Frequency-domain MLPs are More Effective

Learners in Time Series Forecasting

Kun Yi1 , Qi Zhang2 , Wei Fan3 , Shoujin Wang4 , Pengyang Wang5 , Hui He1
Defu Lian6 , Ning An7 , Longbing Cao8 , Zhendong Niu1∗
1
Beijing Institute of Technology, 2 Tongji University, 3 University of Oxford
arXiv:2311.06184v1 [cs.LG] 10 Nov 2023

4
University of Technology Sydney, 5 University of Macau, 6 USTC
7
HeFei University of Technology, 8 Macquarie University
{yikun, hehui617, zniu}@bit.edu.cn, [email protected], [email protected]
[email protected], [email protected], [email protected], [email protected]

Abstract
Time series forecasting has played the key role in different industrial, including
finance, traffic, energy, and healthcare domains. While existing literatures have de-
signed many sophisticated architectures based on RNNs, GNNs, or Transformers,
another kind of approaches based on multi-layer perceptrons (MLPs) are pro-
posed with simple structure, low complexity, and superior performance. However,
most MLP-based forecasting methods suffer from the point-wise mappings and
information bottleneck, which largely hinders the forecasting performance. To
overcome this problem, we explore a novel direction of applying MLPs in the
frequency domain for time series forecasting. We investigate the learned patterns of
frequency-domain MLPs and discover their two inherent characteristic benefiting
forecasting, (i) global view: frequency spectrum makes MLPs own a complete view
for signals and learn global dependencies more easily, and (ii) energy compaction:
frequency-domain MLPs concentrate on smaller key part of frequency components
with compact signal energy. Then, we propose FreTS, a simple yet effective archi-
tecture built upon Frequency-domain MLPs for Time Series forecasting. FreTS
mainly involves two stages, (i) Domain Conversion, that transforms time-domain
signals into complex numbers of frequency domain; (ii) Frequency Learning, that
performs our redesigned MLPs for the learning of real and imaginary part of fre-
quency components. The above stages operated on both inter-series and intra-series
scales further contribute to channel-wise and time-wise dependency learning. Ex-
tensive experiments on 13 real-world benchmarks (including 7 benchmarks for
short-term forecasting and 6 benchmarks for long-term forecasting) demonstrate
our consistent superiority over state-of-the-art methods. Code is available at this
repository: https://github.com/aikunyi/FreTS.

1 Introduction
Time series forecasting has been a critical role in a variety of real-world industries, such as climate
condition estimation [1, 2, 3], traffic state prediction [4, 5, 6], economic analysis [7, 8], etc. In the
early stage, many traditional statistical forecasting methods have been proposed, such as exponential
smoothing [9] and auto-regressive moving averages (ARMA) [10]. Recently, the emerging devel-
opment of deep learning has fostered many deep forecasting models, including Recurrent Neural
Network-based methods (e.g., DeepAR [11], LSTNet [12]), Convolution Neural Network-based meth-
ods (e.g., TCN [13], SCINet [14]), Transformer-based methods (e.g., Informer [15], Autoformer [16]),
and Graph Neural Network-based methods (e.g., MTGNN [17], StemGNN [18], AGCRN [19]), etc.
∗
Corresponding author

37th Conference on Neural Information Processing Systems (NeurIPS 2023).

 (a) Left: time domain. Right: frequency domain. (b) Left: time domain. Right: frequency domain.
Figure 1: Visualizations of the learned patterns of MLPs in the time domain and the frequency
domain (see Appendix B.4). (a) global view: the patterns learned in the frequency domain exhibits
more obvious global periodic patterns than the time domain; (b) energy compaction: learning in the
frequency domain can identify clearer diagonal dependencies and key patterns than the time domain.

While these deep models have achieved promising forecasting performance in certain scenarios,
their sophisticated network architectures would usually bring up expensive computation burden
in training or inference stage. Besides, the robustness of these models could be easily influenced
with a large amount of parameters, especially when the available training data is limited [15, 20].
Therefore, the methods based on multi-layer perceptrons (MLPs) have been recently introduced with
simple structure, low complexity, and superior forecasting performance, such as N-BEATS [21],
LightTS [22], DLinear [23], etc. However, these MLP-based methods rely on point-wise mappings
to capture temporal mappings, which cannot handle global dependencies of time series. Moreover,
they would suffer from the information bottleneck with regard to the volatile and redundant local
momenta of time series, which largely hinders their performance for time series forecasting.
To overcome the above problems, we explore a novel direction of applying MLPs in the frequency
domain for time series forecasting. We investigate the learned patterns of frequency-domain MLPs
in forecasting and have discovered their two key advantages: (i) global view: operating on spectral
components acquired from series transformation, frequency-domain MLPs can capture a more
complete view of signals, making it easier to learn global spatial/temporal dependencies. (ii) energy
compaction: frequency-domain MLPs concentrate on the smaller key part of frequency components
with the compact signal energy, and thus can facilitate preserving clearer patterns while filtering
out influence of noises. Experimentally, we have observed that frequency-domain MLPs capture
much more obvious global periodic patterns than the time-domain MLPs from Figure 1(a), which
highlights their ability to recognize global signals. Also, from Figure 1(b), we easily note a much
more clear diagonal dependency in the learned weights of frequency-domain MLPs, compared with
the more scattered dependency learned by time-domain MLPs. This illustrates the great potential
of frequency-domain MLPs to identify most important features and key patterns while handling
complicated and noisy information.
To fully utilize these advantages, we propose FreTS, a simple yet effective architecture of Frequency-
domain MLPs for Time Series forecasting. The core idea of FreTS is to learn the time series
forecasting mappings in the frequency domain. Specifically, FreTS mainly involves two stages: (i)
Domain Conversion: the original time-domain series signals are first transformed into frequency-
domain spectrum on top of Discrete Fourier Transform (DFT) [24], where the spectrum is composed of
several complex numbers as frequency components, including the real coefficients and the imaginary
coefficients. (ii) Frequency Learning: given the real/imaginary coefficients, we redesign the frequency-
domain MLPs originally for the complex numbers by separately considering the real mappings and
imaginary mappings. The respective real/imaginary parts of output learned by two distinct MLPs are
then stacked in order to recover from frequency components to the final forecasting. Also, FreTS
performs above two stages on both inter-series and intra-series scales, which further contributes to the
channel-wise and time-wise dependencies in the frequency domain for better forecasting performance.
We conduct extensive experiments on 13 benchmarks under different settings, covering 7 benchmarks
for short-term forecasting and 6 benchmarks for long-term forecasting, which demonstrate our
consistent superiority compared with state-of-the-art methods.

2 Related Work
Forecasting in the Time Domain Traditionally, statistical methods have been proposed for forecast-
ing in the time domain, including (ARMA) [10], VAR [25], and ARIMA [26]. Recently, deep learning

2
based methods have been widely used in time series forecasting due to their capability of extracting
nonlinear and complex correlations [27, 28]. These methods have learned the dependencies in the
time domain with RNNs (e.g., deepAR [11], LSTNet [12]) and CNNs (e.g., TCN [13], SCINet [14]).
In addition, GNN-based models have been proposed with good forecasting performance because of
their good abilities to model series-wise dependencies among variables in the time domain, such as
TAMP-S2GCNets [5], AGCRN [19], MTGNN [17], and GraphWaveNet [29]. Besides, Transformer-
based forecasting methods have been introduced due to their attention mechanisms for long-range
dependency modeling ability in the time domain, such as Reformer [20] and Informer [15].

Forecasting in the Frequency Domain Several recent time series forecasting methods have ex-
tracted knowledge of the frequency domain for forecasting [30]. Specifically, SFM [31] decomposes
the hidden state of LSTM into frequencies by Discrete Fourier Transform (DFT). StemGNN [18]
performs graph convolutions based on Graph Fourier Transform (GFT) and computes series corre-
lations based on Discrete Fourier Transform. Autoformer [16] replaces self-attention by proposing
the auto-correlation mechanism implemented with Fast Fourier Transforms (FFT). FEDformer [32]
proposes a DFT-based frequency enhanced attention, which obtains the attentive weights by the
spectrums of queries and keys, and calculates the weighted sum in the frequency domain. CoST [33]
uses DFT to map the intermediate features to frequency domain to enables interactions in representa-
tion. FiLM [34] utilizes Fourier analysis to preserve historical information and remove noisy signals.
Unlike these efforts that leverage frequency techniques to improve upon the original architecture such
as Transformer and GNN, in this paper, we propose a new frequency learning architecture that learns
both channel-wise and time-wise dependencies in the frequency domain.

MLP-based Forecasting Models Several studies have explored the use of MLP-based networks in
time series forecasting. N-BEATS [21] utilizes stacked MLP layers together with doubly residual
learning to process the input data to iteratively forecast the future. DEPTS [35] applies Fourier
transform to extract periods and MLPs for periodicity dependencies for univariate forecasting.
LightTS [22] uses lightweight sampling-oriented MLP structures to reduce complexity and com-
putation time while maintaining accuracy. N-HiTS [36] combines multi-rate input sampling and
hierarchical interpolation with MLPs for univariate forecasting. LTSF-Linear [37] proposes a set
of embarrassingly simple one-layer linear model to learn temporal relationships between input and
output sequences. These studies demonstrate the effectiveness of MLP-based networks in time series
forecasting tasks, and inspire the development of our frequency-domain MLPs in this paper.

3 FreTS
In this section, we elaborate on our proposed novel approach, FreTS, based on our redesigned MLPs
in the frequency domain for time series forecasting. First, we present the detailed frequency learning
architecture of FreTS in Section 3.1, which mainly includes two-fold frequency learners with domain
conversions. Then, we detailedly introduce our redesigned frequency-domain MLPs adopted by
above frequency learners in Section 3.2. Besides, we also theoretically analyze their superior nature
of global view and energy compaction, as aforementioned in Section 1.

Problem Definition Let [X1 , X2 , · · · , XT ] ∈ RN ×T stand for the regularly sampled multi-variate
time series dataset with N series and T timestamps, where Xt ∈ RN denotes the multi-variate
values of N distinct series at timestamp t. We consider a time series lookback window of length-L
at timestamp t as the model input, namely Xt = [Xt−L+1 , Xt−L+2 , · · · , Xt ] ∈ RN ×L ; also, we
consider a horizon window of length-τ at timestamp t as the prediction target, denoted as Yt =
[Xt+1 , Xt+2 , · · · , Xt+τ ] ∈ RN ×τ . Then the time series forecasting formulation is to use historical
observations Xt to predict future values Ŷt and the typical forecasting model fθ parameterized by θ
is to produce forecasting results by Ŷt = fθ (Xt ).
3.1 Frequency Learning Architecture

The frequency learning architecture of FreTS is depicted in Figure 2, which mainly involves Domain
Conversion/Inversion stages, Frequency-domain MLPs, and the corresponding two learners, i.e.,
the Frequency Channel Learner and the Frequency Temporal Learner. Besides, before taken to
learners, we concretely apply a dimension extension block on model input to enhance the model
capability. Specifically, the input lookback window Xt ∈ RN ×L is multiplied with a learnable

3
 Frequency Channel Learner Frequency Temporal Learner
Nxd
L N
FreTS
NxLxd Nxd Lxd Lxd

Frequency-domain

Frequency-domain
NxLxd NxLxd

Dimension
Extension

Conversion

Conversion
Inversion

Inversion
MLP
Domain

Domain

Domain

MLP

Domain
Concat

Concat

FFN
Input 𝓗 𝓏chan 𝓏temp 𝓢 Output
NxL 𝐗 0 0 𝐘෠ Nxτ
𝐇
𝐙 𝐒

L
Shared with L timestamps Shared with N channels
𝓗𝓦+𝓑

Real Part
N chan chan temp temp N
N 𝓦 𝓑 𝓦 𝓑
𝓦 Complex number weight 𝓗𝓦
L L τ
L

Amplitude
𝓑 Complex number bias
𝓑
Amplitude

time dimensions
N Complex Number Imaginary Part
Real Part
Channel dimensions Imaginary Part Operations in the frequency domain
Frequency Frequency

Figure 2: The framework overview of FreTS: the Frequency Channel Learner focuses on modeling
inter-series dependencies with frequency-domain MLPs operating on the channel dimensions; the
Frequency Temporal Learner is to capture the temporal dependencies by performing frequency-
domain MLPs on the time dimensions.

weight vector ϕd ∈ R1×d to obtain a more expressive hidden representation Ht ∈ RN ×L×d , yielding
Ht = Xt × ϕd to bring more semantic information, inspired by word embeddings [38].

Domain Conversion/Inversion The use of Fourier transform enables the decomposition of a

time series signal into its constituent frequencies. This is particularly advantageous for time series
analysis since it benefits to identify periodic or trend patterns in the data, which are often important in
forecasting tasks. As aforementioned in Figure 1(a), learning in the frequency spectrum helps capture
a greater number of periodic patterns. In view of this, we convert the input H into the frequency
domain H by:
Z ∞ Z ∞ Z ∞
H(f ) = H(v)e−j2πf v dv = H(v) cos(2πf v)dv + j H(v) sin(2πf v)dv (1)
−∞ −∞ −∞
where f is the frequency variable,R v is the integral variable, and j is the imaginary unit, which is
∞
defined as the square root of -1; −∞ H(v) cos(2πf v)dv is the real part of H and is abbreviated
R∞
as Re(H); −∞ H(v) sin(2πf v)dv is the imaginary part and is abbreviated as Im(H). Then we
can rewrite H in Equation (1) as: H = Re(H) + jIm(H). Note that in FreTS we operate domain
conversion on both the channel dimension and time dimension, respectively. Once completing the
learning in the frequency domain, we can convert H back into the the time domain using the following
inverse conversion formulation:
Z ∞ Z ∞
j2πf v
H(v) = H(f )e df = (Re(H(f )) + jIm(H(f ))ej2πf v df (2)
−∞ −∞
where we take frequency f as the integral variable. In fact, the frequency spectrum is expressed as a
combination of cos and sin waves in H with different frequencies and amplitudes inferring different
periodic properties in time series signals. Thus examining the frequency spectrum can better discern
the prominent frequencies and periodic patterns in time series. In the following sections, we use
DomainConversion to stand for Equation (1), and DomainInversion for Equation (2) for brevity.

Frequency Channel Learner Considering channel dependencies for time series forecasting is
important because it allows the model to capture interactions and correlations between different
variables, leading to a more accurate predictions. The frequency channel learner enables commu-
nications between different channels; it operates on each timestamp by sharing the same weights
between L timestamps to learn channel dependencies. Concretely, the frequency channel learner
:,(l)
takes Ht ∈ RN ×L×d as input. Given the l-th timestamp Ht ∈ RN ×d , we perform the frequency
channel learner by:
:,(l) :,(l)
Hchan = DomainConversion(chan) (Ht )
:,(l) :,(l)
Zchan = FreMLP(Hchan , W chan , B chan ) (3)
:,(l)
Z:,(l) = DomainInversion(chan) (Zchan )

4
 :,(l) N :,(l)
where Hchan ∈ C 2 ×d is the frequency components of Ht ; DomainConversion(chan) and
DomainInversion(chan) indicates such operations are performed along the channel dimension.
FreMLP are frequency-domain MLPs proposed in Section 3.2, which takes W chan = (Wrchan +
jWichan ) ∈ Cd×d as the complex number weight matrix with Wrchan ∈ Rd×d and Wichan ∈ Rd×d ,
and B chan = (Brchan + jBichan ) ∈ Cd as the biases with Brchan ∈ Rd and Bichan ∈ Rd . And
:,(l) N
Zchan ∈ C 2 ×d is the output of FreMLP, also in the frequency domain, which is conversed back to
time domain as Z:,(l) ∈ RN ×d . Finally, we ensemble Z:,(l) of L timestamps into a whole and output
Zt ∈ RN ×L×d .

Frequency Temporal Learner The frequency temporal learner aims to learn the temporal patterns
in the frequency domain; also, it is constructed based on frequency-domain MLPs conducting on
each channel and it shares the weights between N channels. Specifically, it takes the frequency
(n),:
channel learner output Zt ∈ RN ×L×d as input and for the n-th channel Zt ∈ RL×d , we apply the
frequency temporal learner by:
(n),: (n),:
Ztemp = DomainConversion(temp) (Zt )
(n),: (n),:
Stemp = FreMLP(Ztemp , W temp , B temp ) (4)
(n),: (n),:
S = DomainInversion(temp) (Stemp )
(n),: L (n),:
where Ztemp ∈ C 2 ×d
is the corresponding frequency spectrum of Zt ; DomainConversion(temp)
and DomainInversion(temp) indicates the calculations are applied along the time dimension.
W temp = (Wrtemp + jWitemp ) ∈ Cd×d is the complex number weight matrix with Wrtemp ∈ Rd×d
and Witemp ∈ Rd×d , and B temp = (Brtemp + jBitemp ) ∈ Cd are the complex number biases with
(n),: L
Brtemp ∈ Rd and Bitemp ∈ Rd . Stemp ∈ C 2 ×d is the output of FreMLP and is converted back to the
time domain as S(n),: ∈ RL×d . Finally, we incorporate all channels and output St ∈ RN ×L×d .

Projection Finally, we use the learned channel and temporal dependencies to make predictions for
the future τ timestamps Ŷt ∈ RN ×τ by a two-layer feed forward network (FFN) with one forward
step which can avoid error accumulation, formulated as follows:
Ŷt = σ(St ϕ1 + b1 )ϕ2 + b2 (5)
N ×L×d
where St ∈ R is the output of the frequency temporal learner, σ is the activation function,
ϕ1 ∈ R(L∗d)×dh , ϕ2 ∈ Rdh ×τ are the weights, b1 ∈ Rdh , b2 ∈ Rτ are the biases, and dh is the
inner-layer dimension size.

3.2 Frequency-domain MLPs

As shown in Figure 3, we elaborate our novel frequency-domain MLPs in FreTS that are redesigned
for the complex numbers of frequency components, in order to effectively capture the time series key
patterns with global view and energy compaction, as aforementioned in Section 1.
Definition 1 (Frequency-domain MLPs). Formally, for a complex number input H ∈ Cm×d ,
given a complex number weight matrix W ∈ Cd×d and a complex number bias B ∈ Cd , then the
frequency-domain MLPs can be formulated as:
Y ℓ = σ(Y ℓ−1 W ℓ + B ℓ )
(6)
Y0 = H
where Y ℓ ∈ Cm×d is the final output, ℓ denotes the ℓ-th layer, and σ is the activation function.
As both H and W are complex numbers, according to the rule of multiplication of complex numbers
(details can be seen in Appendix C), we further extend the Equation (6) to:
Y ℓ = σ(Re(Y ℓ−1 )Wrℓ − Im(Y ℓ−1 )Wiℓ + Brℓ ) + jσ(Re(Y ℓ−1 )Wiℓ + Im(Y ℓ−1 )Wrℓ + Biℓ ) (7)
where W ℓ = Wrℓ + jWiℓ and B ℓ = Brℓ + jBiℓ . According to the equation, we implement the MLPs
in the frequency domain (abbreviated as FreMLP) by the separate computation of the real and
imaginary parts of frequency components. Then, we stack them to form a complex number to acquire
the final results. The specific implementation process is shown in Figure 3.

5
Theorem 1. Suppose that H is the representation of raw time series and H is the corresponding
frequency components of the spectrum, then the energy of a time series in the time domain is equal to
the energy of its representation in the frequency domain. Formally, we can express this with above
notations by: Z ∞ Z ∞
|H(v)|2 dv = |H(f )|2 df (8)
−∞ −∞
R∞
where H(f ) = −∞
H(v)e−j2πf v dv, v is the time/channel dimension, f is the frequency dimension.

We include the proof in Appendix D.1. The theorem implies

that if most of the energy of a time series is concentrated in a
small number of frequency components, then the time series
can be accurately represented using only those components.
Accordingly, discarding the others would not significantly
affect the signal’s energy. As shown in Figure 1(b), in the
frequency domain, the energy concentrates on the smaller
part of frequency components, thus learning in the frequency
spectrum can facilitate preserving clearer patterns.
Theorem 2. Given the time series input H and its corre-
sponding frequency domain conversion H, the operations of
frequency-domain MLP on H can be represented as global Figure 3: One layer of the frequency-
convolutions on H in the time domain. This can be given by: domain MLPs.
HW + B = F(H ∗ W + B) (9)
where ∗ is a circular convolution, W and B are the complex number weight and bias, W and B are
the weight and bias in the time domain, and F is DFT.

The proof is shown in Appendix D.2. Therefore, the operations of FreMLP, i.e., HW + B, are
equal to the operations (H ∗ W + B) in the time domain. This implies that the operations of
frequency-domain MLPs can be viewed as global convolutions in the time domain.

4 Experiments
To evaluate the performance of FreTS, we conduct extensive experiments on thirteen real-world time
series benchmarks, covering short-term forecasting and long-term forecasting settings to compare
with corresponding state-of-the-art methods.

Datasets Our empirical results are performed on various domains of datasets, including traffic,
energy, web, traffic, electrocardiogram, and healthcare, etc. Specifically, for the task of short-term
forecasting, we adopt Solar 2 , Wiki [39], Traffic [39], Electricity 3 , ECG [18], METR-LA [40],
and COVID-19 [5] datasets, following previous forecasting literature [18]. For the task of long-
term forecasting, we adopt Weather [16], Exchange [12], Traffic [16], Electricity [16], and ETT
datasets [15], following previous long time series forecasting works [15, 16, 32, 41]. We preprocess
all datasets following [18, 15, 16] and normalize them with the min-max normalization. We split the
datasets into training, validation, and test sets by the ratio of 7:2:1 except for the COVID-19 datasets
with 6:2:2. More dataset details are in Appendix B.1.

Baselines We compare our FreTS with the representative and state-of-the-art models for both short-
term and long-term forecasting to evaluate their effectiveness. For short-term forecasting, we compre
FreTS against VAR [25], SFM [31], LSTNet [12], TCN [13], GraphWaveNet [29], DeepGLO [39],
StemGNN [18], MTGNN [17], and AGCRN [19] for comparison. We also include TAMP-S2GCNets
[5], DCRNN [40] and STGCN [42], which require pre-defined graph structures, for comparison. For
long-term forecasting, we include Informer [15], Autoformer [16], Reformer [20], FEDformer [32],
LTSF-Linear [37], and the more recent PatchTST [41] for comparison. Additional details about the
baselines can be found in Appendix B.2.
2
https://www.nrel.gov/grid/solar-power-data.html
3
https://archive.ics.uci.edu/ml/datasets/ElectricityLoadDiagrams20112014

6
Implementation Details Our model is implemented with Pytorch 1.8 [43], and all experiments
are conducted on a single NVIDIA RTX 3080 10GB GPU. We take MSE (Mean Squared Error) as
the loss function and report MAE (Mean Absolute Errors) and RMSE (Root Mean Squared Errors)
results as the evaluation metrics. For additional implementation details, please refer to Appendix B.3.

4.1 Main Results

Table 1: Short-term forecasting comparison. The best results are in bold, and the second best results
are underlined. Full benchmarks of short-term forecasting are in Appendix F.1.
Models Solar Wiki Traffic ECG Electricity COVID-19
MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE
VAR 0.184 0.234 0.052 0.094 0.535 1.133 0.120 0.170 0.101 0.163 0.226 0.326
SFM 0.161 0.283 0.081 0.156 0.029 0.044 0.095 0.135 0.086 0.129 0.205 0.308
LSTNet 0.148 0.200 0.054 0.090 0.026 0.057 0.079 0.115 0.075 0.138 0.248 0.305
TCN 0.176 0.222 0.094 0.142 0.052 0.067 0.078 0.107 0.057 0.083 0.317 0.354
DeepGLO 0.178 0.400 0.110 0.113 0.025 0.037 0.110 0.163 0.090 0.131 0.169 0.253
Reformer 0.234 0.292 0.047 0.083 0.029 0.042 0.062 0.090 0.078 0.129 0.152 0.209
Informer 0.151 0.199 0.051 0.086 0.020 0.033 0.056 0.085 0.074 0.123 0.200 0.259
Autoformer 0.150 0.193 0.069 0.103 0.029 0.043 0.055 0.081 0.056 0.083 0.159 0.211
FEDformer 0.139 0.182 0.068 0.098 0.025 0.038 0.055 0.080 0.055 0.081 0.160 0.219
GraphWaveNet 0.183 0.238 0.061 0.105 0.013 0.034 0.093 0.142 0.094 0.140 0.201 0.255
StemGNN 0.176 0.222 0.190 0.255 0.080 0.135 0.100 0.130 0.070 0.101 0.421 0.508
MTGNN 0.151 0.207 0.101 0.140 0.013 0.030 0.090 0.139 0.077 0.113 0.394 0.488
AGCRN 0.123 0.214 0.044 0.079 0.084 0.166 0.055 0.080 0.074 0.116 0.254 0.309
FreTS (Ours) 0.120 0.162 0.041 0.074 0.011 0.023 0.053 0.078 0.050 0.076 0.123 0.167

Table 2: Long-term forecasting comparison. We set the lookback window size L as 96 and the
prediction length as τ ∈ {96, 192, 336, 720} except for traffic dataset whose prediction length is
set as τ ∈ {48, 96, 192, 336}. The best results are in bold and the second best are underlined. Full
results of long-term forecasting are included in Appendix F.2.
Models FreTS PatchTST LTSF-Linear FEDformer Autoformer Informer Reformer
Metrics MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE
96 0.032 0.071 0.034 0.074 0.040 0.081 0.050 0.088 0.064 0.104 0.101 0.139 0.108 0.152
Weather

192 0.040 0.081 0.042 0.084 0.048 0.089 0.051 0.092 0.061 0.103 0.097 0.134 0.147 0.201
336 0.046 0.090 0.049 0.094 0.056 0.098 0.057 0.100 0.059 0.101 0.115 0.155 0.154 0.203
720 0.055 0.099 0.056 0.102 0.065 0.106 0.064 0.109 0.065 0.110 0.132 0.175 0.173 0.228
Exchange

96 0.037 0.051 0.039 0.052 0.038 0.052 0.050 0.067 0.050 0.066 0.066 0.084 0.126 0.146
192 0.050 0.067 0.055 0.074 0.053 0.069 0.064 0.082 0.063 0.083 0.068 0.088 0.147 0.169
336 0.062 0.082 0.071 0.093 0.064 0.085 0.080 0.105 0.075 0.101 0.093 0.127 0.157 0.189
720 0.088 0.110 0.132 0.166 0.092 0.116 0.151 0.183 0.150 0.181 0.117 0.170 0.166 0.201
48 0.018 0.036 0.016 0.032 0.020 0.039 0.022 0.036 0.026 0.042 0.023 0.039 0.035 0.053
Traffic

96 0.020 0.038 0.018 0.035 0.022 0.042 0.023 0.044 0.033 0.050 0.030 0.047 0.035 0.054
192 0.019 0.038 0.020 0.039 0.020 0.040 0.022 0.042 0.035 0.053 0.034 0.053 0.035 0.054
336 0.020 0.039 0.021 0.040 0.021 0.041 0.021 0.040 0.032 0.050 0.035 0.054 0.035 0.055
Electricity

96 0.039 0.065 0.041 0.067 0.045 0.075 0.049 0.072 0.051 0.075 0.094 0.124 0.095 0.125
192 0.040 0.064 0.042 0.066 0.043 0.070 0.049 0.072 0.072 0.099 0.105 0.138 0.121 0.152
336 0.046 0.072 0.043 0.067 0.044 0.071 0.051 0.075 0.084 0.115 0.112 0.144 0.122 0.152
720 0.052 0.079 0.055 0.081 0.054 0.080 0.055 0.077 0.088 0.119 0.116 0.148 0.120 0.151
96 0.061 0.087 0.065 0.091 0.063 0.089 0.072 0.096 0.079 0.105 0.093 0.121 0.113 0.143
ETTh1

192 0.065 0.091 0.069 0.094 0.067 0.094 0.076 0.100 0.086 0.114 0.103 0.137 0.120 0.148
336 0.070 0.096 0.073 0.099 0.070 0.097 0.080 0.105 0.088 0.119 0.112 0.145 0.124 0.155
720 0.082 0.108 0.087 0.113 0.082 0.108 0.090 0.116 0.102 0.136 0.125 0.157 0.126 0.155
96 0.052 0.077 0.055 0.082 0.055 0.080 0.063 0.087 0.081 0.109 0.070 0.096 0.065 0.089
ETTm1

192 0.057 0.083 0.059 0.085 0.060 0.087 0.068 0.093 0.083 0.112 0.082 0.107 0.081 0.108
336 0.062 0.089 0.064 0.091 0.065 0.093 0.075 0.102 0.091 0.125 0.090 0.119 0.100 0.128
720 0.069 0.096 0.070 0.097 0.072 0.099 0.081 0.108 0.093 0.126 0.115 0.149 0.132 0.163

Short-Term Time Series Forecasting Table 1 presents the forecasting accuracy of our FreTS
compared to thirteen baselines on six datasets, with an input length of 12 and a prediction length
of 12. The best results are highlighted in bold and the second-best results are underlined. From the
table, we observe that FreTS outperforms all baselines on MAE and RMSE across all datasets, and
on average it makes improvement of 9.4% on MAE and 11.6% on RMSE. We credit this to the fact
that FreTS explicitly models both channel and temporal dependencies, and it flexibly unifies channel
and temporal modeling in the frequency domain, which can effectively capture the key patterns with
the global view and energy compaction. We further report the complete benchmarks of short-term
forecasting under different steps on different datasets (including METR-LA dataset) in Appendix F.1.

7
Long-term Time Series Forecasting Table 2 showcases the long-term forecasting results of FreTS
compared to six representative baselines on six benchmarks with various prediction lengths. For
the traffic dataset, we select 48 as the lookback window size L with the prediction lengths τ ∈
{48, 96, 192, 336}. For the other datasets, the input lookback window length is set to 96 and the
prediction length is set to τ ∈ {96, 192, 336, 720}. The results demonstrate that FreTS outperforms
all baselines on all datasets. Quantitatively, compared with the best results of Transformer-based
models, FreTS has an average decrease of more than 20% in MAE and RMSE. Compared with more
recent LSTF-Linear [37] and the SOTA PathchTST [41], FreTS can still outperform them in general.
In addition, we provide further comparison of FreTS and other baselines and report performance
under different lookback window sizes in Appendix F.2. Combining Tables 1 and 2, we can conclude
that FreTS achieves competitive performance in both short-term and long-term forecasting task.

4.2 Model Analysis

Frequency Channel and Temporal Table 3: Ablation studies of frequency channel and temporal
Learners We analyze the effects learners in both short-term and long-term forecasting. ’I/O’
of frequency channel and temporal indicates lookback window sizes/prediction lengths.
learners in Table 3 in both short-term
Tasks Short-term Long-term
and long-term experimental settings.
Dataset Electricity METR-LA Exchange Weather
We consider two variants: FreCL: I/O 12/12 12/12 96/336 96/336
we remove the frequency temporal Metrics MAE RMSE MAE RMSE MAE RMSE MAE RMSE
learner from FreTS, and FreTL: we FreCL 0.054 0.080 0.086 0.168 0.067 0.086 0.051 0.094
remove the frequency channel learner FreTL 0.058 0.086 0.085 0.167 0.065 0.085 0.047 0.091
from FreTS. From the comparison, FreTS 0.050 0.076 0.080 0.166 0.062 0.082 0.046 0.090
we observe that the frequency chan-
nel learner plays a more important role in short-term forecasting. In long-term forecasting, we note
that the frequency temporal learner is more effective than the frequency channel learner. In Appendix
E.1, we also conduct the experiments and report performance on other datasets. Interestingly, we find
out the channel learner would lead to the worse performance in some long-term forecasting cases. A
potential explanation is that the channel independent strategy [41] brings more benefit to forecasting.

FreMLP vs. MLP We further study the effectiveness of FreMLP in time series forecasting. We
use FreMLP to replace the original MLP component in the existing SOTA MLP-based models (i.e.,
DLinear and NLinear [37]), and compare their performances with the original DLinear and NLinear
under the same experimental settings. The experimental results are presented in Table 4. From the
table, we easily observe that for any prediction length, the performance of both DLinear and NLinear
models has been improved after replacing the corresponding MLP component with our FreMLP.
Quantitatively, incorporating FreMLP into the DLinear model brings an average improvement of
6.4% in MAE and 11.4% in RMSE on the Exchange dataset, and 4.9% in MAE and 3.5% in RMSE
on the Weather dataset. A similar improvement has also been achieved on the two datasets with regard
to NLinear, according to Table 4. These results confirm the effectiveness of FreMLP compared to
MLP again and we include more implementation details and analysis in Appendix B.5.

Table 4: Ablation study on the Exchange and Weather datasets with a lookback window size of 96
and the prediction length τ ∈ {96, 192, 336, 720}. DLinear (FreMLP)/NLinear (FreMLP) means
that we replace the MLPs in DLinear/NLinear with FreMLP. The best results are in bold.
Datasets Exchange Weather
Lengths 96 192 336 720 96 192 336 720
Metrics MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE
DLinear 0.037 0.051 0.054 0.072 0.071 0.095 0.095 0.119 0.041 0.081 0.047 0.089 0.056 0.098 0.065 0.106
DLinear (FreMLP) 0.036 0.049 0.053 0.071 0.063 0.071 0.086 0.101 0.038 0.078 0.045 0.086 0.055 0.097 0.061 0.100
NLinear 0.037 0.051 0.051 0.069 0.069 0.093 0.115 0.146 0.037 0.081 0.045 0.089 0.052 0.098 0.058 0.106
NLinear (FreMLP) 0.036 0.050 0.049 0.067 0.067 0.091 0.109 0.139 0.035 0.076 0.043 0.084 0.050 0.094 0.057 0.103

4.3 Efficiency Analysis

The complexity of our proposed FreTS is O(N log N + L log L). We perform efficiency comparisons
with some state-of-the-art GNN-based methods and Transformer-based models under different
numbers of variables N and prediction lengths τ , respectively. On the Wiki dataset, we conduct
experiments over N ∈ {1000, 2000, 3000, 4000, 5000} under the same lookback window size of 12

8
 )UH76 )UH76  )UH76  )UH76
$*&51
6WHP*11
$*&51
 6WHP*11  )('IRUPHU
$XWRIRUPHU 
)('IRUPHU
$XWRIRUPHU

7UDLQLQJ7LPHVHSRFK

7UDLQLQJ7LPHVHSRFK
07*11 07*11  ,QIRUPHU 
,QIRUPHU

3DUDPHWHUV0


3DUDPHWHUV0
 
 
  
 
   
 

                  
9DULDEOH1XPEHUV1 9DULDEOH1XPEHUV1 3UHGLFWLRQOHQJWKV 3UHGLFWLRQOHQJWKV

(a) Parameters (left) and training time (right) under (b) Parameters (left) and training time (right) under
different variable numbers different prediction lengths

Figure 4: Efficiency analysis (model parameters and training time) on the Wiki and Exchange dataset.
(a) The efficiency comparison under different number of variables: the number of variables is enlarged
from 1000 to 5000 with the input window size as 12 and the prediction length as 12 on Wiki dataset.
(b) The efficiency comparison under the prediction lengths: we conduct experiments with prediction
lengths prolonged from 96 to 480 under the same window size of 96 on the Exchange dataset.

and prediction length of 12, as shown in Figure 4(a). From the figure, we can find that: (1) The
amount of FreTS parameters is agnostic to N . (2) Compared with AGCRN, FreTS incurs an average
30% reduction of the number of parameters and 20% reduction of training time. On the Exchange
dataset, we conduct experiments on different prediction lengths τ ∈ {96, 192, 336, 480} with the
same input length of 96. The results are shown in Figure 4(b). It demonstrates: (1) Compared with
Transformer-based methods (FEDformer [32], Autoformer [16], and Informer [15]), FreTS reduces
the number of parameters by at least 3 times. (2) The training time of FreTS is averagely 3 times
faster than Informer, 5 times faster than Autoformer, and more than 10 times faster than FEDformer.
These show our great potential in real-world deployment.

4.4 Visualization Analysis

In Figure 5, we visualize the learned

weights W in FreMLP on the Traffic
dataset with a lookback window size of
48 and prediction length of 192. As the
weights W are complex numbers, we pro-
vide visualizations of the real part Wr (pre-
sented in (a)) and the imaginary part Wi
(presented in (b)) separately. From the fig-
(a) The real part Wr (b) The imaginary part Wi
ure, we can observe that both the real and
imaginary parts play a crucial role in learn- Figure 5: Visualizing learned weights of FreMLP on
ing process: the weight coefficients of the the Traffic dataset. Wr represents the real part of W,
real or imaginary part exhibit energy aggre- and Wi represents the imaginary part.
gation characteristics (clear diagonal pat-
terns) which can facilitate to learn the significant features. In Appendix E.2, we further conduct a
detailed analysis on the effects of the real and imaginary parts in different contexts of forecasting, and
the effects of the two parts in the FreMLP. We examine their individual contributions and investigate
how they influence the final performance. Additional visualizations of the weights on different
datasets with various settings, as well as visualizations of global periodic patterns, can be found in
Appendix G.1 and Appendix G.2, respectively.

5 Conclusion Remarks
In this paper, we explore a novel direction and make a new attempt to apply frequency-domain MLPs
for time series forecasting. We have redesigned MLPs in the frequency domain that can effectively
capture the underlying patterns of time series with global view and energy compaction. We then
verify this design by a simple yet effective architecture, FreTS, built upon the frequency-domain
MLPs for time series forecasting. Our comprehensive empirical experiments on seven benchmarks of
short-term forecasting and six benchmarks of long-term forecasting have validated the superiority
of our proposed methods. Simple MLPs have several advantages and lay the foundation of modern
deep learning, which have great potential for satisfied performance with high efficiency. We hope this
work can facilitate more future research of MLPs on time series modeling.

9
Acknowledgments and Disclosure of Funding
The work was supported in part by the National Key Research and Development Program of China
under Grant 2020AAA0104903 and 2019YFB1406300, and National Natural Science Foundation of
China under Grant 62072039 and 62272048.

References
[1] Edward N Lorenz. Empirical orthogonal functions and statistical weather prediction, volume 1.
Massachusetts Institute of Technology, Department of Meteorology Cambridge, 1956.
[2] Yu Zheng, Xiuwen Yi, Ming Li, Ruiyuan Li, Zhangqing Shan, Eric Chang, and Tianrui Li.
Forecasting fine-grained air quality based on big data. In KDD, pages 2267–2276, 2015.
[3] Wei Fan, Pengyang Wang, Dongkun Wang, Dongjie Wang, Yuanchun Zhou, and Yanjie Fu.
Dish-ts: A general paradigm for alleviating distribution shift in time series forecasting. In AAAI,
pages 7522–7529. AAAI Press, 2023.
[4] Hui He, Qi Zhang, Simeng Bai, Kun Yi, and Zhendong Niu. CATN: cross attentive tree-aware
network for multivariate time series forecasting. In AAAI, pages 4030–4038. AAAI Press, 2022.
[5] Yuzhou Chen, Ignacio Segovia-Dominguez, Baris Coskunuzer, and Yulia Gel. TAMP-s2GCNets:
Coupling time-aware multipersistence knowledge representation with spatio-supra graph con-
volutional networks for time-series forecasting. In International Conference on Learning
Representations, 2022.
[6] Hui He, Qi Zhang, Shoujin Wang, Kun Yi, Zhendong Niu, and Longbing Cao. Learning
informative representation for fairness-aware multivariate time-series forecasting: A group-
based perspective. IEEE Transactions on Knowledge and Data Engineering, pages 1–13,
2023.
[7] Benjamin F King. Market and industry factors in stock price behavior. the Journal of Business,
39(1):139–190, 1966.
[8] Adebiyi A Ariyo, Adewumi O Adewumi, and Charles K Ayo. Stock price prediction using the
arima model. In 2014 UKSim-AMSS 16th international conference on computer modelling and
simulation, pages 106–112. IEEE, 2014.
[9] Charles C Holt. Forecasting trends and seasonal by exponentially weighted moving averages.
ONR Memorandum, 52(2), 1957.
[10] Peter Whittle. Prediction and regulation by linear least-square methods. English Universities
Press, 1963.
[11] David Salinas, Valentin Flunkert, Jan Gasthaus, and Tim Januschowski. Deepar: Probabilistic
forecasting with autoregressive recurrent networks. International Journal of Forecasting,
36(3):1181–1191, 2020.
[12] Guokun Lai, Wei-Cheng Chang, Yiming Yang, and Hanxiao Liu. Modeling long- and short-term
temporal patterns with deep neural networks. In SIGIR, pages 95–104, 2018.
[13] Shaojie Bai, J. Zico Kolter, and Vladlen Koltun. An empirical evaluation of generic convolu-
tional and recurrent networks for sequence modeling. CoRR, abs/1803.01271, 2018.
[14] Minhao Liu, Ailing Zeng, Muxi Chen, Zhijian Xu, Qiuxia Lai, Lingna Ma, and Qiang Xu.
Scinet: time series modeling and forecasting with sample convolution and interaction. Advances
in Neural Information Processing Systems, 35:5816–5828, 2022.
[15] Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai
Zhang. Informer: Beyond efficient transformer for long sequence time-series forecasting. In
AAAI, pages 11106–11115, 2021.
[16] Haixu Wu, Jiehui Xu, Jianmin Wang, and Mingsheng Long. Autoformer: Decomposition
transformers with auto-correlation for long-term series forecasting. In NeurIPS, pages 22419–
22430, 2021.
[17] Zonghan Wu, Shirui Pan, Guodong Long, Jing Jiang, Xiaojun Chang, and Chengqi Zhang.
Connecting the dots: Multivariate time series forecasting with graph neural networks. In KDD,
pages 753–763, 2020.

10
[18] Defu Cao, Yujing Wang, Juanyong Duan, Ce Zhang, Xia Zhu, Congrui Huang, Yunhai Tong,
Bixiong Xu, Jing Bai, Jie Tong, and Qi Zhang. Spectral temporal graph neural network for
multivariate time-series forecasting. In NeurIPS, 2020.
[19] Lei Bai, Lina Yao, Can Li, Xianzhi Wang, and Can Wang. Adaptive graph convolutional
recurrent network for traffic forecasting. In NeurIPS, 2020.
[20] Nikita Kitaev, Lukasz Kaiser, and Anselm Levskaya. Reformer: The efficient transformer. In
ICLR, 2020.
[21] Boris N Oreshkin, Dmitri Carpov, Nicolas Chapados, and Yoshua Bengio. N-beats: Neural basis
expansion analysis for interpretable time series forecasting. arXiv preprint arXiv:1905.10437,
2019.
[22] Tianping Zhang, Yizhuo Zhang, Wei Cao, Jiang Bian, Xiaohan Yi, Shun Zheng, and Jian
Li. Less is more: Fast multivariate time series forecasting with light sampling-oriented mlp
structures. arXiv preprint arXiv:2207.01186, 2022.
[23] Ailing Zeng, Muxi Chen, Lei Zhang, and Qiang Xu. Are transformers effective for time series
forecasting? arXiv preprint arXiv:2205.13504, 2022.
[24] Duraisamy Sundararajan. The discrete Fourier transform: theory, algorithms and applications.
World Scientific, 2001.
[25] Mark W. Watson. Vector autoregressions and cointegration. Working Paper Series, Macroeco-
nomic Issues, 4, 1993.
[26] Dimitros Asteriou and Stephen G Hall. Arima models and the box–jenkins methodology.
Applied Econometrics, 2(2):265–286, 2011.
[27] Bryan Lim and Stefan Zohren. Time-series forecasting with deep learning: a survey. Philosoph-
ical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences,
379(2194):20200209, feb 2021.
[28] José Torres, Dalil Hadjout, Abderrazak Sebaa, Francisco Martínez-Álvarez, and Alicia Troncoso.
Deep learning for time series forecasting: A survey. Big Data, 9, 12 2020.
[29] Zonghan Wu, Shirui Pan, Guodong Long, Jing Jiang, and Chengqi Zhang. Graph wavenet for
deep spatial-temporal graph modeling. In IJCAI, pages 1907–1913, 2019.
[30] Kun Yi, Qi Zhang, Longbing Cao, Shoujin Wang, Guodong Long, Liang Hu, Hui He, Zhendong
Niu, Wei Fan, and Hui Xiong. A survey on deep learning based time series analysis with
frequency transformation. CoRR, abs/2302.02173, 2023.
[31] Liheng Zhang, Charu C. Aggarwal, and Guo-Jun Qi. Stock price prediction via discovering
multi-frequency trading patterns. In KDD, pages 2141–2149, 2017.
[32] Tian Zhou, Ziqing Ma, Qingsong Wen, Xue Wang, Liang Sun, and Rong Jin. FEDformer:
Frequency enhanced decomposed transformer for long-term series forecasting. In ICML, 2022.
[33] Gerald Woo, Chenghao Liu, Doyen Sahoo, Akshat Kumar, and Steven C. H. Hoi. Cost:
Contrastive learning of disentangled seasonal-trend representations for time series forecasting.
In ICLR. OpenReview.net, 2022.
[34] Tian Zhou, Ziqing Ma, Xue Wang, Qingsong Wen, Liang Sun, Tao Yao, Wotao Yin, and Rong
Jin. Film: Frequency improved legendre memory model for long-term time series forecasting.
2022.
[35] Wei Fan, Shun Zheng, Xiaohan Yi, Wei Cao, Yanjie Fu, Jiang Bian, and Tie-Yan Liu. DEPTS:
deep expansion learning for periodic time series forecasting. In ICLR. OpenReview.net, 2022.
[36] Cristian Challu, Kin G. Olivares, Boris N. Oreshkin, Federico Garza, Max Mergenthaler, and
Artur Dubrawski. N-hits: Neural hierarchical interpolation for time series forecasting. CoRR,
abs/2201.12886, 2022.
[37] Ailing Zeng, Muxi Chen, Lei Zhang, and Qiang Xu. Are transformers effective for time series
forecasting? 2023.
[38] Tomás Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word
representations in vector space. In ICLR (Workshop Poster), 2013.

11
[39] Rajat Sen, Hsiang-Fu Yu, and Inderjit S. Dhillon. Think globally, act locally: A deep neural
network approach to high-dimensional time series forecasting. In NeurIPS, pages 4838–4847,
2019.
[40] Yaguang Li, Rose Yu, Cyrus Shahabi, and Yan Liu. Diffusion convolutional recurrent neural
network: Data-driven traffic forecasting. In ICLR (Poster), 2018.
[41] Yuqi Nie, Nam H. Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. A time series is
worth 64 words: Long-term forecasting with transformers. In International Conference on
Learning Representations, 2023.
[42] Bing Yu, Haoteng Yin, and Zhanxing Zhu. Spatio-temporal graph convolutional networks: A
deep learning framework for traffic forecasting. In IJCAI, pages 3634–3640, 2018.
[43] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan,
Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas
Köpf, Edward Z. Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy,
Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style,
high-performance deep learning library. In NeurIPS, pages 8024–8035, 2019.

12
A Notations

Table 5: Notation.

multivariate time series with a lookback window of L

Xt
at timestamps t, Xt ∈ RN ×L
Xt the multivariate values of N distinct series at timestamp t, Xt ∈ RN
the prediction target with a horizon window of length τ
Yt
at timestamps t, Yt ∈ RN ×τ
Ht the hidden representation of Xt , Ht ∈ RN ×L×d

Zt the output of the frequency channel learner, Zt ∈ RN ×L×d

St the output of the frequency temporal learner, St ∈ RN ×L×d

Hchan the domain conversion of Ht on channel dimensions, Hchan ∈ CN ×L×d

Zchan the FreMLP output of Hchan , Zchan ∈ CN ×L×d

Ztemp the domain conversion of Zt on temporal dimensions, Ztemp ∈ CN ×L×d

Stemp the FreMLP output of Ztemp , Stemp ∈ CN ×L×d

the complex number weight matrix of FreMLP in the frequency
W chan
channel learner, W chan ∈ Cd×d
the complex number bias of FreMLP in the frequency channel
B chan
learner, B chan ∈ Cd
the complex number weight matrix of FreMLP in the frequency
W temp
temporal learner, W temp ∈ Cd×d
the complex number bias of FreMLP in the frequency
B temp
temporal learner, B temp ∈ Cd

B Experimental Details

B.1 Datasets

We adopt thirteen real-world benchmarks in the experiments to evaluate the accuracy of short-term
and long-term forecasting. The details of the datasets are as follows:
Solar4 : It is about the solar power collected by National Renewable Energy Laboratory. We choose
the power plant data points in Florida as the data set which contains 593 points. The data is collected
from 01/01/2006 to 31/12/2016 with the sampling interval of every 1 hour.
Wiki [39]: It contains a number of daily views of different Wikipedia articles and is collected from
1/7/2015 to 31/12/2016. It consists of approximately 145k time series and we randomly choose 5k
from them as our experimental data set.
Traffic [39]: It contains hourly traffic data from 963 San Francisco freeway car lanes for short-term
forecasting settings while it contains 862 car lanes for long-term forecasting. It is collected since
01/01/2015 with a sampling interval of every 1 hour.
ECG5 : It is about Electrocardiogram(ECG) from the UCR time-series classification archive. It
contains 140 nodes and each node has a length of 5000.

4
https://www.nrel.gov/grid/solar-power-data.html
5
http://www.timeseriesclassification.com/description.php?Dataset=ECG5000

13
Electricity6 : It contains electricity consumption of 370 clients for short-term forecasting while
it contains electricity consumption of 321 clients for long-term forecasting. It is collected since
01/01/2011. The data sampling interval is every 15 minutes.
COVID-19 [5]: It is about COVID-19 hospitalization in the U.S. state of California (CA) from
01/02/2020 to 31/12/2020 provided by the Johns Hopkins University with the sampling interval of
every day.
METR-LA7 : It contains traffic information collected from loop detectors in the highway of Los
Angeles County. It contains 207 sensors which are from 01/03/2012 to 30/06/2012 and the data
sampling interval is every 5 minutes.
Exchange8 : It contains the collection of the daily exchange rates of eight foreign countries including
Australia, British, Canada, Switzerland, China, Japan, New Zealand, and Singapore ranging from
1990 to 2016 and the data sampling interval is every 1 day.
Weather9 : It collects 21 meteorological indicators, such as humidity and air temperature, from the
Weather Station of the Max Planck Biogeochemistry Institute in Germany in 2020. The data sampling
interval is every 10 minutes.
ETT10 : It is collected from two different electric transformers labeled with 1 and 2, and each of them
contains 2 different resolutions (15 minutes and 1 hour) denoted with m and h. We use ETTh1 and
ETTm1 as our long-term forecasting benchmarks.

B.2 Baselines

We adopt eighteen representative and state-of-the-art baselines for comparison including LSTM-based
models, GNN-based models, and Transformer-based models. We introduce these models as follows:
VAR [25]: VAR is a classic linear autoregressive model. We use the Statsmodels library (https:
//www.statsmodels.org) which is a Python package that provides statistical computations to
realize the VAR.
DeepGLO [39]: DeepGLO models the relationships among variables by matrix factorization and
employs a temporal convolution neural network to introduce non-linear relationships. We download
the source code from: https://github.com/rajatsen91/deepglo. We use the recommended
configuration as our experimental settings for Wiki, Electricity, and Traffic datasets. For the COVID-
19 dataset, the vertical and horizontal batch size is set to 64, the rank of the global model is set to 64,
the number of channels is set to [32, 32, 32, 1], and the period is set to 7.
LSTNet [12]: LSTNet uses a CNN to capture inter-variable relationships and an RNN to discover
long-term patterns. We download the source code from: https://github.com/laiguokun/
LSTNet. In our experiment, we use the recommended configuration where the number of CNN
hidden units is 100, the kernel size of the CNN layers is 4, the dropout is 0.2, the RNN hidden units
is 100, the number of RNN hidden layers is 1, the learning rate is 0.001 and the optimizer is Adam.
TCN [13]: TCN is a causal convolution model for regression prediction. We download the source code
from: https://github.com/locuslab/TCN. We utilize the same configuration as the polyphonic
music task exampled in the open source code where the dropout is 0.25, the kernel size is 5, the
number of hidden units is 150, the number of levels is 4 and the optimizer is Adam.
Informer [15]: Informer leverages an efficient self-attention mechanism to encode the dependen-
cies among variables. We download the source code from: https://github.com/zhouhaoyi/
Informer2020. We use the recommended configuration as the experimental settings where the
dropout is 0.05, the number of encoder layers is 2, the number of decoder layers is 1, the learning
rate is 0.0001, and the optimizer is Adam.
Reformer [20]: Reformer combines the modeling capacity of a Transformer with an architecture that
can be executed efficiently on long sequences and with small memory use. We download the source
6
https://archive.ics.uci.edu/ml/datasets/ElectricityLoadDiagrams20112014
7
https://github.com/liyaguang/DCRNN
8
https://github.com/laiguokun/multivariate-time-series-data
9
https://www.bgc-jena.mpg.de/wetter/
10
https://github.com/zhouhaoyi/ETDataset

14
code from: https://github.com/thuml/Autoformer. We use the recommended configuration
as the experimental settings.
Autoformer [16]: Autoformer proposes a decomposition architecture by embedding the series
decomposition block as an inner operator, which can progressively aggregate the long-term trend part
from intermediate prediction. We download the source code from: https://github.com/thuml/
Autoformer. We use the recommended configuration as the experimental settings.
FEDformer [32]: FEDformer proposes an attention mechanism with low-rank approximation in
frequency and a mixture of expert decomposition to control the distribution shifting. We download the
source code from: https://github.com/MAZiqing/FEDformer. We use FEB-f as the Frequency
Enhanced Block and select the random mode with 64 as the experimental mode.
SFM [31]: On the basis of the LSTM model, SFM introduces a series of different frequency compo-
nents in the cell states. We download the source code from: https://github.com/z331565360/
State-Frequency-Memory-stock-prediction. We follow the recommended configuration as
the experimental settings where the learning rate is 0.01, the frequency dimension is 10, the hidden
dimension is 10 and the optimizer is RMSProp.
StemGNN [18]: StemGNN leverages GFT and DFT to capture dependencies among variables in
the frequency domain. We download the source code from: https://github.com/microsoft/
StemGNN. We use the recommended configuration of stemGNN as our experiment setting where the
optimizer is RMSProp, the learning rate is 0.0001, the number of stacked layers is 5, and the dropout
rate is 0.5.
MTGNN [17]: MTGNN proposes an effective method to exploit the inherent dependency relation-
ships among multiple time series. We download the source code from: https://github.com/
nnzhan/MTGNN. Because the experimental datasets have no static features, we set the parameter
load_static_feature to false. We construct the graph by the adaptive adjacency matrix and add the
graph convolution layer. Regarding other parameters, we follow the recommended settings.
GraphWaveNet [29]: GraphWaveNet introduces an adaptive dependency matrix learning to cap-
ture the hidden spatial dependency. We download the source code from: https://github.com/
nnzhan/Graph-WaveNet. Since our datasets have no prior defined graph structures, we use only
adaptive adjacent matrix. We add a graph convolutional layer and randomly initialize the adjacent
matrix. We adopt the recommended setting as its experimental configuration where the learning rate
is 0.001, the dropout is 0.3, the number of epochs is 50, and the optimizer is Adam.
AGCRN [19]: AGCRN proposes a data-adaptive graph generation module for discovering spatial
correlations from data. We download the source code from: https://github.com/LeiBAI/AGCRN.
We follow the recommended settings where the embedding dimension is 10, the learning rate is 0.003,
and the optimizer is Adam.
TAMP-S2GCNets [5]: TAMP-S2GCNets explores the utility of MP to enhance knowledge represen-
tation mechanisms within the time-aware DL paradigm. We download the source code from: https:
//www.dropbox.com/sh/n0ajd5l0tdeyb80/AABGn-ejfV1YtRwjf_L0AOsNa?dl=0. TAMP-
S2GCNets require a pre-defined graph topology and we use the California State topology provided
by the source code as input. We adopt the recommended settings as the experimental configuration
for COVID-19.
DCRNN [40]: DCRNN uses bidirectional graph random walk to model spatial dependency and
recurrent neural network to capture the temporal dynamics. We download the source code from:
https://github.com/liyaguang/DCRNN. We use the recommended configuration as our experi-
mental settings with the batch size is 64, the learning rate is 0.01, the input dimension is 2 and the
optimizer is Adam. DCRNN requires a pre-defined graph structure and we use the adjacency matrix
as the pre-defined structure provided by the METR-LA dataset.
STGCN [42]: STGCN integrates graph convolution and gated temporal convolution through spatial-
temporal convolutional blocks. We download the source code from: https://github.com/
VeritasYin/STGCN_IJCAI-18. We follow the recommended settings as our experimental config-
uration where the batch size is 50, the learning rate is 0.001 and the optimizer is Adam. STGCN
requires a pre-defined graph structure and we leverage the adjacency matrix as the pre-defined
structure provided by the METR-LA dataset.
LTSF-Linear [37]: LTSF-Linear proposes a set of embarrassingly simple one-layer linear models to

15
learn temporal relationships between input and output sequences. We download the source code from:
https://github.com/cure-lab/LTSF-Linear. We use it as our long-term forecasting baseline
and follow the recommended settings as experimental configuration.
PatchTST [41]: PatchTST proposes an effective design of Transformer-based models for time series
forecasting tasks by introducing two key components: patching and channel-independent structure.
We download the source code from: https://github.com/PatchTST. We use it as our long-term
forecasting baseline and adhere to the recommended settings as the experimental configuration.

B.3 Implementation Details

By default, both the frequency channel and temporal learners contain one layer of FreMLP with
the embedding size d of 128, and the hidden size dh is set to 256. For short-term forecasting, the
batch size is set to 32 for Solar, METR-LA, ECG, COVID-19, and Electricity datasets. And for Wiki
and Traffic datasets, the batch size is set to 4. For the long-term forecasting, except for the lookback
window size, we follow most of the experimental settings of LTSF-Linear [37]. The lookback window
size is set to 96 which is recommended by FEDformer [32] and Autoformer [16]. In Appendix
F.2, we also use 192 and 336 as the lookback window size to conduct experiments and the results
demonstrate that FreTS outperforms other baselines as well. For the longer prediction lengths (e.g.,
336, 720), we use the channel independence strategy and contain only the frequency temporal learner
in our model. For some datasets, we carefully tune the hyperparameters including the batch size and
learning rate on the validation set, and we choose the settings with the best performance. We tune the
batch size over {4, 8, 16, 32}.

B.4 Visualization Settings

The Visualization Method for Global View. We follow the visualization methods in LTSF-
Linear [37] to visualize the weights learned in the time domain on the input (corresponding to
the left side of Figure 1(a)). For the visualization of the weights learned on the frequency spectrum,
we first transform the input into the frequency domain and select the real part of the input frequency
spectrum to replace the original input. Then, we learn the weights and visualize them in the same
manner as in the time domain. The right side of Figure 1(a) shows the weights learned on the Traffic
dataset with a lookback window of 96 and a prediction length of 96, Figure 9 displays the weights
learned on the Traffic dataset with a lookback window of 72 and a prediction length of 336, and
Figure 10 is the weights learned on the Electricity dataset with a lookback window of 96 and a
prediction length of 96.
The Visualization Method for Energy Compaction. Since the learned weights W = Wr + jWi ∈
Cd×d of the frequency-domain MLPs are complex numbers, we visualize the corresponding real part
Wr and imaginary part Wi , respectively. We normalize them by the calculation of 1/ max(W) ∗ W
and visualize the normalization values. The right side of Figure 1(b) is the real part of W learned
on the Traffic dataset with a lookback window of 48 and a prediction length of 192. To visualize
the corresponding weights learned in the time domain, we replace the frequency spectrum of input
Ztemp ∈ CN ×L×d with the original time domain input Ht ∈ RN ×L×d and perform calculations in
the time domain with a weight W ∈ Rd×d , as depicted in the left side of Figure 1(b).

B.5 Ablation Experimental Settings

DLinear decomposes a raw data input into a trend component and a seasonal component, and two one-
layer linear layers are applied to each component. In the ablation study part, we replace the two linear
layers with two different frequency-domain MLPs (corresponding to DLinear (FreMLP) in Table 4),
and compare their accuracy using the same experimental settings recommended in LTSF-Linear [37].
NLinear subtracts the input by the last value of the sequence. Then, the input goes through a linear
layer, and the subtracted part is added back before making the final prediction. We replace the linear
layer with a frequency-domain MLP (corresponding to NLinear (FreMLP) in Table 4), and compare
their accuracy using the same experimental settings recommended in LTSF-Linear [37].

16
C Complex Multiplication
For two complex number values Z1 = (a + jb) and Z2 = (c + jd), where a and c is the real
part of Z1 and Z2 respectively, b and d is the imaginary part of Z1 and Z2 respectively. Then the
multiplication of Z1 and Z2 is calculated by:
Z1 Z2 = (a + jb)(c + jd) = ac + j 2 bd + jad + jbc = (ac − bd) + j(ad + bc) (10)
where j 2 = −1.

D Proof
D.1 Proof of Theorem 1

Theorem 1. Suppose that H is the representation of raw time series and H is the corresponding
frequency components of the spectrum, then the energy of a time series in the time domain is equal to
the energy of its representation in the frequency domain. Formally, we can express this with above
notations by: Z ∞ Z ∞
|H(v)|2 dv = |H(f )|2 df (11)
−∞ −∞
R∞
where H(f ) = −∞
H(v)e−j2πf v dv, v is the time/channel dimension, f is the frequency dimension.

Proof. Given the representation of raw time series H ∈ RN ×L×d , let us consider performing
integration in either the N dimension (channel dimension) or the L dimension (temporal dimension),
denoted as the integral over v, then
Z ∞ Z ∞
|H(v)|2 dv = H(v)H∗ (v)dv
−∞ −∞
R∞
where H∗ (v) is the conjugate of H(v). According to IDFT, H∗ (v) = −∞ H∗ (f )e−j2πf v df , we
can obtain
Z ∞ Z ∞ Z ∞
|H(v)|2 dv = H(v)[ H∗ (f )e−j2πf v df ]dv
−∞ −∞ −∞
Z ∞ Z ∞
= H∗ (f )[ H(v)e−j2πf v dv]df
−∞ −∞
Z ∞
= H∗ (f )H(f )df
−∞
Z ∞
= |H(f )|2 df
−∞

Proved.

Therefore, the energy of a time series in the time domain is equal to the energy of its representation
in the frequency domain.

D.2 Proof of Theorem 2

Theorem 2. Given the time series input H and its corresponding frequency domain conversion H,
the operations of frequency-domain MLP on H can be represented as global convolutions on H in
the time domain. This can be given by:
HW + B = F(H ∗ W + B) (12)
where ∗ is a circular convolution, W and B are the complex number weight and bias, W and B are
the weight and bias in the time domain, and F is DFT.

17
Proof. Suppose that we conduct operations in the N (i.e., channel dimension) or L (i.e., temporal
dimension) dimension, then
Z ∞
F(H(v) ∗ W (v)) = (H(v) ∗ W (v))e−j2πf v dv
−∞
R∞
According to convolution theorem, H(v) ∗ W (v) = −∞
(H(τ )W (v − τ ))dτ , then
Z ∞ Z ∞
F(H(v) ∗ W (v)) = (H(τ )W (v − τ ))e−j2πf v dτ dv
−∞ −∞
Z ∞ Z ∞
= W (v − τ )e−j2πf v dvH(τ )dτ
−∞ −∞

Let x = v − τ , then
Z ∞ Z ∞
F(H(v) ∗ W (v)) = W (x)e−j2πf (x+τ ) dxH(τ )dτ
−∞ −∞
Z ∞ Z ∞
= W (x)e−j2πf x e−j2πf τ dxH(τ )dτ
−∞ −∞
Z ∞ Z ∞
−j2πf τ
= H(τ )e dτ W (x)e−j2πf x dx
−∞ −∞
= H(f )W(f )
Accordingly, (H(v) ∗ W (v)) in the time domain is equal to (H(f )W(f )) in the frequency domain.
Therefore, the operations of FreMLP (HW + B) in the channel (i.e., v = N ) or temporal dimension
(i.e., v = L), are equal to the operations (H ∗ W + B) in the time domain. This implies that
frequency-domain MLPs can be viewed as global convolutions in the time domain. Proved.

E Further Analysis
E.1 Ablation Study

In this section, we further analyze the effects of the frequency channel and temporal learners with
different prediction lengths on ETTm1 and ETTh1 datasets. The results are shown in Table 6. It
demonstrates that with the prediction length increasing, the frequency temporal learner shows more
effective than the channel learner. Especially, when the prediction length is longer (e.g., 336, 720),
the channel learner will lead to worse performance. The reason is that when the prediction lengths
become longer, the model with the channel learner is likely to overfit data during training. Thus for
long-term forecasting with longer prediction lengths, the channel independence strategy may be more
effective, as described in PatchTST [41].

Table 6: Ablation studies of the frequency channel and temporal learners in long-term forecasting.
’I/O’ indicates lookback window sizes/prediction lengths.
Dataset ETTm1 ETTh1
I/O 96/96 96/192 96/336 96/720 96/96 96/192 96/336 96/720
Metrics MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE
FreCL 0.053 0.078 0.059 0.085 0.067 0.095 0.097 0.125 0.063 0.089 0.067 0.093 0.071 0.097 0.087 0.115
FreTL 0.053 0.078 0.058 0.084 0.062 0.089 0.069 0.096 0.061 0.087 0.065 0.091 0.070 0.096 0.082 0.108
FreTS 0.052 0.077 0.057 0.083 0.064 0.092 0.071 0.099 0.063 0.089 0.066 0.092 0.072 0.098 0.086 0.113

E.2 Impacts of Real/Imaginary Parts

To investigate the effects of real and imaginary parts, we conduct experiments on Exchange and
ETTh1 datasets under different prediction lengths L ∈ {96, 192} with the lookback window of 96.
Furthermore, we analyze the effects of Wr and Wi in the weights W = Wr + jWi of FreMLP. In
this experiment, we only use the frequency temporal learner in our model. The results are shown in

18
Table 7. In the table, Inputreal indicates that we only feed the real part of the input into the network,
and Inputimag indicates that we only feed the imaginary part of the input into the network. W(Wr )
denotes that we set Wi to 0 and W(Wi ) denotes that we set Wr to 0. From the table, we can observe
that both the real part and imaginary part of input are indispensable and the real part is more important
to the imaginary part, and the real part of W plays a more significant role for the model performances.

Table 7: Investigation the impacts of real/imaginary parts

Dataset Exchange ETTh1

I/O 96/96 96/192 96/96 96/192
Metrics MAE RMSE MAE RMSE MAE RMSE MAE RMSE
Inputreal 0.048 0.062 0.058 0.074 0.080 0.111 0.083 0.113
Inputimag 0.143 0.185 0.143 0.184 0.130 0.156 0.130 0.156
W(Wr ) 0.039 0.053 0.051 0.067 0.063 0.089 0.067 0.093
W(Wi ) 0.143 0.184 0.142 0.184 0.116 0.138 0.117 0.139
FreTS 0.037 0.051 0.050 0.067 0.061 0.087 0.065 0.091

E.3 Parameter Sensitivity

We further perform extensive experiments on the ECG dataset to evaluate the sensitivity of the input
length L and the embedding dimension size d. (1) Input length: We tune over the input length with the
value {6, 12, 18, 24, 30, 36, 42, 50, 60} on the ECG dataset and the prediction length is 12, and the
result is shown in Figure 6(a). From the figure, we can find that with the input length increasing, the
performance first becomes better because the long input length may contain more pattern information,
and then it decreases due to data redundancy or overfitting. (2) Embedding size: We choose the
embedding size over the set {32, 64, 128, 256, 512} on the ECG dataset. The results are shown in
Figure 6(b). It shows that the performance first increases and then decreases with the increase of the
embedding size because a large embedding size improves the fitting ability of our FreTS but may
easily lead to overfitting especially when the embedding size is too large.

(a) Input window length (b) Embedding size

Figure 6: The parameter sensitivity analyses of FreTS.

F Additional Results
F.1 Multi-Step Forecasting

To further evaluate the performance of our FreTS in multi-step forecasting, we conduct more
experiments on METR-LA and COVID-19 datasets with the input length of 12 and the prediction
lengths of {3, 6, 9, 12}, and the results are shown in Tables 8 and 9, respectively. In this experiment,
we only select the state-of-the-art (i.e., GNN-based and Transformer-based) models as the baselines
since they perform better than other models, such as RNN and TCN. Among these baselines, STGCN,
DCRNN, and TAMP-S2GCNets require pre-defined graph structures. The results demonstrate that

19
FreTS outperforms other baselines, including those models with pre-defined graph structures, at
all steps. This further confirms that FreTS has strong capabilities in capturing channel-wise and
time-wise dependencies.

Table 8: Multi-step short-term forecasting results comparison on the METR-LA dataset with the
input length of 12 and the prediction length of τ ∈ {3, 6, 9, 12}. We highlight the best results in bold
and the second best results are underline.
Length 3 6 9 12
Metrics MAE RMSE MAE RMSE MAE RMSE MAE RMSE
Reformer 0.086 0.154 0.097 0.176 0.107 0.193 0.118 0.206
Informer 0.082 0.156 0.094 0.176 0.108 0.193 0.125 0.214
Autoformer 0.087 0.149 0.091 0.162 0.106 0.178 0.099 0.184
FEDformer 0.064 0.127 0.073 0.145 0.079 0.160 0.086 0.175
DCRNN 0.160 0.204 0.191 0.243 0.216 0.269 0.241 0.291
STGCN 0.058 0.133 0.080 0.177 0.102 0.209 0.128 0.238
GraphWaveNet 0.180 0.366 0.184 0.375 0.196 0.382 0.202 0.386
MTGNN 0.135 0.294 0.144 0.307 0.149 0.328 0.153 0.316
StemGNN 0.052 0.115 0.069 0.141 0.080 0.162 0.093 0.175
AGCRN 0.062 0.131 0.086 0.165 0.099 0.188 0.109 0.204
FreTS 0.050 0.113 0.066 0.140 0.076 0.158 0.080 0.166

Table 9: Multi-step short-term forecasting results comparison on the COVID-19 dataset with the
input length of 12 and the prediction length of τ ∈ {3, 6, 9, 12}. We highlight the best results in bold
and the second best results are underline.

Length 3 6 9 12
Metrics MAE RMSE MAE RMSE MAE RMSE MAE RMSE
Reformer 0.212 0.282 0.139 0.186 0.148 0.197 0.152 0.209
Informer 0.234 0.312 0.190 0.245 0.184 0.242 0.200 0.259
Autoformer 0.212 0.280 0.144 0.191 0.152 0.201 0.159 0.211
FEDformer 0.246 0.328 0.169 0.242 0.175 0.247 0.160 0.219
GraphWaveNet 0.092 0.129 0.133 0.179 0.171 0.225 0.201 0.255
StemGNN 0.247 0.318 0.344 0.429 0.359 0.442 0.421 0.508
AGCRN 0.130 0.172 0.171 0.218 0.224 0.277 0.254 0.309
MTGNN 0.276 0.379 0.446 0.513 0.484 0.548 0.394 0.488
TAMP-S2GCNets 0.140 0.190 0.150 0.200 0.170 0.230 0.180 0.230
FreTS 0.071 0.103 0.093 0.131 0.109 0.148 0.124 0.164

F.2 Long-Term Forecasting under Varying Lookback Window

In Table 10, we present the long-term forecasting results of our FreTS and other baselines
(PatchTST [41], LTSF-linear [37], FEDformer [32], Autoformer [16], Informer [15], and Re-
former [20]) under different lookback window lengths L ∈ {96, 192, 336} on the Exchange dataset.
The prediction lengths are {96, 192, 336, 720}. From the table, we can observe that our FreTS
outperforms all baselines in all settings and achieves significant improvements than FEDformer [32],
Autoformer [16], Informer [15], and Reformer [20]. It verifies the effectiveness of our FreTS in
learning informative representation under different lookback window.

G Visualizations

G.1 Weight Visualizations for Energy Compaction

We further visualize the weights W = Wr + jWi in the frequency temporal learner under different
settings, including different lookback window sizes and prediction lengths, on the Traffic and
Electricity datasets. The results are illustrated in Figures 7 and 8. These figures demonstrate that

20
Table 10: Long-term forecasting results comparison with different lookback window lengths L ∈
{96, 192, 336}. The prediction lengths are as τ ∈ {96, 192, 336, 720}. The best results are in bold
and the second best results are underlined.
Models FreTS PatchTST LTSF-Linear FEDformer Autoformer Informer Reformer
Metrics MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE
96 0.037 0.051 0.039 0.052 0.038 0.052 0.050 0.067 0.050 0.066 0.066 0.084 0.126 0.146
192 0.050 0.067 0.055 0.074 0.053 0.069 0.064 0.082 0.063 0.083 0.068 0.088 0.147 0.169
96

336 0.062 0.082 0.071 0.093 0.064 0.085 0.080 0.105 0.075 0.101 0.093 0.127 0.157 0.189
720 0.088 0.110 0.132 0.166 0.092 0.116 0.151 0.183 0.150 0.181 0.117 0.170 0.166 0.201
96 0.036 0.050 0.037 0.051 0.038 0.051 0.067 0.086 0.066 0.085 0.109 0.131 0.123 0.143
192 0.051 0.068 0.052 0.070 0.053 0.070 0.080 0.101 0.080 0.102 0.144 0.172 0.139 0.161
192

336 0.066 0.087 0.072 0.097 0.073 0.096 0.093 0.122 0.099 0.129 0.141 0.177 0.155 0.181
720 0.088 0.110 0.099 0.128 0.098 0.122 0.190 0.222 0.191 0.224 0.173 0.210 0.159 0.193
96 0.038 0.052 0.039 0.053 0.040 0.055 0.088 0.113 0.088 0.110 0.137 0.169 0.128 0.148
192 0.053 0.070 0.055 0.071 0.055 0.072 0.103 0.133 0.104 0.133 0.161 0.195 0.138 0.159
336

336 0.071 0.092 0.074 0.099 0.077 0.100 0.123 0.155 0.127 0.159 0.156 0.193 0.156 0.179
720 0.082 0.108 0.100 0.129 0.087 0.110 0.210 0.242 0.211 0.244 0.173 0.210 0.168 0.205

the weight coefficients of the real or imaginary part exhibit energy aggregation characteristics (clear
diagonal patterns) which can facilitate frequency-domain MLPs in learning the significant features.

(a) Wr under I/O=48/192 (b) Wr under I/O=48/336 (c) Wr under I/O=72/336

(d) Wi under I/O=48/192 (e) Wi under I/O=48/336 (f) Wi under I/O=72/336

Figure 7: The visualizations of the weights W in the frequency temporal learner on the Traffic dataset.
’I/O’ denotes lookback window sizes/prediction lengths. Wr and Wi are the real and imaginary parts
of W, respectively.

G.2 Weight Visualizations for Global View

To verify the characteristics of a global view of learning in the frequency domain, we perform
additional experiments on the Traffic and Electricity datasets and compare the weights learned on
the input in the time domain with those learned on the input frequency spectrum. The results are
presented in Figures 9 and 10. The left side of the figures displays the weights learned on the input
in the time domain, while the right side shows those learned on the real part of the input frequency
spectrum. From the figures, we can observe that the patterns learned on the input frequency spectrum
exhibit more obvious periodic patterns compared to the time domain. This is attributed to the global
view characteristics of the frequency domain. Furthermore, we visualize the predictions of FreTS on

21
 0 0 1.0 0 1.00
0.8 0.75
20 20 20
0.5 0.50
0.6
40 40 40 0.25
0.4
0.0
60 0.2 60 60 0.00
0.25
0.0 0.5
80 80 80
0.50
0.2
100 100 1.0 100 0.75
0.4
1.00
120 0.6 120 120
1.5
0 20 40 60 80 100 120 0 20 40 60 80 100 120 0 20 40 60 80 100 120

(a) Wr under I/O=96/96 (b) Wr under I/O=96/336 (c) Wr under I/O=96/720

0 0 1.0 0 1.0

0.75 0.8 0.8

20 20 20
0.50 0.6 0.6
40 40 40 0.4
0.25 0.4
60 60 60 0.2
0.00 0.2
0.0
80 0.25 80 0.0 80
0.2
0.50 0.2
100 100 100 0.4
0.75 0.4
0.6
120 120 120
1.00 0.6
0 20 40 60 80 100 120 0 20 40 60 80 100 120 0 20 40 60 80 100 120

(d) Wi under I/O=96/96 (e) Wi under I/O=96/336 (f) Wi under I/O=96/720

Figure 8: The visualizations of the weights W in the frequency temporal learner on the Electricity
dataset. ’I/O’ denotes lookback window sizes/prediction lengths. Wr and Wi are the real and
imaginary parts of W, respectively.

the Traffic and Electricity datasets, as depicted in Figures 11 and 12, which show that FreTS exhibit
a good ability to fit cyclic patterns. In summary, these results demonstrate that FreTS has a strong
capability to capture the global periodic patterns, which benefits from the global view characteristics
of the frequency domain.

0.15
0.8

0.10
0.6
0.05
0 0
25 0.4 25
0.00
50 50
0 50 100 150 200 250 300 0.2 0 50 100 150 200 250 300 0.05

0.0 0.10

0.15

(a) Learned on the input (b) Learned on the frequency spectrum

Figure 9: Visualization of the weights (L × τ ) on the Traffic dataset with lookback window size of
72 and prediction length of 336.

22
 0 0.10 0 0.100
0.075
20 0.08 20 0.050
0.06 0.025
40 40
0.04 0.000
0.025
60 0.02 60
0.050
0.00 0.075
80 80
0.02 0.100
0 20 40 60 80 0 20 40 60 80

(a) Learned on the input (b) Learned on the frequency spectrum

Figure 10: Visualization of the weights (L × τ ) on the Electricity dataset with lookback window size
of 96 and prediction length of 96.

GroundTruth GroundTruth
0.08 Prediction 0.08 Prediction

0.06 0.06

0.04 0.04

0.02 0.02

0.00 0.00
0 20 40 60 80 0 20 40 60 80 100 120 140

(a) I/O=48/48 (b) I/O=48/96

GroundTruth GroundTruth
0.08 Prediction 0.08 Prediction

0.06 0.06

0.04 0.04

0.02 0.02

0.00 0.00
0 50 100 150 200 250 0 50 100 150 200 250 300 350 400

(c) I/O=48/192 (d) I/O=48/336

Figure 11: Visualizations of predictions (forecast vs. actual) on the Traffic dataset. ’I/O’ denotes
lookback window sizes/prediction lengths.

23
 0.80
GroundTruth GroundTruth
Prediction 0.60 Prediction
0.75

0.70
0.55
0.65

0.60 0.50

0.55
0.45
0.50

0.45
0.40
0 25 50 75 100 125 150 175 200 0 50 100 150 200 250 300

(a) I/O=96/96 (b) I/O=96/192

0.65 GroundTruth
0.60 Prediction
0.60

0.55
0.55

0.50 0.50

0.45 0.45

GroundTruth
Prediction 0.40
0.40
0 100 200 300 400 0 100 200 300 400 500 600 700 800

(c) I/O=96/336 (d) I/O=96/720

Figure 12: Visualizations of predictions (forecast vs. actual) on the Electricity dataset. ’I/O’ denotes
lookback window sizes/prediction lengths.

24