Applying Informer for Option Pricing: A Transformer-Based Approach
Feliks Ba
´
nka
a
and Jarosław A. Chudziak
b
The Faculty of Electronics and Information Technology, Warsaw University of Technology, Poland
Keywords:
Option Pricing, Transformers, Neural Networks, Time Series Forecasting, Deep Learning.
Abstract:
Accurate option pricing is essential for effective trading and risk management in financial markets, yet it
remains challenging due to market volatility and the limitations of traditional models like Black-Scholes.
In this paper, we investigate the application of the Informer neural network for option pricing, leveraging
its ability to capture long-term dependencies and dynamically adjust to market fluctuations. This research
contributes to the field of financial forecasting by introducing Informer’s efficient architecture to enhance
prediction accuracy and provide a more adaptable and resilient framework compared to existing methods.
Our results demonstrate that Informer outperforms traditional approaches in option pricing, advancing the
capabilities of data-driven financial forecasting in this domain.
1 INTRODUCTION
Option pricing is a cornerstone of modern finance, es-
sential for developing trading strategies and managing
risk. Options enable traders and investors to hedge
against potential losses or speculate on price move-
ments. A call (put) option grants the holder the right,
but not the obligation, to buy (sell) an asset at a speci-
fied price before the contract expires. Accurate option
pricing models shape critical decisions in hedging and
risk management, directly affecting trading portfolio
profitability and stability.
Early theoretical frameworks, such as the
Black–Scholes (Black and Scholes, 1973; Merton,
1973) and the Heston (Heston, 1993) models, offered
valuable mathematical foundations but often rely on
simplifying assumptions (e.g., constant volatility).
These assumptions do not always hold in real-world
markets, where sudden shifts in macroeconomic con-
ditions or sentiment can lead to rapid changes in asset
prices (Bollerslev, 1986). Over the past few decades,
machine learning techniques—such as LSTM-based
neural networks (Hochreiter and Schmidhuber, 1997;
Yue Liu, 2023; Bao et al., 2017)—have demonstrated
improved adaptability by capturing non-linearities
and sequential dependencies. Yet, their effective-
ness can be limited when handling very long time
sequences, which demand more efficient and robust
architectures.
Transformer-based models, originally devised for
natural language processing (A. Vaswani and Polo-
a
https://orcid.org/0009-0005-1973-5861
b
https://orcid.org/0000-0003-4534-8652
sukhin, 2017), have shown promise in overcoming
these challenges by leveraging self-attention mecha-
nisms that allow for parallelized long-sequence pro-
cessing. Recent advances, such as the Informer
model (H. Zhou and Zhang, 2021), have introduced
more efficient attention mechanisms geared toward
time-series data. However, their application within
option pricing remains underexplored, motivating the
present study to investigate whether Informer’s long-
horizon capability and computational efficiency can
produce more accurate predictions in option pricing
tasks.
This paper contributes to the field of financial
modeling by evaluating the application of the In-
former architecture for predicting option prices, lever-
aging its efficient attention mechanism and long-
sequence modeling capabilities to enhance predic-
tion accuracy and adaptability to market fluctuations.
Informer’s ability to handle long-term dependencies
makes it an ideal candidate for modeling complex fi-
nancial data, offering a more advanced approach com-
pared to traditional models like Black-Scholes (Black
and Scholes, 1973; Merton, 1973) and Heston (He-
ston, 1993), as well as existing machine learning
models such as LSTM (Hochreiter and Schmidhuber,
1997; Yue Liu, 2023). The contributions of this paper
are as follows:
• We apply the Informer architecture to option pric-
ing, leveraging its long-sequence modeling capa-
bilities and self-attention mechanisms to enhance
prediction accuracy.
• We benchmark the model against traditional and
machine learning-based approaches, evaluating
1270
Ba
´
nka, F. and Chudziak, J. A.
Applying Informer for Option Pricing: A Transformer-Based Approach.
DOI: 10.5220/0013320900003890
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 17th International Conference on Agents and Artificial Intelligence (ICAART 2025) - Volume 3, pages 1270-1277
ISBN: 978-989-758-737-5; ISSN: 2184-433X
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.

its performance in high-volatility scenarios.
• We present an analysis of Informer’s predictive
accuracy and trading profitability on historical
data.
The remainder of this paper is organized as fol-
lows: Section 2 discusses related work, focusing on
traditional and machine learning approaches to option
pricing and the emerging role of Transformers in fi-
nance. Section 3 outlines the Informer-based method-
ology applied to option pricing. Section 4 presents
the experimental setup and results, and Section 5 con-
cludes with a summary and potential directions for
future research.
2 RELATED WORK
The foundational models for option pricing, such as
the Black-Scholes model (Black and Scholes, 1973;
Merton, 1973) and the binomial model (Cox et al.,
1979), have been pivotal in shaping early financial
derivatives pricing. These models introduced critical
concepts such as risk-neutral valuation but often rest
on simplifying assumptions, such as constant volatil-
ity, which do not align with real-world market condi-
tions. The introduction of stochastic volatility mod-
els, such as the Heston model (Heston, 1993), of-
fered more flexibility by allowing volatility to vary
as a stochastic process.
Despite improvements like stochastic volatility
in the Heston model (Heston, 1993), traditional
models remain limited in capturing the rapid shifts
and complex dependencies of modern financial mar-
kets (Jones, 2019; Assaf Eisdorfer and Zhdanov,
2022). This has motivated the exploration of adaptive
machine-learning approaches capable of modeling in-
tricate relationships and dynamic patterns in financial
data (Gatheral, 2006; Christoffersen, 2009).
Recurrent architectures, such as Long Short-
Term Memory (LSTM) networks and Gated Recur-
rent Units (GRU), became popular due to their abil-
ity to capture temporal dependencies in sequential
data (Mintarya et al., 2023; Hochreiter and Schmid-
huber, 1997; Yue Liu, 2023). However, these models
encounter scalability challenges when dealing with
long-term dependencies or high-frequency data, of-
ten leading to computational inefficiencies (Miko-
laj Binkowski and Donnat, 2018; Bryan Lim and
Roberts, 2019). While modular and hybrid neural
networks have been employed to integrate financial
indicators and better capture non-linearities, issues of
scalability and interpretability persist (Amilon, 2003;
N. Gradojevic and Kukolj, 2009).
Input
Sequence
Output
Sequence
Encoder
Multi-head ProbSparse Self-attention
Self-attention distilling
Generative-Style Decoder
Multi-head Attention
Masked Multi-head ProbSparse Self-attention
Figure 1: Informer model - concepcual overview. Based
on (Szydlowski and Chudziak, 2024a).
Transformers, initially developed for natural lan-
guage processing (A. Vaswani and Polosukhin, 2017),
introduced self-attention mechanisms that bypass the
limitations of recurrent models, allowing for the cap-
ture of long-term dependencies without the vanish-
ing gradient problem. Szydlowski (Szydlowski and
Chudziak, 2024b; Wawer et al., 2024) applied the
Hidformer model to stock market prediction, demon-
strating its effectiveness in handling long sequences
and capturing complex market patterns. Informer,
introduced by Zhou et al. (H. Zhou and Zhang,
2021) and illustrated in Figure 2, marked a signif-
icant advancement for time-series analysis with its
ProbSparse self-attention mechanism, reducing the
time and memory complexity of processing long se-
quences to O(L logL) for input length L. Wang
et al. (C. Wang and Zhang, 2022) demonstrated In-
former’s application in predicting stock market in-
dices, showcasing its ability to outperform traditional
deep learning models (e.g., CNN, RNN, LSTM) by
effectively capturing relevant information while filter-
ing out noise—a common challenge in financial time
series. Informer’s robust multi-head attention mecha-
nism allowed for the extraction of key features, lead-
ing to significantly higher prediction accuracy, partic-
ularly in short-term forecasting.
While studies have applied Transformer-based ar-
chitectures to option pricing, including the generic
Transformer model used by Guo and Tian (Guo and
Tian, 2022) and Sagen’s investigation of the Temporal
Applying Informer for Option Pricing: A Transformer-Based Approach
1271

Fusion Transformer (TFT) (Sagen, 2024), the appli-
cation of Informer has not been explored in this do-
main. Given Informer’s strengths in long-sequence
modeling and handling high-dimensional data effi-
ciently, this paper seeks to evaluate its potential for
enhancing predictive accuracy and computational ef-
ficiency in the complex landscape of option pricing.
3 MODEL ARCHITECTURE
In this section, we outline the architecture of the
Informer-based model employed for option pricing.
The Informer model is chosen for its ability to handle
long sequences efficiently and capture dependencies
over varying time scales through its unique attention
mechanisms and architectural optimizations (H. Zhou
and Zhang, 2021; C. Wang and Zhang, 2022). This
is essential in financial applications where complex
temporal relationships can influence outcomes signif-
icantly.
Figure 2: Informer model overview. Copied from(H. Zhou
and Zhang, 2021).
3.1 Data and Feature Engineering
For effective model training, we select input features
known to be crucial for option pricing. These features
include the underlying asset price, implied volatility,
time to maturity, strike price, and an indicator for the
type of option (call or put). Each of these factors
provides valuable insights into how option prices re-
spond to market conditions. Volatility measures indi-
cate market uncertainty (Hull, 2006; Kolm and Ritter,
2019), while time to maturity and strike price are fun-
damental in assessing the intrinsic and extrinsic value
of the option (Jones, 2019; Black and Scholes, 1973;
Merton, 1973). Normalization is applied to standard-
ize the data, ensuring all features are on a comparable
scale:
x
norm
t
=
x
t
−x
min
x
max
−x
min
(1)
where x
norm
t
represents the normalized feature value
at time t, and x
max
and x
min
denote the maximum and
minimum feature values, respectively. This approach
keeps all features within the range [0, 1], aiding in
model stability and faster convergence during train-
ing.
The input data is structured as a time-series se-
quence with a moving window approach, where T
x
past data points are used to predict T
y
future option
prices or metrics. This sequential setup helps cap-
ture dependencies over different time horizons and
enables the model to account for short-term fluctua-
tions as well as long-term trends.
3.2 Proposed Model Architecture
The Informer-based model extends the standard
Transformer architecture by incorporating enhance-
ments tailored to the challenges of time-series fore-
casting in financial applications. It consists of two
main components - the encoder and the decoder,
which exchange information through self-attention
mechanisms and encoder-decoder attention modules,
as we can see in Figure 2. This section details each of
these components and the overall data flow and token
construction procedure.
3.2.1 Encoder
The encoder is responsible for extracting meaningful
temporal dependencies from the input sequence.
It includes an embedding layer, a ProbSparse self-
attention mechanism, a feedforward sub-layer, and a
self-attention distilling step to reduce computational
overhead.
Embedding Layer. Each time step in the raw data
is represented as a token, which is a set of features
(e.g., strike price, time to maturity). The embedding
layer projects these tokens into a dense vector space
of fixed dimension, enabling the network to learn
hidden interactions across features.
ProbSparse Self-Attention Mechanism. This atten-
tion mechanism aims to identify and focus on the
most informative queries in the attention calculation,
as illustrated in Figure 3. Instead of computing at-
tention scores for all L queries and keys, it selects a
subset of queries based on the Kullback-Leibler di-
vergence (KLD) between the query distribution and a
ICAART 2025 - 17th International Conference on Agents and Artificial Intelligence
1272

predefined sparse distribution. Formally:
Attention(Q, K, V ) = Softmax

Q
⊤
K
√
d
k

V (2)
where Q, K, V are the query, key, and value matri-
ces, and d
k
is the dimension of the keys. By select-
ing only the top-U queries (with U ≪ L), complexity
is reduced from O(L
2
) to approximately O(L log L),
making the model scalable for long sequences.
Figure 3: Illustration of the ProbSparse Attention mecha-
nism. Adapted from(H. Zhou and Zhang, 2021).
Feedforward Sub-Layer. The output of the attention
sub-layer is passed through a fully connected feedfor-
ward network with a hidden dimensionality D
FF
:
FeedForward(x) = ReLU(W
1
x + b
1
)W
2
+ b
2
, (3)
where W
1
, W
2
are weight matrices, and b
1
, b
2
are
biases. A larger dimension D
FF
allows the model to
capture intricate patterns.
Self-Attention Distilling. To improve efficiency, the
Informer applies a distilling mechanism at the end of
each layer, pooling or downsampling the sequence to
retain only the most critical tokens. Mathematically,
Z
l
= Pooling(X
l
), X
l+1
= SelfAttention(Z
l
),
where X
l
is the layer-l input. This process concen-
trates the model’s capacity on dominant features,
enhancing performance while mitigating overfitting.
Encoder Output. The final encoder output, denoted
by E
t
, is a contextually enriched representation of the
input tokens and will be passed to the decoder for gen-
erating forecasts.
3.2.2 Decoder
The decoder produces the target sequence by lever-
aging both the encoder output and partially known
future labels. It features a self-attention sub-layer,
encoder-decoder attention, and a feedforward
network. Unlike the traditional approach, which
decodes one step at a time, the Informer employs a
generative-style approach to predict all future steps
simultaneously.
Decoder Input Preparation. To provide the de-
coder with partial knowledge of the future horizon,
the model concatenates the most recent T
label
observed
values with placeholder zeros for the T
y
unknown time
steps. This can be expressed as:
D
t
= [y
t−T
label
+1
, . . . , y
t
, 0, . . . , 0 ].
During training, the first portion corresponds to
known labels or ground truth values, while zeros
mark positions to be predicted.
Attention Modules and Feedforward Sub-Layer.
In the decoder, self-attention accounts for dependen-
cies among known and future positions in D
t
, while
encoder-decoder attention utilizes E
t
(the encoder
output) as keys and values to incorporate previously
extracted temporal structure. It also applies a feedfor-
ward sub-layer similar to that in the encoder.
Generative-Style Decoding. Finally, the decoder
produces the entire predicted sequence in one forward
pass:
ˆ
Y
t
= Decoder(E
t
, D
t
).
This approach improves inference speed compared to
autoregressive decoding, which is advantageous for
time-sensitive financial applications.
Decoder Output. The vector
ˆ
Y
t
constitutes the
model’s forecast for the T
y
future time steps.
3.3 Model Workflow
The overall workflow begins by converting each time
step into a token that bundles relevant features. These
tokens are then passed to the embedding layer, which
maps them into a continuous space of dimension
d
model
. The encoder applies ProbSparse self-attention,
feedforward transformations, and self-attention dis-
tilling to capture critical dependencies with reduced
computational overhead. Its final output E
t
, enriched
with temporal context, is transferred to the decoder.
In parallel, the decoder constructs its input D
t
by
combining partially known labels from the predic-
tion window with placeholder zeros. Self-attention
in the decoder identifies dependencies among these
elements, while encoder-decoder attention integrates
signals from E
t
. The generative-style decoding step
then yields a full multi-step forecast in a single
pass, producing
ˆ
Y
t
. This hierarchical design is es-
pecially suited to financial time-series forecasting,
where long-range dependencies and efficient compu-
tation are both critical.
Applying Informer for Option Pricing: A Transformer-Based Approach
1273

4 EXPERIMENTS
The experiments conducted aim to evaluate the
robustness and predictive power of the proposed
Informer-based model in the context of option pric-
ing. A thorough comparison is established using
baseline models that encompass traditional and ma-
chine learning-based methods.
4.1 Dataset and Data Preparation
We use a dataset comprising eight years of histori-
cal option contracts for Apple Inc. (AAPL), sourced
from publicly available financial databases, covering
the period from January 4, 2016, to March 31, 2023.
The dataset includes both call and put options with
varying strike prices, expiration dates, and moneyness
levels, providing a diverse and comprehensive foun-
dation for analysis.
To improve data quality and ensure relevance, the
preprocessing stage included the application of strict
selection criteria. Options with a time-to-maturity
(TTM) below 30 days were excluded, as such short-
term contracts are typically highly volatile and specu-
lative (Heston, 1993). Furthermore, only options with
a moneyness ratio (the ratio of the underlying asset’s
price to the strike price) between 0.6 and 1.3 were
included, as near-the-money options are more liquid
and exhibit more reliable pricing (Bakshi et al., 2000).
Contracts with insufficient data points or low trading
volume were also removed to maintain robustness and
integrity. The dataset is split into training, validation,
and test sets, with 70% of the data allocated for train-
ing, 15% for validation, and the remaining 15% for
testing (Matsunaga and Suzumura, 2019). This split
ensures that the model is evaluated on unseen data,
simulating real-world conditions where future predic-
tions depend on past training.
4.2 Model Configuration and Training
Strategy
The Informer model is configured to handle complex
time-series data with the following parameters: the
input sequence length is set to 30 days (T
x
= 30), and
predictions are made over a 30-day horizon (T
y
= 30).
The architecture includes one encoder layer and two
decoder layers with a label length of 5 days, each
featuring three attention heads. The embedding di-
mension (D
MODEL
) is set to 32, balancing computa-
tional efficiency and model expressiveness. The feed-
forward network dimension is set to 8, with a dropout
rate of 0.06 to prevent overfitting. The model employs
full attention with a factor of 3, suitable for capturing
temporal patterns effectively in financial time-series
data. The training process employs a batch size of
64 and utilizes the Adam optimizer (Kingma and Ba,
2014) with an initial learning rate of 0.0001. Train-
ing proceeds over 300 epochs, with early stopping ap-
plied based on validation loss, using a patience of 30
epochs. A weighted mean squared error (MSE) loss
function is used, prioritizing accuracy across the en-
tire 30-day prediction horizon. Hyperparameters, in-
cluding the number of layers, attention heads, embed-
ding dimension, learning rate, and dropout rate, were
fine-tuned via random search.
4.3 Evaluation Metrics
The performance of the Informer model is evaluated
using a comprehensive set of metrics to ensure a ro-
bust evaluation (Ruf and Wang, 2020):
Prediction Accuracy: The model’s outputs are
compared with the ground truth on the validation set
to evaluate the prediction accuracy. Two commonly
used indicators are employed: Mean Absolute Er-
ror (MAE), which measures the average magnitude
of prediction errors, and Root Mean Squared Error
(RMSE), which emphasizes larger errors to capture
prediction variance. Lower values of both metrics in-
dicate better model performance.
Final-Day Evaluation: We focus on final-day
evaluation because it highlights the model’s ability to
make accurate long-term predictions, which is crucial
for strategic financial decision-making (Kristoufek,
2012). To measure this, we use Direction Accuracy
(DA), which measures the percentage of sequences
where the predicted and actual price changes have
the same direction, and Final-Day MAE, which cal-
culates the MAE between predicted and actual prices
specifically on the last day.
Return Calculation: The trading effectiveness of
the model is evaluated using a simple strategy based
on the predicted price at the end of each sequence.
For a given sequence, if the predicted price ( ˆy
t+30
)
is higher than the starting price (y
t
), a long position
is taken; otherwise, a short position is assumed. The
return for the sequence is calculated as:
R = ln
y
t+30
y
t
×sign( ˆy
t+30
−y
t
) (4)
where y
t+30
is the true price at the prediction horizon,
y
t
is the starting price, and ˆy
t+30
is the predicted price.
The cumulative net value (NV) aggregates returns
across all sequences in the dataset, starting from an
initial value of 1:
NV = 1 +
N
∑
i=1
R
i
(5)
ICAART 2025 - 17th International Conference on Agents and Artificial Intelligence
1274

where N is the total number of sequences.
By combining predictive accuracy metrics (MAE
and RMSE) with trading performance (NV), this eval-
uation framework captures both the statistical preci-
sion and the practical utility of the model in financial
applications.
To benchmark the performance of the Informer-
based model, we compare it against several estab-
lished baseline models, including the Black-Scholes
model, the Heston model, and the simple LSTM-
based model. These models, ranging from traditional
finance to advanced machine learning, help evaluate
how the Informer performs in option pricing, high-
lighting its strengths and areas for improvement.
4.4 Results and Analysis
The results of the experiments demonstrate that the
Informer model consistently outperforms all other
models, both in terms of prediction accuracy and
final-day evaluation metrics.
Table 1: Overall prediction metrics for all models.
Model MAE RMSE
Informer 2.7145 3.6766
LSTM 3.9343 5.0373
Black-Scholes 4.1765 5.3840
Heston 4.1282 5.3565
Figure 4: Comparison of Informer and LSTM predictions
on a longer period for an option contract.
Table 1 highlights the overall prediction metrics,
Mean Absolute Error (MAE) and Root Mean Squared
Error (RMSE). The Informer model achieves the low-
est MAE (2.7145) and RMSE (3.6766) among all
models, demonstrating its superior ability to pre-
dict option prices with high accuracy. The LSTM
model, while a competitive machine-learning ap-
proach, exhibits a significantly higher MAE (3.9343)
and RMSE (5.0373). Traditional models like Black-
Scholes and Heston, despite their widespread use in
finance, perform worse than the machine-learning-
based methods. The Black-Scholes model has a
slightly lower MAE (4.1765) compared to the Hes-
ton model (4.1282), but both models fail to capture
complex market dynamics as effectively as the In-
former. Figure 4 further illustrates the comparative
performance of the Informer and LSTM models on a
longer prediction period, highlighting the Informer’s
ability to track trends more closely.
Table 2: Final-day evaluation metrics for all models.
Model DA (%) Final-Day MAE
Informer 54.43 2.9709
LSTM 52.19 4.0900
Black-Scholes 52.53 4.6880
Heston 51.74 4.6861
Table 2 presents the final-day evaluation met-
rics, including Direction Accuracy (DA) and Final-
Day MAE. The Informer achieves the highest DA
(54.43%) and the lowest Final-Day MAE (2.9709),
showcasing its ability to predict both the direction
and final value of option prices with superior preci-
sion. The LSTM model, while demonstrating a rea-
sonable DA (52.19%), exhibits a higher Final-Day
MAE (4.0900), indicating less reliability in final price
predictions. Among the traditional models, Black-
Scholes performs slightly better than Heston, achiev-
ing a DA of 52.53% compared to 51.74%, but both
models have significantly higher Final-Day MAE val-
ues, exceeding 4.68.
Table 3: Performance of the trading strategy for Apple op-
tions based on final cumulative net value.
Model Net Value
Informer 1.30
LSTM 1.21
Heston 1.15
Black-Scholes 1.14
In trading performance, the Informer achieved the
highest cumulative net value (NV), outperforming all
models, as shown in Table 3. With a final NV of
1.30, the Informer model demonstrates its superior
ability to generate profitable trading strategies by ac-
curately predicting directional movements over a 30-
day horizon. The LSTM model follows with an NV of
Applying Informer for Option Pricing: A Transformer-Based Approach
1275

Figure 5: Example of upward trend Informer prediction for
one option contract.
1.21, while the traditional models, Heston and Black-
Scholes, lag slightly behind with NVs of 1.15 and
1.14, respectively.
As we can see on Figures 5 to 7 the In-
former’s predictions remain stable across different
trend types—upward, downward, and mixed. This
stability highlights the potential of the Informer
model as a valuable tool for investors, providing re-
liable insights to navigate diverse market conditions
effectively.
Figure 6: Example of downward trend Informer prediction
for one option contract.
Figure 7: Example of mixed trend Informer prediction for
one option contract.
5 CONCLUSION AND FUTURE
WORK
Our research demonstrates that the Informer model,
with its specialized attention mechanisms and
generative-style decoder, outperforms traditional
models like Black-Scholes and Heston, as well as re-
current neural networks such as LSTM, in predicting
option prices and capturing long-term dependencies
in financial data. The Informer not only achieved
the lowest MAE and RMSE across all tested models
but also generated the highest cumulative net value
in trading evaluations, outperforming all other mod-
els and demonstrating its practical value in optimizing
trading strategies.
This paper contributes to the field of option pric-
ing by implementing the Informer model for option
trading and evaluating its performance against other
established models.
This study demonstrates the potential of the In-
former model in enhancing option pricing predic-
tions, yet there are several avenues for further ex-
ploration. Future work could involve incorporat-
ing reinforcement learning (RL) strategies to dynami-
cally adjust trading decisions based on model predic-
tions (Szydlowski and Chudziak, 2024a), improving
adaptability in real-time trading environments. Ad-
ditionally, applying the Informer architecture within
a broader portfolio management framework could re-
veal insights into its effectiveness in balancing risk
and return across diverse financial instruments. An-
ICAART 2025 - 17th International Conference on Agents and Artificial Intelligence
1276

other promising direction would be to test and refine
trading strategies based on model outputs, such as
mean-reversion or momentum-based approaches, to
assess the practical profitability and robustness of In-
former in real-world trading applications (Chudziak
and Wawer, 2024).
REFERENCES
A. Vaswani, N. Shazeer, N. P. J. U. L. J. A. G. L. K. and
Polosukhin, I. (2017). Attention is all you need. Ad-
vances in Neural Information Processing Systems, 30.
Amilon, H. (2003). A neural network versus black-scholes:
A comparison of pricing and hedging performances.
Journal of Forecasting, 22(4):317–335.
Assaf Eisdorfer, R. S. and Zhdanov, A. (2022). Maturity
driven mispricing of options. Journal of Financial and
Quantitative Analysis, 57(2):514–542.
Bakshi, G., Cao, C., and Chen, Z. (2000). Pricing and
hedging long-term options. Journal of Econometrics,
94(1):277–318.
Bao, W., Yue, J., and Rao, Y. (2017). A deep learning
framework for financial time series using stacked au-
toencoders and long-short term memory. PLOS ONE,
12(7):1–24.
Black, F. and Scholes, M. (1973). The pricing of options
and corporate liabilities. In Journal of Political Econ-
omy. Journal of Political Economy.
Bollerslev, T. (1986). Generalized autoregressive condi-
tional heteroscedasticity. Journal of Econometrics,
31(3):307–327.
Bryan Lim, S. Z. and Roberts, S. (2019). Enhancing
time-series momentum strategies using deep neural
networks. The Journal of Financial Data Science,
1(4):19–38.
C. Wang, Y. Chen, S. Z. and Zhang, Q. (2022). Stock market
index prediction using deep transformer model. Ex-
pert Systems with Applications, 208:118128.
Christoffersen, P. F. (2009). Elements of Financial Risk
Management. Academic Press, San Diego, CA.
Chudziak, J. A. and Wawer, M. (2024). Elliottagents: A
natural language-driven multi-agent system for stock
market analysis and prediction. In Proceedings of the
38th Pacific Asia Conference on Language, Informa-
tion and Computation, Tokyo, Japan, (in press).
Cox, J. C., Ross, S. A., and Rubinstein, M. (1979). Option
pricing: A simplified approach. Journal of Financial
Economics, 7:229–263.
Gatheral, J. (2006). The Volatility Surface: A Practitioner’s
Guide. Wiley, Hoboken, NJ.
Guo, T. and Tian, B. (2022). The study of option pricing
problems based on transformer model. In Proceedings
of the IEEE Conference. IEEE.
H. Zhou, S. Zhang, J. P. S. Z. J. L. H. X. and Zhang, W.
(2021). Informer: Beyond efficient transformer for
long sequence time-series forecasting. In Proceedings
of the AAAI Conference on Artificial Intelligence, vol-
ume 35, pages 11106–11115.
Heston, S. L. (1993). A closed-form solution for options
with stochastic volatility with applications to bond and
currency options. The Review of Financial Studies,
6(2):327–343.
Hochreiter, S. and Schmidhuber, J. (1997). Long short-term
memory. Neural Computation, 9:1735–1780.
Hull, J. C. (2006). Options, Futures, and Other Derivatives.
Pearson Prentice Hall.
Jones, C. M. (2019). Volatility estimation and financial mar-
kets. Journal of Financial Markets, 42:12–36.
Kingma, D. P. and Ba, J. (2014). Adam: A method for
stochastic optimization. arXiv, 1412.6980.
Kolm, P. N. and Ritter, G. (2019). Dynamic replication
and hedging: A reinforcement learning approach. The
Journal of Financial Data Science, 1(1):159–171.
Kristoufek, L. (2012). Fractal markets hypothesis and
the global financial crisis: Scaling, investment hori-
zons and liquidity. Advances in Complex Systems,
15(06):1250065.
Matsunaga, D. and Suzumura, T. (2019). Long-term rolling
window for stock market predictions. arXiv preprint,
1911.05009.
Merton, R. C. (1973). Theory of rational option pricing. The
Bell Journal of Economics and Management Science,
4(1):141–183.
Mikolaj Binkowski, G. M. and Donnat, P. (2018). Au-
toregressive convolutional neural networks for asyn-
chronous time series. In International Conference on
Machine Learning, pages 580–589. PMLR.
Mintarya, L. N., Halim, J. N., Angie, C., Achmad, S., and
Kurniawan, A. (2023). Machine learning approaches
in stock market prediction: A systematic literature re-
view. Procedia Computer Science, 216:96–102. 7th
International Conference on Computer Science and
Computational Intelligence 2022.
N. Gradojevic, B. G. and Kukolj, S. (2009). Option pric-
ing with modular neural networks. Neural Networks,
22(5):716–723.
Ruf, J. and Wang, W. (2020). Neural networks for option
pricing and hedging: a literature review.
Sagen, L. K. (2024). Applied option pricing using trans-
formers. Master’s thesis, Norwegian University of
Science and Technology (NTNU).
Szydlowski, K. L. and Chudziak, J. A. (2024a). Toward
predictive stock trading with hidformer integrated into
reinforcement learning strategy. In Proceedings of the
36th International Conference on Tools for Artificial
Intelligence (ICTAI 2024), Herndon, VA, USA.
Szydlowski, K. L. and Chudziak, J. A. (2024b).
Transformer-style neural network in stock price fore-
casting. In Proceedings of the 21th International Con-
ference on Modeling Decisions for Artificial Intelli-
gence (MDAI 2024), Tokyo, Japan.
Wawer, M., Chudziak, J. A., and Niewiadomska-
Szynkiewicz, E. (2024). Large language models and
the elliott wave principle: A multi-agent deep learn-
ing approach to big data analysis in financial markets.
Applied Sciences, 14(24).
Yue Liu, X. Z. (2023). Option pricing using lstm:
A perspective of realized skewness. Mathematics,
11(2):314.
Applying Informer for Option Pricing: A Transformer-Based Approach
1277