Published as a conference paper at ICLR 2021

A T RANSFORMER - BASED F RAMEWORK FOR M ULTI -

VARIATE T IME S ERIES R EPRESENTATION L EARNING

Anonymous Authors
Paper under double-blind review

A BSTRACT
In this work we propose for the first time a transformer-based framework for un-
supervised representation learning of multivariate time series. Pre-trained models
can be potentially used for downstream tasks such as regression and classifica-
tion, forecasting and missing value imputation. We evaluate our models on several
benchmark datasets for multivariate time series regression and classification and
show that they exceed current state-of-the-art performance, even when the number
of training samples is very limited, while at the same time offering computational
efficiency. We show that unsupervised pre-training of our transformer models
offers a substantial performance benefit over fully supervised learning, even with-
out leveraging additional unlabeled data, i.e., by reusing the same data samples
through the unsupervised objective.

1 I NTRODUCTION
Multivariate time series (MTS) are an important type of data that is ubiquitous in a wide variety
of domains, including science, medicine, finance, engineering and industrial applications. Despite
the recent abundance of MTS data in the much touted era of “Big Data”, the availability of labeled
data in particular is far more limited: extensive data labeling is often prohibitively expensive or
impractical, as it may require much time and effort, special infrastructure or domain expertise. For
this reason, in all aforementioned domains there is great interest in methods which can offer high
accuracy by using only a limited amount of labeled data or by leveraging the existing plethora of
unlabeled data.
There is a large variety of modeling approaches for univariate and multivariate time series, with deep
learning models recently challenging or replacing the state of the art in tasks such as forecasting,
regression and classification (De Brouwer et al., 2019; Tan et al., 2020a; Fawaz et al., 2019b). How-
ever, unlike in domains such as Computer Vision or Natural Language Processing (NLP), the dom-
inance of deep learning for time series is far from established: in fact, non-deep learning methods
such as TS-CHIEF (Shifaz et al., 2020), HIVE-COTE (Lines et al., 2018), and ROCKET (Dempster
et al., 2020) currently hold the record on time series regression and classification dataset bench-
marks (Tan et al., 2020a; Bagnall et al., 2017), matching or even outperforming sophisticated deep
architectures such as InceptionTime (Fawaz et al., 2019a) and ResNet (Fawaz et al., 2019b).
In this work, we investigate, for the first time, the use of a transformer encoder for unsupervised
representation learning of multivariate time series, as well as for the tasks of time series regression
and classification. Transformers are an important, recently developed class of deep learning models,
which were first proposed for the task of natural language translation (Vaswani et al., 2017) but have
since come to monopolize the state-of-the-art performance across virtually all NLP tasks (Raffel
et al., 2019). A key factor for the widespread success of transformers in NLP is their aptitude for
learning how to represent natural language through unsupervised pre-training (Brown et al., 2020;
Raffel et al., 2019; Devlin et al., 2018). Besides NLP, transformers have also set the state of the art
in several domains of sequence generation, such as polyphonic music composition (Huang et al.,
2018).
Transformer models are based on a multi-headed attention mechanism that offers several key ad-
vantages and renders them particularly suitable for time series data (see Appendix section A.4 for
details). Inspired by the impressive results attained through unsupervised pre-training of transformer
models in NLP, as our main contribution, in the present work we develop a generally applicable

1
Published as a conference paper at ICLR 2021

methodology (framework) that can leverage unlabeled data by first training a transformer model
to extract dense vector representations of multivariate time series through an input denoising (au-
toregressive) objective. The pre-trained model can be subsequently applied to several downstream
tasks, such as regression, classification, imputation, and forecasting. Here, we apply our framework
for the tasks of multivariate time series regression and classification on several public datasets and
demonstrate that transformer models can convincingly outperform all current state-of-the-art model-
ing approaches, even when only having access to a very limited amount of training data samples (on
the order of hundreds of samples), an unprecedented success for deep learning models. Importantly,
despite common preconceptions about transformers from the domain of NLP, where top performing
models have billions of parameters and require days to weeks of pre-training on many parallel GPUs
or TPUs, we also demonstrate that our models, using at most hundreds of thousands of parameters,
can be trained even on CPUs, while training them on GPUs allows them to be trained as fast as even
the fastest and most accurate non-deep learning based approaches.

2 R ELATED W ORK

Regression and classification of time series: Currently, non-deep learning methods such as TS-
CHIEF (Shifaz et al., 2020), HIVE-COTE (Lines et al., 2018), and ROCKET (Dempster et al.,
2020) constitute the state of the art for time series regression and classification based on evaluations
on public benchmarks (Tan et al., 2020a; Bagnall et al., 2017), followed by CNN-based deep archi-
tectures such as InceptionTime (Fawaz et al., 2019a) and ResNet (Fawaz et al., 2019b). ROCKET,
which on average is the best ranking method, is a fast method that involves training a linear classi-
fier on top of features extracted by a flat collection of numerous and various random convolutional
kernels. HIVE-COTE and TS-CHIEF (itself inspired by Proximity Forest (Lucas et al., 2019)),
are very sophisticated methods which incorporate expert insights on time series data and consist
of large, heterogeneous ensembles of classifiers utilizing shapelet transformations, elastic similarity
measures, spectral features, random interval and dictionary-based techniques; however, these meth-
ods are highly complex, involve significant computational cost, cannot benefit from GPU hardware
and scale poorly to datasets with many samples and long time series; moreover, they have been
developed for and only been evaluated on univariate time series.
Unsupervised learning for multivariate time series: Recent work on unsupervised learning for
multivariate time series has predominantly employed autoencoders, trained with an input recon-
struction objective and implemented either as Multi-Layer Perceptrons (MLP) or RNN (most com-
monly, LSTM) networks. As interesting variations of the former, Kopf et al. (2019) and Fortuin
et al. (2019) additionally incorporated Variational Autoencoding into this approach, but focused on
clustering and the visualization of shifting sample topology with time. As an example of the lat-
ter, Malhotra et al. (2017) presented a multi-layered RNN sequence-to-sequence autoencoder, while
Lyu et al. (2018) developed a multi-layered LSTM with an attention mechanism and evaluated both
an input reconstruction (autoencoding) as well as a forecasting loss for unsupervised representation
learning of Electronic Healthcare Record multivariate time series.
As a novel take on autoencoding, and with the goal of dealing with missing data, Bianchi et al.
(2019) employ a stacked bidirectional RNN encoder and stacked RNN decoder to reconstruct the
input, and at the same time use a user-provided kernel matrix as prior information to condition
internal representations and encourage learning similarity-preserving representations of the input.
They evaluate the method on the tasks of missing value imputation and classification of time series
under increasing “missingness” of values.
A distinct approach is followed by Zhang et al. (2019), who use a composite convolutional - LSTM
network with attention and a loss which aims at reconstructing correlation matrices between the
variables of the multivariate time series input. They use and evaluate their method only for the task
of anomaly detection.
Finally, Jansen et al. (2018) rely on a triplet loss and the idea of temporal proximity (the loss rewards
similarity of representations between proximal segments and penalizes similarity between distal
segments of the time series) for unsupervised representation learning of non-speech audio data.
This idea is explored further by Franceschi et al. (2019), who combine the triplet loss with a deep
causal dilated CNN, in order to make the method effective for very long time series.

2
Published as a conference paper at ICLR 2021

Transformer models for time series: Recently, a full encoder-decoder transformer architecture
was employed for univariate time series forecasting: Li et al. (2019) showed superior performance
compared to the classical statistical method ARIMA, the recent matrix factorization method TRMF,
an RNN-based autoregressive model (DeepAR) and an RNN-based state space model (DeepState)
on 4 public forecasting datasets, while Wu et al. (2020) used a transformer to forecast influenza
prevalence and similarly showed performance benefits compared to ARIMA, an LSTM and a GRU
Seq2Seq model with attention, and Lim et al. (2020) used a transformer for multi-horizon uni-
variate forecasting, supporting interpretation of temporal dynamics. Finally, Ma et al. (2019) use an
encoder-decoder architecture with a variant of self-attention for imputation of missing values in mul-
tivariate, geo-tagged time series and outperform classic as well as the state-of-the-art, RNN-based
imputation methods on 3 public and 2 competition datasets for imputation.
By contrast, our work aspires to generalize the use of transformers from solutions to specific gen-
erative tasks (which require the full encoder-decoder architecture) to a framework which allows for
unsupervised pre-training and with minor modifications can be readily used for a wide variety of
downstream tasks; this is analogous to the way BERT (Devlin et al., 2018) converted a translation
model into a generic framework based on unsupervised learning, an approach which has become a
de facto standard and established the dominance of transformers in NLP.

3 M ETHODOLOGY

3.1 BASE MODEL

At the core of our method lies a transformer encoder, as described in the original transformer work
by Vaswani et al. (2017); however, we do not use the decoder part of the architecture. A schematic
diagram of the generic part of our model, common across all considered tasks, is shown in Figure 1.
We refer the reader to the original work for a detailed description of the transformer model, and here
present the proposed changes that make it compatible with multivariate time series data, instead of
sequences of discrete word indices.
In particular, each training sample X ∈ Rw×m , which is a multivariate time series of length w
and m different variables, constitutes a sequence of w feature vectors xt ∈ Rm : X ∈ Rw×m =
[x1 , x2 , . . . , xw ]. The original feature vectors xt are first normalized (for each dimension, we sub-
tract the mean and divide by the variance across the training set samples) and then linearly projected
onto a d-dimensional vector space, where d is the dimension of the transformer model sequence
element representations (typically called model dimension):

ut = Wp xt + bp (1)

where Wp ∈ Rd×m , bp ∈ Rd are learnable parameters and ut ∈ Rd , t = 0, . . . , w are the model

input vectors1 . These will become the queries, keys and values of the self-attention layer, after
adding the positional encodings and multiplying by the corresponding matrices.
We note that the above formulation also covers the univariate time series case, i.e., m = 1, although
we only evaluate our approach on multivariate time series in the scope of this work. We additionally
note that the input vectors ut need not necessarily be obtained from the (transformed) feature vectors
at a time step t: because the computational complexity of the model scales as O(w2 ) and the number
of parameters2 as O(w) with the input sequence length w, to obtain ut in case the granularity
(temporal resolution) of the data is very fine, one may instead use a 1D-convolutional layer with 1
input and d output channels and kernels Ki of size (k, m), where k is the width in number of time
steps and i the output channel:

XX
ut i = u(t, i) = x(t + j, h)Ki (j, h), i = 1, . . . , d (2)
j h

1
Although equation 1 shows the operation for a single time step for clarity, all input vectors are embedded
concurrently by a single matrix-matrix multiplication
2
Specifically, the learnable positional encoding, batch normalization and output layers

3
Published as a conference paper at ICLR 2021

Figure 1: Left: Generic model architecture, common to all tasks. The feature vector xt at each time
step t is linearly projected to a vector ut of the same dimensionality d as the internal representation
vectors of the model and is fed to the first self-attention layer to form the keys, queries and values
after adding a positional encoding. Right: Training setup of the unsupervised pre-training task.
We mask a proportion r of each variable sequence in the input independently, such that across each
variable, time segments of mean length lm are masked, each followed by an unmasked segment
of mean length lu = 1−r r lm . Using a linear layer on top of the final vector representations zt , at
each time step the model tries to predict the full, uncorrupted input vectors xt ; however, only the
predictions on the masked values are considered in the Mean Squared Error loss.

In this way, one may control the temporal resolution by using a stride or dilation factor greater than
1. Moreover, although in the present work we only used equation 1, one may use equation 2 as an
input to compute the keys and queries and equation 1 to compute the values of the self-attention
layer. This is particularly useful in the case of univariate time series, where self-attention would
otherwise match (consider relevant/compatible) all time steps which share similar values for the
independent variable, as noted by Li et al. (2019).
Finally, since the transformer is a feed-forward architecture that is insensitive to the ordering of in-
put, in order to make it aware of the sequential nature of the time series, we add positional encodings
Wpos ∈ Rw×d to the input vectors U ∈ Rw×d = [u1 , . . . , uw ]: U 0 = U + Wpos .
Instead of deterministic, sinusoidal encodings, which were originally proposed by Vaswani et al.
(2017), we use fully learnable positional encodings, as we observed that they perform better for all
datasets presented in this work. Based on the performance of our models, we also observe that the
positional encodings generally appear not to significantly interfere with the numerical information
of the time series, similar to the case of word embeddings; we hypothesize that this is because they
are learned so as to occupy a different, approximately orthogonal, subspace to the one in which the
projected time series samples reside. This approximate orthogonality condition is much easier to
satisfy in high dimensional spaces.
An important consideration regarding time series data is that individual samples may display con-
siderable variation in length. This issue is effectively dealt with in our framework: after setting a
maximum sequence length w for the entire dataset, shorter samples are padded with arbitrary values,
and we generate a padding mask which adds a large negative value to the attention scores for the
padded positions, before computing the self-attention distribution with the softmax function. This
forces the model to completely ignore padded positions, while allowing the parallel processing of
samples in large minibatches.

4
Published as a conference paper at ICLR 2021

Transformers in NLP use layer normalization after computing self-attention and after the feed-
forward part of each encoder block, leading to significant performance gains over batch normaliza-
tion, as originally proposed by Vaswani et al. (2017). However, here we instead use batch normaliza-
tion, because it can mitigate the effect of outlier values in time series, an issue that does not arise in
NLP word embeddings. Additionally, the inferior performance of batch normalization in NLP has
been mainly attributed to extreme variation in sample length (i.e., sentences in most tasks) (Shen
et al., 2020), while in the datasets we examine this variation is much smaller. In Table 11 of the
Appendix we show that batch normalization can indeed offer a very significant performance benefit
over layer normalization, while the extent can vary depending on dataset characteristics.

3.2 R EGRESSION AND CLASSIFICATION

The base model architecture presented in Section 3.1 and depicted in Figure 1 can be used for the
purposes of regression and classification with the following modification: the final representation
vectors zt ∈ Rd corresponding to all time steps are concatenated into a single vector z̄ ∈ Rd·w =
[z1 ; . . . ; zw ], which serves as the input to a linear output layer with parameters Wo ∈ Rn×(d·w) ,
bo ∈ Rn , where n is the number of scalars to be estimated for the regression problem (typically
n = 1), or the number of classes for the classification problem:

ŷ = Wo z̄ + bo (3)

In the case of regression, the loss for a single data sample will simply be the squared error L =
2
kŷ − yk , where y ∈ Rn are the ground truth values. We clarify that regression in the context of
this work means predicting a numeric value for a given sequence (time series sample). This numeric
value is of a different nature than the numerical data appearing in the time series: for example,
given a sequence of simultaneous temperature and humidity measurements of 9 rooms in a house,
as well as weather and climate data such as temperature, pressure, humidity, wind speed, visibility
and dewpoint, we wish to predict the total energy consumption in kWh of a house for that day. The
parameter n corresponds to the number of scalars (or the dimensionality of a vector) to be estimated.
In the case of classification, the predictions ŷ will additionally be passed through a softmax function
to obtain a distribution over classes, and its cross-entropy with the categorical ground truth labels
will be the sample loss.
Finally, when fine-tuning the pre-trained models, we allow training of all weights; instead, freezing
all layers except for the output layer would be equivalent to using static, pre-extracted time-series
representations of the time series. In Table 12 in the Appendix we show the trade-off in terms of
speed and performance when using a fully trainable model versus static representations.

3.3 U NSUPERVISED PRE - TRAINING

As a task for the unsupervised pre-training of our model we consider the autoregressive task of de-
noising the input: specifically, we set part of the input to 0 and ask the model to predict the masked
values. The corresponding setup is depicted in the right part of Figure 1. A binary noise mask
M ∈ Rw×m , is created independently for each training sample, and the input is masked by elemen-
twise multiplication: X̃ = M  X. On average, a proportion r of each mask column of length w
(corresponding to a single variable in the multivariate time series) is set to 0 by alternating between
segments of 0s and 1s. We choose the state transition probabilities such that each masked segment
(sequence of 0s) has a length that follows a geometric distribution with mean lm and is succeeded by
an unmasked segment (sequence of 1s) of mean length lu = 1−r r lm . We chose lm = 3 for all pre-
sented experiments. The reason why we wish to control the length of the masked sequence, instead
of simply using a Bernoulli distribution with parameter r to set all mask elements independently at
random, is that very short masked sequences (e.g., of 1 masked element) in the input can often be
trivially predicted with good approximation by replicating the immediately preceding or succeeding
values or by the average thereof. In order to obtain enough long masked sequences with relatively
high likelihood, a very high masking proportion r would be required, which would render the over-
all task detrimentally challenging. Following the process above, at each time step on average r · m
variables will be masked. We empirically chose r = 0.15 for all presented experiments. This input
masking process is different from the “cloze type” masking used by NLP models such as BERT,

5
Published as a conference paper at ICLR 2021

where a special token and thus word embedding vector replaces the original word embedding, i.e.,
the entire feature vector at affected time steps. We chose this masking pattern because it encour-
ages the model to learn to attend both to preceding and succeeding segments in individual variables,
as well as to existing contemporary values of the other variables in the time series, and thereby to
learn to model inter-dependencies between variables. In Table 10 in the Appendix we show that this
masking scheme is more effective than other possibilities for denoising the input.
Using a linear layer with parameters Wo ∈ Rm×d , bo ∈ Rm on top of the final vector repre-
sentations zt ∈ Rd , for each time step the model concurrently outputs its estimate x̂t of the full,
uncorrupted input vectors xt ; however, only the predictions on the masked values (with indices in
the set M ≡ {(t, i) : mt,i = 0}, where mt,i are the elements of the mask M), are considered in the
Mean Squared Error loss for each data sample:

x̂t = Wo zt + bo (4)
1 XX 2
LMSE = (x̂(t, i) − x(t, i)) (5)
|M |
(t,i)∈M

This objective differs from the one used by denoising autoencoders, where the loss considers re-
construction of the entire input, under (typically Gaussian) noise corruption. Also, we note that the
approach described above differs from simple dropout on the input embeddings, both with respect to
the statistical distributions of masked values, as well as the fact that here the masks also determine
the loss function. In fact, we additionally use a dropout of 10% when training all of our supervised
and unsupervised models.

4 E XPERIMENTS & R ESULTS

In the experiments reported below we use the predefined training - test set splits of the benchmark
datasets and train all models long enough to ensure convergence. We do this to account for the fact
that training the transformer models in a fully supervised way typically requires more epochs than
fine-tuning the ones which have already been pre-trained using the unsupervised methodology of
Section 3.3. Because the benchmark datasets are very heterogeneous in terms of number of samples,
dimensionality and length of the time series, as well as the nature of the data itself, we observed that
we can obtain better performance by a cursory tuning of hyperparameters (such as the number of
encoder blocks, the representation dimension, number of attention heads or dimension of the feed-
forward part of the encoder blocks) separately for each dataset. To select hyperparameters, for each
dataset we randomly split the training set in two parts, 80%-20%, and used the 20% as a validation
set for hyperparameter tuning. After fixing the hyperparameters, the entire training set was used to
train the model again, which was finally evaluated on the test set. A set of hyperparameters which
has consistently good performance on all datasets is shown in Table 14 in the Appendix, alongside
the hyperparameters that we have found to yield the best performance for each dataset (Tables 15,
16, 17, 18.

4.1 R EGRESSION

We select a diverse range of 6 datasets from the Monash University, UEA, UCR Time Series Regres-
sion Archive Tan et al. (2020a) in a way so as to ensure diversity with respect to the dimensionality
and length of time series samples, as well as the number of samples (see Appendix Table 3 for
dataset characteristics). Table 1 shows the Root Mean Squared Error achieved by of our models,
named TST for “Time Series Transformer”, including a variant trained only through supervision,
and one first pre-trained on the same training set in an unsupervised way. We compare them with
the currently best performing models as reported in the archive. Our transformer models rank first
on all but two of the examined datasets, for which they rank second. They thus achieve an average
rank of 1.33, setting them clearly apart from all other models; the overall second best model, XG-
Boost, has an average rank of 3.5, ROCKET (which outperformed ours on one dataset) on average
ranks in 5.67th place and Inception (which outperformed ours on the second dataset) also has an
average rank of 5.67. On average, our models attain 30% lower RMSE than the mean RMSE among

6
Published as a conference paper at ICLR 2021

all models, and approx. 16% lower RMSE than the overall second best model (XGBoost), with ab-
solute improvements varying among datasets from approx. 4% to 36%. We note that all other deep
learning methods achieve performance close to the middle of the ranking or lower. In Table 1 we
report the ”average relative difference from mean” metric rj for each model j, the over N datasets:
N M
1 X R(i, j) − R̄i 1 X
rj = , R̄i = R(i, k)
N i=1 R̄i M
k=1

, where R(i, j) is the RMSE of model j on dataset i and M is the number of models.
Importantly, we also observe that the pre-trained transformer models outperform the fully super-
vised ones in 3 out of 6 datasets. This is interesting, because no additional samples are used for
pre-training: the benefit appears to originate from reusing the same training samples for learning
through an unsupervised objective. To further elucidate this observation, we investigate the follow-
ing questions:
Q1: Given a partially labeled dataset of a certain size, how will additional labels affect perfor-
mance? This pertains to one of the most important decisions that data owners face, namely, to what
extent will further annotation help. To clearly demonstrate this effect, we choose the largest dataset
we have considered from the regression archive (12.5k samples), in order to avoid the variance in-
troduced by small set sizes. The left panel of Figure 2 (where each marker is an experiment) shows
how performance on the entire test set varies with an increasing proportion of labeled training set
data used for supervised learning. As expected, with an increasing proportion of available labels
performance improves both for a fully supervised model, as well as the same model that has been
first pre-trained on the entire training set through the unsupervised objective and then fine-tuned. In-
terestingly, not only does the pre-trained model outperform the fully supervised one, but the benefit
persists throughout the entire range of label availability, even when the models are allowed to use all
labels; this is consistent with our previous observation on Table 1 regarding the advantage of reusing
samples.
Q2: Given a labeled dataset, how will additional unlabeled samples affect performance? In
other words, to what extent does unsupervised learning make it worth collecting more data, even
if no additional annotations are available? This question differs from the above, as we now only
scale the availability of data samples for unsupervised pre-training, while the number of labeled
samples is fixed. The right panel of Figure 2 (where each marker is an experiment) shows that,
for a given number of labels (shown as a percentage of the totally available labels), the more data
samples are used for unsupervised learning, the lower the error achieved (note that the horizontal
axis value 0 corresponds to fully supervised training only, while all other values to unsupervised
pre-training followed by supervised fine-tuning). This trend is more linear in the case of supervised
learning on 20% of the labels (approx. 2500). Likely due to a small sample (here, meaning set)
effect, in the case of having only 10% of the labels (approx. 1250) for supervised learning, the error
first decreases rapidly as we use more samples for unsupervised pre-training, and then momentarily
increases, before it decreases again (for clarity, the same graphs are shown separately in Figure 3
in the Appendix). Consistent with our observations above, it is interesting to again note that, for a
given number of labeled samples, even reusing a subset of the same samples for unsupervised pre-
training improves performance: for the 1250 labels (blue diamonds of the right panel of Figure 2 or
left panel of Figure 3 in the Appendix) this can be observed in the horizontal axis range [0, 0.1], and
for the 2500 labels (blue diamonds of the right panel of Figure 2 or right panel of Figure 3 in the
Appendix) in the horizontal axis range [0, 0.2].

4.2 C LASSIFICATION

We select a set of 11 multivariate datasets from the UEA Time Series Classification Archive (Bag-
nall et al., 2018) with diverse characteristics in terms of the number, dimensionality and length of
time series samples, as well as the number of classes (see Appendix Table 4). As this archive is new,
there have not been many reported model evaluations; we follow Franceschi et al. (2019) and use
as a baseline the best performing method studied by the creators of the archive, DTWD (dimension-
Dependent DTW), together with the method proposed by Franceschi et al. (2019) themselves (a
dilation-CNN leveraging unsupervised and supervised learning). Additionally, we use the publicly
available implementations Tan et al. (2020b) of ROCKET, which is currently the top performing

7
 Published as a conference paper at ICLR 2021

70 70
68
68
66

64
66
62
RMSE

RMSE
60 64
58
62
56 Unsup. pretrained 0.1 of labels
54
Supervised 60 0.2 of labels
52

50 58
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Availability of labels Use of data for unsupervised pre-training

Figure 2: Dataset: BeijingPM25Quality. Left: Root Mean Squared Error of a fully supervised
transformer (orange circles) and the same model pre-trained (blue diamonds) on the training set
through the unsupervised objective and then fine-tuned on available labels, versus the proportion of
labeled data in the training set. Right: Root Mean Squared Error of a given model as a function of
the number of samples (here, shown as a proportion of the total number of samples in the training
set) used for unsupervised pre-training. For supervised learning, two levels of label availability
are depicted: 10% (purple circles) and 20% (green squares) of all training data labels. Note that a
horizontal axis value of 0 means fully supervised learning only, while all other values correspond to
unsupervised pre-training followed by supervised fine-tuning.

Root MSE
Ours
Random 5-NN 1-NN- 5-NN- TST TST
Dataset SVR XGBoost 1-NN-ED Rocket FCN ResNet Inception
˜Forest -ED DTWD DTWD (sup. only) (pretrained)
AppliancesEnergy 3.457 3.455 3.489 5.231 4.227 6.036 4.019 2.299 2.865 3.065 4.435 2.228 2.375
BenzeneConcentr. 4.790 0.855 0.637 6.535 5.844 4.983 4.868 3.360 4.988 4.061 1.584 0.517 0.494
BeijingPM10 110.574 94.072 93.138 139.229 115.669 139.134 115.502 120.057 94.348 95.489 96.749 91.344 86.866
BeijingPM25 75.734 63.301 59.495 88.193 74.156 88.256 72.717 62.769 59.726 64.462 62.227 60.357 53.492
LiveFuelMoisture 43.021 44.657 44.295 58.238 46.331 57.111 46.290 41.829 47.877 51.632 51.539 42.607 43.138
IEEEPPG 36.301 32.109 31.487 33.208 27.111 37.140 33.572 36.515 34.325 33.150 23.903 25.042 27.806
Avg Rel. diff. from mean 0.097 -0.172 -0.197 0.377 0.152 0.353 0.124 -0.048 0.021 0.005 -0.108 -0.301 -0.303
Avg Rank 7.166 4.5 3.5 10.833 8 11.167 7.667 5.667 6.167 6.333 5.666 1.333

Table 1: Performance on multivariate regression datasets, in terms of Root Mean Squared Error.
Bold indicates best values, underlining indicates second best.

model for univariate time series and one of the best in our regression evaluation, and XGBoost,
which is one of the most commonly used models for univariate and multivariate time series, and
also the best baseline model in our regression evaluation (Section 4.1). Finally, we did not find any
reported evaluations of RNN-based models on any of the UCR/UEA archives, possibly because of
a common perception for long training and inference times, as well as difficulty in training (Fawaz
et al., 2019b); therefore, we implemented a stacked LSTM model and also include it in the compar-
ison. The performance of the baselines alongside our own models are shown in Table 2 in terms of
accuracy, to allow comparison with reported values.
It can be seen that our models performed best on 7 out of the 11 datasets, achieving an average rank
of 1.7, followed by ROCKET, which performed best on 3 datasets and on average ranked 2.3th.
The dilation-CNN (Franceschi et al., 2019) and XGBoost, which performed best on the remaining
1 dataset, tied and on average ranked 3.7th and 3.8th respectively. Interestingly, we observe that
all datasets on which ROCKET outperformed our model were very low dimensional (specifically,
3-dimensional). Although our models still achieved the second best performance for UWaveG-
estureLibrary, in general we believe that this indicates a relative weakness of our current models
when dealing with very low dimensional time series. As discussed in Section 3.1, this may be
due to the problems introduced by a low-dimensional representation space to the attention mecha-
nism, as well as the added positional embeddings; to mitigate this issue, in future work we intend
to use a 1D-convolutional layer to extract more meaningful representations of low-dimensional in-
put features (see Section 3.1). Conversely, our models performed particularly well on very high-

8
 Published as a conference paper at ICLR 2021

Ours
Dataset TST (pretrained) TST (sup. only) Rocket XGBoost LSTM Frans. et al DTW D
EthanolConcentration 0.326 0.337 0.452 0.437 0.323 0.289 0.323
FaceDetection 0.689 0.681 0.647 0.633 0.577 0.528 0.529
Handwriting 0.359 0.305 0.588 0.158 0.152 0.533 0.286
Heartbeat 0.776 0.776 0.756 0.732 0.722 0.756 0.717
JapaneseVowels 0.997 0.994 0.962 0.865 0.797 0.989 0.949
InsectWingBeat 0.687 0.684 - 0.369 0.176 0.16 -
PEMS-SF 0.896 0.919 0.751 0.983 0.399 0.688 0.711
SelfRegulationSCP1 0.922 0.925 0.908 0.846 0.689 0.846 0.775
SelfRegulationSCP2 0.604 0.589 0.533 0.489 0.466 0.556 0.539
SpokenArabicDigits 0.998 0.993 0.712 0.696 0.319 0.956 0.963
UWaveGestureLibrary 0.913 0.903 0.944 0.759 0.412 0.884 0.903
Avg Accuracy
0.748 0.742 0.725 0.659 0.486 0.703 0.669
(excl. InsectWingBeat)
Avg Rank 1.7 2.3 3.8 5.4 3.7 4.1

Table 2: Accuracy on multivariate classification datasets. Bold indicates best and underlining
second best values. A dash indicates that the corresponding method failed to run on this dataset.

dimensional datasets (FaceDetection, HeartBeat, InsectWingBeat, PEMS-SF), and/or datasets with

relatively more training samples. As a characteristic example, on InsectWingBeat (which is by far
the largest dataset with 30k samples and contains time series of 200 dimensions and highly irregular
length) our model reached an accuracy of 0.689, while all other methods performed very poorly - the
second best was XGBoost with an accuracy of 0.369. However, we note that our model performed
exceptionally well also on datasets with only a couple of hundred samples, which in fact constitute
8 out of the 11 examined datasets.
Finally, we observe that the pre-trained transformer models performed better than the fully super-
vised ones in 8 out of 11 datasets, sometimes by a substantial margin.Again, no additional samples
were available for unsupervised pre-training, so the benefit appears to originate from reusing the
same samples.

5 C ONCLUSION

In this work we propose a novel framework for multivariate time series representation learning
based on the transformer encoder architecture. The framework includes an unsupervised pre-training
scheme, which we show that can offer substantial performance benefits over fully supervised learn-
ing, even without leveraging additional unlabeled data, i.e., by reusing the same data samples. By
evaluating our framework on several public multivariate time series datasets from various domains
and with diverse characteristics, we demonstrate that it is currently the best performing method
for regression and classification, even for datasets where only a few hundred training samples are
available.

R EFERENCES
A. Bagnall, J. Lines, A. Bostrom, J. Large, and E. Keogh. The great time series classification
bake off: a review and experimental evaluation of recent algorithmic advances. Data Mining and
Knowledge Discovery, 31:606–660, 2017.

Anthony Bagnall, Hoang Anh Dau, Jason Lines, Michael Flynn, James Large, Aaron Bostrom, Paul
Southam, and Eamonn Keogh. The UEA multivariate time series classification archive, 2018.
arXiv:1811.00075 [cs, stat], October 2018. arXiv: 1811.00075.

Iz Beltagy, Matthew E. Peters, and Arman Cohan. Longformer: The Long-Document Transformer.
arXiv:2004.05150 [cs], April 2020. URL http://arxiv.org/abs/2004.05150. arXiv:
2004.05150.

9
Published as a conference paper at ICLR 2021

Filippo Maria Bianchi, Lorenzo Livi, Karl Øyvind Mikalsen, Michael Kampffmeyer, and Robert
Jenssen. Learning representations of multivariate time series with missing data. Pattern
Recognition, 96:106973, December 2019. ISSN 0031-3203. doi: 10.1016/j.patcog.2019.106973.
T. Brown, B. Mann, Nick Ryder, Melanie Subbiah, J. Kaplan, P. Dhariwal, Arvind Neelakantan,
Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, G. Krüger,
Tom Henighan, R. Child, Aditya Ramesh, D. Ziegler, Jeffrey Wu, Clemens Winter, Christopher
Hesse, Mark Chen, E. Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, J. Clark, Christopher
Berner, Sam McCandlish, A. Radford, Ilya Sutskever, and Dario Amodei. Language models are
few-shot learners. ArXiv, abs/2005.14165, 2020.
Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc V. Le, and Ruslan Salakhutdinov.
Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context. arXiv:1901.02860
[cs, stat], June 2019. URL http://arxiv.org/abs/1901.02860. arXiv: 1901.02860.
Edward De Brouwer, Jaak Simm, Adam Arany, and Yves Moreau. GRU-ODE-Bayes: Continuous
Modeling of Sporadically-Observed Time Series. In H. Wallach, H. Larochelle, A. Beygelzimer,
F. d\textquotesingle Alché-Buc, E. Fox, and R. Garnett (eds.), Advances in Neural Information
Processing Systems 32, pp. 7379–7390. Curran Associates, Inc., 2019.
Angus Dempster, Franccois Petitjean, and Geoffrey I. Webb. ROCKET: exceptionally fast and accu-
rate time series classification using random convolutional kernels. Data Mining and Knowledge
Discovery, 2020. doi: 10.1007/s10618-020-00701-z.
Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of Deep
Bidirectional Transformers for Language Understanding. CoRR, abs/1810.04805, 2018. eprint:
1810.04805.
H. Fawaz, B. Lucas, G. Forestier, Charlotte Pelletier, D. Schmidt, Jonathan Weber, Geoffrey I. Webb,
L. Idoumghar, Pierre-Alain Muller, and Franccois Petitjean. InceptionTime: Finding AlexNet for
Time Series Classification. ArXiv, 2019a. doi: 10.1007/s10618-020-00710-y.
Hassan Fawaz, Germain Forestier, Jonathan Weber, Lhassane Idoumghar, and Pierre-Alain Muller.
Deep learning for time series classification: a review. Data Mining and Knowledge Discovery, 33
(4):917–963, July 2019b. ISSN 1573-756X. doi: 10.1007/s10618-019-00619-1.
Vincent Fortuin, M. Hüser, Francesco Locatello, Heiko Strathmann, and G. Rätsch. SOM-VAE:
Interpretable Discrete Representation Learning on Time Series. ICLR, 2019.
Jean-Yves Franceschi, Aymeric Dieuleveut, and Martin Jaggi. Unsupervised Scalable Represen-
tation Learning for Multivariate Time Series. In H. Wallach, H. Larochelle, A. Beygelzimer,
F. d\textquotesingle Alché-Buc, E. Fox, and R. Garnett (eds.), Advances in Neural Information
Processing Systems 32, pp. 4650–4661. Curran Associates, Inc., 2019.
Sepp Hochreiter. The Vanishing Gradient Problem During Learning Recurrent Neural Nets and
Problem Solutions. Int. J. Uncertain. Fuzziness Knowl.-Based Syst., 6(2):107–116, April 1998.
ISSN 0218-4885. doi: 10.1142/S0218488598000094. Place: River Edge, NJ, USA Publisher:
World Scientific Publishing Co., Inc.
Cheng-Zhi Anna Huang, Ashish Vaswani, Jakob Uszkoreit, Ian Simon, Curtis Hawthorne, Noam
Shazeer, Andrew M Dai, Matthew D Hoffman, Monica Dinculescu, and Douglas Eck. Music
transformer: Generating music with long-term structure. In International Conference on Learning
Representations, 2018.
A. Jansen, M. Plakal, Ratheet Pandya, D. Ellis, Shawn Hershey, Jiayang Liu, R. C. Moore, and R. A.
Saurous. Unsupervised Learning of Semantic Audio Representations. 2018 IEEE International
Conference on Acoustics, Speech and Signal Processing (ICASSP), 2018. doi: 10.1109/ICASSP.
2018.8461684.
A. Kopf, Vincent Fortuin, Vignesh Ram Somnath, and M. Claassen. Mixture-of-Experts Variational
Autoencoder for clustering and generating from similarity-based representations. ICLR 2019,
2019.

10
Published as a conference paper at ICLR 2021

Shiyang Li, Xiaoyong Jin, Yao Xuan, Xiyou Zhou, Wenhu Chen, Yu-Xiang Wang, and Xifeng
Yan. Enhancing the locality and breaking the memory bottleneck of transformer on time series
forecasting. In Advances in Neural Information Processing Systems, pp. 5243–5253, 2019.

Bryan Lim, Sercan O. Arik, Nicolas Loeff, and Tomas Pfister. Temporal fusion transformers for
interpretable multi-horizon time series forecasting, 2020.

J. Lines, Sarah Taylor, and Anthony J. Bagnall. Time Series Classification with HIVE-COTE. ACM
Trans. Knowl. Discov. Data, 2018. doi: 10.1145/3182382.

Benjamin Lucas, Ahmed Shifaz, Charlotte Pelletier, Lachlan O’Neill, Nayyar Zaidi, Bart Goethals,
Francois Petitjean, and Geoffrey I. Webb. Proximity Forest: An effective and scalable distance-
based classifier for time series. Data Mining and Knowledge Discovery, 33(3):607–635, May
2019. ISSN 1384-5810, 1573-756X. doi: 10.1007/s10618-019-00617-3. arXiv: 1808.10594.

Xinrui Lyu, Matthias Hueser, Stephanie L. Hyland, George Zerveas, and Gunnar Raetsch. Improving
Clinical Predictions through Unsupervised Time Series Representation Learning. In Proceedings
of the NeurIPS 2018 Workshop on Machine Learning for Health, 2018. eprint: 1812.00490.

J. Ma, Zheng Shou, Alireza Zareian, Hassan Mansour, A. Vetro, and S. Chang. Cdsa:
Cross-dimensional self-attention for multivariate, geo-tagged time series imputation. ArXiv,
abs/1905.09904, 2019.

P. Malhotra, T. Vishnu, L. Vig, Puneet Agarwal, and G. Shroff. TimeNet: Pre-trained deep recurrent
neural network for time series classification. ESANN, 2017.

Razvan Pascanu, Tomas Mikolov, and Yoshua Bengio. On the difficulty of training recurrent neural
networks. In Proceedings of the 30th International Conference on International Conference on
Machine Learning - Volume 28, ICML’13, pp. III–1310–III–1318, Atlanta, GA, USA, June 2013.
JMLR.org.

Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi
Zhou, W. Li, and Peter J. Liu. Exploring the Limits of Transfer Learning with a Unified Text-to-
Text Transformer. ArXiv, abs/1910.10683, 2019.

Sheng Shen, Zhewei Yao, Amir Gholami, Michael W. Mahoney, and Kurt Keutzer. PowerNorm:
Rethinking Batch Normalization in Transformers. arXiv:2003.07845 [cs], June 2020. arXiv:
2003.07845.

Ahmed Shifaz, Charlotte Pelletier, F. Petitjean, and Geoffrey I. Webb. TS-CHIEF: a scalable and
accurate forest algorithm for time series classification. Data Mining and Knowledge Discovery,
2020. doi: 10.1007/s10618-020-00679-8.

C. Tan, C. Bergmeir, François Petitjean, and Geoffrey I. Webb. Monash University, UEA, UCR
Time Series Regression Archive. ArXiv, 2020a.

Chang Wei Tan, Christoph Bergmeir, Francois Petitjean, and Geoffrey I Webb. Time series regres-
sion. arXiv preprint arXiv:2006.12672, 2020b.

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez,
Łukasz Kaiser, and Illia Polosukhin. Attention is All you Need. In I. Guyon, U. V. Luxburg,
S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett (eds.), Advances in Neural
Information Processing Systems 30, pp. 5998–6008. Curran Associates, Inc., 2017.

Neo Wu, Bradley Green, Xue Ben, and Shawn O’Banion. Deep transformer models for time series
forecasting: The influenza prevalence case, 2020.

Chuxu Zhang, Dongjin Song, Yuncong Chen, Xinyang Feng, C. Lumezanu, Wei Cheng, Jingchao
Ni, B. Zong, H. Chen, and Nitesh V. Chawla. A Deep Neural Network for Unsupervised Anomaly
Detection and Diagnosis in Multivariate Time Series Data. In AAAI, 2019. doi: 10.1609/aaai.
v33i01.33011409.

11
Published as a conference paper at ICLR 2021

A A PPENDIX

A.1 A DDITIONAL POINTS & F UTURE W ORK

Execution time for training: While a precise comparison in terms of training time is well out of
scope for the present work, in Section A.3 of the Appendix we demonstrate that our transformer-
based method is economical in terms of its use of computational resources. However, alternative
self-attention schemes, such as sparse attention patterns (Li et al., 2019), recurrence (Dai et al.,
2019) or compressed (global-local) attention (Beltagy et al., 2020), can help drastically reduce the
O(w2 ) complexity of the self-attention layers with respect to the time series length w, which is the
main performance bottleneck.
Imputation and forecasting: The model and training process described in Section 3.3 is exactly the
setup required to perform imputation of missing values, without any modifications, and we observed
that it was possible to achieve very good results following this method; as a rough indication, our
models could reach Root Mean Square Errors very close to 0 when asked to perform the input
denoising (autoregressive) task on the test set, after being subjected to unsupervised pre-training on
the training set. We also show example results of imputation on one of the datasets presented in this
work in Figure 5. However, we defer a systematic quantitative comparison with the state of the art to
future work. Furthermore, we note that one may simply use different patterns of masking to achieve
different objectives, while the rest of the model and setup remain the same. For example, using a
mask which conceals the last part of all variables simultaneously, one may perform forecasting (see
Figure 4 in Appendix), while for longer time series one may additionally perform this process within
a sliding window. Again, we defer a systematic investigation to future work.
Extracted representations: The representations zt extracted by the transformer models can be
used directly for evaluating similarity between time series, clustering, visualization and any other
use cases where time series representations are used in practice. A valuable benefit offered by
transformers is that representations can be independently addressed for each time step; this means
that, for example, a greater weight can be placed at the beginning, middle or end of the time series,
which allows to selectively compare time series, visualize temporal evolution of samples etc.

Dataset Train Size Test Size Length Dimension Missing Values

AppliancesEnergy 96 42 144 24 No
BenzeneConcentration 3433 5445 240 8 Yes
BeijingPM10Quality 12432 5100 24 9 Yes
BeijingPM25Quality 12432 5100 24 9 Yes
LiveFuelMoistureContent 3493 1510 365 7 No
IEEEPPG 1768 1328 1000 5 No

Table 3: Multivariate Regression Datasets

Dataset TrainSize TestSize NumDimensions SeriesLength NumClasses

EthanolConcentration 261 263 3 1751 4
FaceDetection 5890 3524 144 62 2
Handwriting 150 850 3 152 26
Heartbeat 204 205 61 405 2
InsectWingbeat 30000 20000 200 30 10
JapaneseVowels 270 370 12 29 9
PEMS-SF 267 173 963 144 7
SelfRegulationSCP1 268 293 6 896 2
SelfRegulationSCP2 200 180 7 1152 2
SpokenArabicDigits 6599 2199 13 93 10
UWaveGestureLibrary 120 320 3 315 8

Table 4: Multivariate Classification Datasets

12
Published as a conference paper at ICLR 2021

Standard deviation
Dataset
Supervised TST Pre-trained TST
AppliancesEnergy 0.240 0.163
BenzeneConcentration 0.031 0.092
BeijingPM10Quality 0.689 0.813
BeijingPM25Quality 0.189 0.253
LiveFuelMoistureContent 0.735 0.013
IEEEPPG 1.079 1.607

Table 5: Standard deviation of the Root Mean Square Error displayed by the Time Series Trans-
former models on multivariate regression datasets

A.2 C RITERIA FOR DATASET SELECTION

We select a diverge range of datasets from the Monash University, UEA, UCR Time Series Regres-
sion and Classification Archives, in a way so as to ensure diversity with respect to the dimensionality
and length of time series samples, as well as the number of samples. Additionally, we have tried to
include both ”easy” and ”difficult” datasets (where the baselines perform very well or less well). In
the following we provide a more detailed rationale for each of the selected multivariate datasets.
EthanolConcentration: very low dimensional, very few samples, moderate length, large number
of classes, challenging
FaceDetection: very high dimensional, many samples, very short length, minimum number of
classes
Handwriting: very low dimensional, very few samples, moderate length, large number of classes
Heartbeat: high dimensional, very few samples, moderate length, minimum number of classes
JapaneseVowels: very heterogeneous sample length, moderate num. dimensions, very few samples,
very short length, moderate number of classes, all baselines perform well
InsectWingBeat: very high dimensional, many samples, very short length, moderate number of
classes, very challenging
PEMS-SF: extremely high dimensional, very few samples, moderate length, moderate number of
classes
SelfRegulationSCP1: Few dimensions, very few samples, long length, minimum number of classes,
baselines perform well
SelfRegulationSCP2: similar to SelfRegulationSCP1, but challenging
SpokenArabicDigits: Moderate number of dimensions, many samples, very heterogeneous length,
moderate number of classes, most baselines perform well
UWaveGestureLibrary: very low dimensional, very few samples, moderate length, moderate num-
ber of classes, baselines perform well

A.3 E XECUTION TIME

We recorded the times required for training our fully supervised models until convergence on a
GPU, as well as for the currently fastest and top performing (in terms of classification accuracy and
regression error) baseline methods, ROCKET and XGBoost on a CPU. These have been shown to be
orders of magnitude faster than methods such as TS-CHIEF, Proximity Forest, Elastic Ensembles,
DTW and HIVE-COTE, but also deep learning based methods Dempster et al. (2020). Although
XGBoost and ROCKET are incomparably faster than the transformer on a CPU, as can be seen in
Table 7 in the Appendix, exploiting commercial GPUs and the parallel processing capabilities of a
transformer typically enables as fast (and sometimes faster) training times as these (currently fastest
available) methods. In practice, despite allowing for many hundreds of epochs, using a GPU we
never trained our models longer than 3 hours on any of the examined datasets.

13
 Published as a conference paper at ICLR 2021

Standard deviation
Dataset
Supervised TST Pre-trained TST
EthanolConcentration 0.024 0.002
FaceDetection 0.007 0.006
Handwriting 0.020 0.006
Heartbeat 0.018 0.018
InsectWingbeat 0.003 0.026
JapaneseVowels 0.000 0.0016
PEMS-SF 0.017 0.003
SelfRegulationSCP1 0.005 0.006
SelfRegulationSCP2 0.020 0.003
SpokenArabicDigits 0.0003 0.001
UWaveGestureLibrary 0.005 0.003

Table 6: Standard deviation of accuracy displayed by the Time Series Transformer models on mul-
tivariate classification datasets

Dataset Rocket XGBoost TST (GPU)

EthanolConcentration 41.937 3.760 34.72
FaceDetection 279.033 57.832 67.8
Handwriting 6.705 1.836 134.4
Heartbeat 35.825 3.013 2.57
InsectWingBeat - 64.883 4565
JapaneseVowels 5.032 0.527 4.71
PEMS-SF 369.198 150.879 341
SelfRegulationSCP1 30.578 0.967 3.46
SelfRegulationSCP2 28.286 1.213 97.3
SpokenArabicDigits 65.143 3.129 73.2
UWaveGestureLibrary 3.078 0.636 2.90

Table 7: Total training time (time until maximum accuracy is recorded) in seconds: for the fastest
currently available methods (Rocket, XGBoost) on the same CPU, as well as for our fully supervised
transformer models on a GPU. On a CPU, training for our model is typically at least an order of
magnitude slower.

As regards deep learning models, LSTMs are well known to be slow, as they require O(w) sequential
operations (where w is the length of the time series) for each sample, with the complexity per layer
scaling as O(N · d2 ), where d is the internal representation dimension (hidden state size). We refer
the reader to the original transformer paper (Vaswani et al., 2017) for a detailed discussion about
how tranformers compare to Convolutional Neural Networks in terms of computational efficiency.

Ranks
Random 1-NN 5-NN 1-NN- 5-NN- Rel diff from
Dataset Name TST SVR XGBoost Rocket FCN ResNet Inception
Forest -ED -ED DTWD DTWD 2nd best
AppliancesEnergy 1 6 5 7 11 9 12 8 2 3 4 10 -0.361
BenzeneConcentration 1 7 3 2 12 11 9 8 5 10 6 4 -0.225
BeijingPM10Quality 1 7 3 2 12 9 11 8 10 4 5 6 -0.067
BeijingPM25Quality 1 10 6 2 11 9 12 8 5 3 7 4 -0.101
LiveFuelMoistureContent 2 3 5 4 12 7 11 6 1 8 10 9 -0.038
IEEEPPG 2 10 5 4 7 3 12 8 11 9 6 1

Table 8: Relative ranks of all methods on Regression datasets. Last column indicates relative diff of
our method from the 2nd best, whenever the rank is 1.

14
 Published as a conference paper at ICLR 2021

Problem TST Rocket XGBoost LSTM Franseschi et al DTW D Rel. Diff from 2nd best
EthanolConcentration 4 1 2 3 6 5
FaceDetection 1 2 3 4 6 5 -0.041
Handwriting 3 1 5 6 2 4
Heartbeat 1 2 4 5 3 6 -0.02
JapaneseVowels 1 3 5 6 2 4 -0.008
PEMS-SF 2 3 1 6 5 4
SelfRegulationSCP1 1 2 3 6 4 5 -0.017
SelfRegulationSCP2 1 4 5 6 2 3 -0.055
SpokenArabicDigits 1 4 5 6 3 2 -0.035
UWaveGestureLibrary 2 1 5 6 4 3

Table 9: Relative ranks of methods on the Classification datasets. The last column is the relative diff
of our method from the 2nd best, whenever we rank 1st.

Dataset Task (Metric) Sep., Bern. Sync., Bern. Sep., Stateful Sync., Stateful
Heartbeat Classif. (Accuracy) 0.761 0.756 0.776 0.751
InsectWingbeat Classif. (Accuracy) 0.641 0.632 0.687 0.689
SpokenArabicDigits Classif. (Accuracy) 0.994 0.994 0.998 0.996
PEMS-SF Classif. (Accuracy) 0.873 0.879 0.896 0.879
BenzeneConcentration Regress. (RMSE) 0.681 0.493 0.494 0.684
BeijingPM25Quality Regress. (RMSE) 57.241 59.529 53.492 59.632
LiveFuelMoistureContent Regress. (RMSE) 44.398 43.519 43.138 43.420

Table 10: Comparison of four different input value masking schemes evaluated for unsupervised
learning on 4 classification and 3 regression datasets. Two of the variants involve separately gen-
erating the mask for each variable, and two involve a single distribution over “time steps”, applied
synchronously to all variables. Also, two of the variants involve sampling each “time step” inde-
pendently based on a Bernoulli distribution with parameter p = r = 15%, while the remaining two
involve using a Markov chain with two states, “masked” or “unmasked”, with different transition
1 r
probabilities pm = lm and pu = pm 1−r , such that the masked sequences follow a geometric dis-
tribution with a mean length of lm = 3 and each variable is masked on average by r = 15%. We
observe that our proposed scheme, separately masking each variable through stateful generation,
performs consistently well and shows the overall best performance across all examined datasets.

Dataset Task (Metric) LayerNorm BatchNorm

Heartbeat Classif. (Accuracy) 0.741 0.776
InsectWingbeat Classif. (Accuracy) 0.658 0.684
SpokenArabicDigits Classif. (Accuracy) 0.993 0.993
PEMS-SF Classif. (Accuracy) 0.832 0.919
BenzeneConcentration Regress. (RMSE) 2.053 0.516
BeijingPM25Quality Regress. (RMSE) 61.082 60.357
LiveFuelMoistureContent Regress. (RMSE) 42.993 42.607

Table 11: Performance comparison between using layer normalization and batch normalization in
our supervised transformer model. The batch size is 128.

15
Published as a conference paper at ICLR 2021

70 68

68 66
66 64
64
62
62
60

RMSE
RMSE

60
58
58
56
56
54 54

52 52

50 50
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Use of data for unsupervised pre-training Use of data for unsupervised pre-training

Figure 3: Root Mean Squared Error of a given model as a function of the number of samples (here,
shown as a proportion of the total number of samples in the training set) used for unsupervised pre-
training. Two levels of label availability (used for supervised learning) are depicted: 10% (left panel)
and 20% (right panel) of all training data labels. Note that a horizontal axis value of 0 means fully
supervised learning only, while all other values correspond to unsupervised pre-training followed by
supervised fine-tuning.

Static Fine-tuned
Dataset Task (Metric)
Metric Epoch time (s) Metric Epoch time (s)
Heartbeat Classif. (Accuracy) 0.756 0.082 0.776 0.14
InsectWingbeat Classif. (Accuracy) 0.236 4.52 0.687 6.21
SpokenArabicDigits Classif. (Accuracy) 0.996 1.29 0.998 2.00
PEMS-SF Classif. (Accuracy) 0.844 0.208 0.896 0.281
BenzeneConcentration Regress. (RMSE) 4.684 0.697 0.494 1.101
BeijingPM25Quality Regress. (RMSE) 65.608 1.91 53.492 2.68
LiveFuelMoistureContent Regress. (RMSE) 48.724 1.696 43.138 3.57

Table 12: Performance comparison between allowing all layers of a pre-trained transformer to be
fine-tuned, versus using static (“extracted”) representations of the time series as input to the output
layer (which is equivalent to freezing all model layers except for the output layer). The per-epoch
training time on a GPU is also shown.

forecasting synchronous noise

xm xm

... ...

x2 x2
x1 x1
t1 t2 t3 tw t1 t2 t3 tw
𝒙1 𝒙𝑤 𝒙1 𝒙𝑤

Figure 4: Masking schemes within our transformer encoder framework: for implementation of fore-
casting objective (left), for an alternative unsupervised learning objective involving a single noise
distribution over time steps, applied synchronously to all variables (right).

16
 Published as a conference paper at ICLR 2021

True, input
True, masked
2 Prediction
Concentration [mg/m3]

0 50 100 150 200 250

time (hours)

Dimensions
2 4
2 2
1 2
2
0
0 0 0
1 0

0 50 100 150 200 250 0 50 100 150 200 250 0 50 100 150 200 250 0 50 100 150 200 250 0 50 100 150 200 250
4
2 1 4
Samples

2
1 0 2 2
1
0 0 1 0 0
1 1
2
0 50 100 150 200 250 0 50 100 150 200 250 0 50 100 150 200 250 0 50 100 150 200 250 0 50 100 150 200 250
4 2
2
2 2 1
2
0 0 0
0 0 1
2
2 2 2
0 50 100 150 200 250 0 50 100 150 200 250 0 50 100 150 200 250 0 50 100 150 200 250 0 50 100 150 200 250
4
4
2 2 2 2
2
0 0 0 0
0
2 2 2 2
0 50 100 150 200 250 0 50 100 150 200 250 0 50 100 150 200 250 0 50 100 150 200 250 0 50 100 150 200 250

Figure 5: Top: Imputation of missing values in the test set of BenzeneConcentration dataset. The
continuous blue line is the ground truth signal, the light blue circles indicate the values hidden
from the model and the orange dots its prediction. We observe that imputed values approximate
true values very well, even in cases of rapid transitions and in cases where many contiguous values
are missing. Bottom: Same, shown for 5 different dimensions of the time series (here, these are
concentrations of different substances) as columns and 4 different, randomly selected samples as
rows.

17
Published as a conference paper at ICLR 2021

A.4 A DVANTAGES OF TRANSFORMERS

Transformer models are based on a multi-headed attention mechanism that offers several key advan-
tages and renders them particularly suitable for time series data:

• They can concurrently take into account long contexts of input sequence elements and
learn to represent each sequence element by selectively attending to those input sequence
elements which the model considers most relevant. They do so without position-dependent
prior bias; this is to be contrasted with RNN-based models: a) even bi-directional RNNs
treat elements in the middle of the input sequence differently from elements close to the two
endpoints, and b) despite careful design, even LSTM (Long Short Term Memory) and GRU
(Gated Recurrent Unit) networks practically only retain information from a limited number
of time steps stored inside their hidden state (vanishing gradient problem (Hochreiter, 1998;
Pascanu et al., 2013)), and thus the context used for representing each sequence element is
inevitably local.
• Multiple attention heads can consider different representation subspaces, i.e., multiple as-
pects of relevance between input elements. For example, in the context of a signal with
two frequency components, 1/T1 and 1/T2 , one attention head can attend to neighboring
time points, while another one may attend to points spaced a period T1 before the currently
examined time point, a third to a period T2 before, etc. This is to be contrasted with at-
tention mechanisms in RNN models, which learn a single global aspect/mode of relevance
between sequence elements.
• After each stage of contextual representation (i.e., transformer encoder layer), attention is
redistributed over the sequence elements, taking into account progressively more abstract
representations of the input elements as information flows from the input towards the out-
put. By contrast, RNN models with attention use a single distribution of attention weights
to extract a representation of the input, and most typically attend over a single layer of
representation (hidden states).

A.5 H YPERPARAMETERS

Parameter Value
activation gelu
dropout 0.1
learning rate 0.001
pos. encoding learnable

Table 13: Common (fixed) hyperparameters used for all transformer models.

18
Published as a conference paper at ICLR 2021

Parameter Value
dim. model 128
dim. feedforward 256
num. heads 16
num. encoder blocks 3
batch size 128

Table 14: Hyperparameter configuration that performs reasonably well for all transformer models.

Dataset num. blocks num. heads dim. model dim. feedforward

AppliancesEnergy 3 8 128 512
BenzeneConcentration 3 8 128 256
BeijingPM10Quality 3 8 64 256
BeijingPM25Quality 3 8 64 (128) 256
LiveFuelMoistureContent 3 8 64 256
IEEEPPG 3 8 512 512

Table 15: Supervised TST model hyperparameters for the multivariate regression datasets

Dataset num. blocks num. heads dim. model dim. feedforward

AppliancesEnergy 3 16 128 512
BenzeneConcentration 1 8 128 256
BeijingPM10Quality 3 8 64 256
BeijingPM25Quality 3 8 128 256
LiveFuelMoistureContent 3 8 64 256
IEEEPPG 4 16 512 512

Table 16: Unsupervised TST model hyperparameters for the multivariate regression datasets

Dataset num. blocks num. heads dim. model dim. feedforward

EthanolConcentration 1 8 64 256
FaceDetection 3 8 128 256
Handwriting 1 8 128 256
Heartbeat 1 8 64 256
JapaneseVowels 3 8 128 256
PEMS-SF 1 8 128 512
SelfRegulationSCP1 3 8 128 256
SelfRegulationSCP2 3 8 128 256
SpokenArabicDigits 3 8 64 256
UWaveGestureLibrary 3 16 256 256

Table 17: Supervised TST model hyperparameters for the multivariate classification datasets

Dataset num. blocks num. heads dim. model dim. feedforward

EthanolConcentration 1 8 64 256
FaceDetection 3 8 128 256
Handwriting 3 16 64 256
Heartbeat 1 8 64 256
JapaneseVowels 3 8 128 256
PEMS-SF 1 8 256 512
SelfRegulationSCP1 3 16 256 512
SelfRegulationSCP2 3 8 256 512
SpokenArabicDigits 3 8 64 256
UWaveGestureLibrary 3 16 256 512

Table 18: Unsupervised TST model hyperparameters for the multivariate classification datasets

19