < 10−3 for each test). The optimal combination
and the bottom-up methods are not significantly different (p-value = 0.599).

22
 Brand 1
1,000
True
Forecast
800

600
Sales

400

200

0
0 50 100 150 200 250 300 350
Day
Brand 2

True
500 Forecast

400
Sales

300

200

100

0 50 100 150 200 250 300 350

Day

Figure 5: NND out-of-sample forecasts for Brand 1 and Brand 2.

23
 Brand 3
600
True
Forecast

400
Sales

200

0
0 50 100 150 200 250 300 350
Day
Brand 4
800
True
Forecast
600
Sales

400

200

0
0 50 100 150 200 250 300 350
Day

Figure 6: NND out-of-sample forecasts for Brand 3 and Brand 4.

24
 Meter 11 Meter 16
40 True True
Forecast 50 Forecast

30 40
Power

Power
20 30

20
10
0 100 200 300 400 0 100 200 300 400
Time Time
Meter 18 Meter 19

True 40 True
30
Forecast Forecast

30
Power

Power

20

20
10
10

0 100 200 300 400 0 100 200 300 400

Time Time
Figure 7: NND out-of-sample forecasts for meter 11, 16, 18 and 19 (first 72 hours).

25
 Summarizing, we get similar results on both datasets, and this is particu-
larly significant due to the different characteristics of the two datasets. In the
sales data, the bottom-level series are extremely noisy and hard to forecast
(as confirmed by the poor performance of the bottom-up method). On the
other hand, the electricity demand data display seasonality at the bottom-
level, as confirmed by the good performance of the bottom-up method. In
both cases, we generate coherent forecasts (in all our experiments the maxi-
mum violation of the aggregation constraint is less then 10−3 ) and we improve
the overall accuracy at any level of the hierarchy. This confirms the general
viability of our approach.

6. Conclusions
In this paper, we propose a machine learning method for forecasting hier-
archical time series. We formulate the disaggregation problem as a non-linear
regression problem and we solve it with a deep neural network. The network
architecture is able to jointly learn the structure of the hierarchy and gen-
erate coherent forecasts, thanks to the neural network’s ability to extract
meaningful features from the aggregate series and combine them with the
dynamics of the individual series. Furthermore, differently from top-down
approaches, our method allows to easily incorporate any external informa-
tion that affect the time series to disaggregate with no loss of information.
Results demonstrate that our method does not only increase the average
forecasting accuracy of the hierarchy but also addresses the need of building
an automated procedure generating coherent forecasts for many time series
at the same time. Our procedure fulfills the need of scalable algorithms to
automate the process of forecasting hierarchical time series with the aim to
increase the forecasting accuracy at any level. We stress that, differently
from the recently proposed optimal forecast reconciliation approach, in our
method, forecast reconciliation is performed inside the learning process with-
out the need of generating base forecast for all the series of the hierarchy. As
all the top-down approaches, our method highly relies on accurate top-level
forecasts. However, this assumption is often satisfied in hierarchical time
series since the top-level series are in general periodic and less noisy (since
they are the sum of many sub-level components) compared to individual se-
ries at the bottom-level. In summary, our machine learning approach uses
all the relevant information available in the hierarchical structure. This is
important, as particular aggregation levels may reveal hidden features of the

26
data, not easily identifiable at other levels, that are of interest to the user
and needed to be modeled.

References
[1] Abadi, M., Agarwal, A., Barham, P., Brevdo, E., Chen, Z., Citro, C.,
Corrado, G. S., Davis, A., Dean, J., Devin, M., Ghemawat, S., Goodfel-
low, I., Harp, A., Irving, G., Isard, M., Jia, Y., Jozefowicz, R., Kaiser,
L., Kudlur, M., Levenberg, J., Mané, D., Monga, R., Moore, S., Murray,
D., Olah, C., Schuster, M., & Shlens, J. (2015). TensorFlow: Large-scale
machine learning on heterogeneous systems.
[2] Athanasopoulos, G., Ahmed, R. A., & Hyndman, R. J. (2009). Hierar-
chical forecasts for australian domestic tourism. International Journal
of Forecasting, 25 , 146–166.
[3] Caruana, R., Lawrence, S., & Giles, L. (2000). Overfitting in neural nets:
Backpropagation, conjugate gradient, and early stopping. In Proceedings
of the 13th International Conference on Neural Information Processing
Systems NIPS’00 (pp. 381–387). MIT Press.
[4] Dunn, D. M., Williams, W. H., & Dechaine, T. L. (1976). Aggregate
versus subaggregate models in local area forecasting. Journal of the
American Statistical Association, 71 , 68–71.
[5] Gross, C. W., & Sohl, J. E. (1990). Disaggregation methods to expedite
product line forecasting. Journal of forecasting, 9 , 233–254.
[6] Hyndman, R., & Athanasopoulos, G. (2014). Optimally reconciling fore-
casts in a hierarchy. Foresight: The International Journal of Applied
Forecasting, (pp. 42–48).
[7] Hyndman, R., & Athanasopoulos, G. (2018). Forecasting: principles
and practice. OTexts.
[8] Hyndman, R., Lee, A., Wang, E., & Wickramasuriya, S. (2018). hts:
Hierarchical and Grouped Time Series. R package version 5.1.5.
[9] Hyndman, R. J., Ahmed, R. A., Athanasopoulos, G., & Shang, H. L.
(2011). Optimal combination forecasts for hierarchical time series. Com-
putational statistics & data analysis, 55 , 2579–2589.

27
[10] Hyndman, R. J., & Khandakar, Y. (2008). Automatic time series fore-
casting: the forecast package for R. Journal of Statistical Software, 26 ,
1–22.

[11] Hyndman, R. J., & Koehler, A. B. (2006). Another look at measures of

forecast accuracy. International journal of forecasting, 22 , 679–688.

[12] Hyndman, R. J. et al. (2006). Another look at forecast-accuracy met-

rics for intermittent demand. Foresight: The International Journal of
Applied Forecasting, 4 , 43–46.

[13] Ismail Fawaz, H., Forestier, G., Weber, J., Idoumghar, L., & Muller,
P.-A. (2019). Deep learning for time series classification: a review. Data
Mining and Knowledge Discovery, 33 , 917963.

[14] Kingma, D. P., & Ba, J. (2014). Adam: A method for stochastic opti-
mization.

[15] Nenova, Z. D., & May, J. H. (2016). Determining an optimal hierarchical

forecasting model based on the characteristics of the data set: Technical
note. Journal of Operations Management, 44 , 62 – 68.

[16] Nespoli, L., Medici, V., Lopatichki, K., & Sossan, F. (2019). Hierarchical
demand forecasting benchmark for the distribution grid.

[17] Shlifer, E., & Wolff, R. W. (1979). Aggregation and proration in fore-
casting. Management Science, 25 , 594–603.

[18] Timmermann, A. (2006). Forecast combinations. Handbook of economic

forecasting, 1 , 135–196.

[19] Wickramasuriya, S. L., Athanasopoulos, G., & Hyndman, R. J. (2019).

Optimal forecast reconciliation for hierarchical and grouped time series
through trace minimization. Journal of the American Statistical Asso-
ciation, 114 , 804–819.

28