ARTICLE

https://doi.org/10.1057/s41599-023-01836-2 OPEN

Quantum Monte Carlo simulations for estimating

FOREX markets: a speculative attacks experience
David Alaminos1 ✉, M. Belén Salas2,3 & Manuel Á. Fernández-Gámez2,3

The foreign exchange markets, renowned as the largest financial markets globally, also stand
out as one of the most intricate due to their substantial volatility, nonlinearity, and irregular
1234567890():,;

nature. Owing to these challenging attributes, various research endeavors have been
undertaken to effectively forecast future currency prices in foreign exchange with precision.
The studies performed have built models utilizing statistical methods, being the Monte Carlo
algorithm the most popular. In this study, we propose to apply Auxiliary-Field Quantum
Monte Carlo to increase the precision of the FOREX markets models from different sample
sizes to test simulations in different stress contexts. Our findings reveal that the imple-
mentation of Auxiliary-Field Quantum Monte Carlo significantly enhances the accuracy of
these models, as evidenced by the minimal error and consistent estimations achieved in the
FOREX market. This research holds valuable implications for both the general public and
financial institutions, empowering them to effectively anticipate significant volatility in
exchange rate trends and the associated risks. These insights provide crucial guidance for
future decision-making processes.

1 Department of Business, Universitat de Barcelona, Barcelona, Spain. 2 Department of Finance and Accounting, Universidad de Málaga, Málaga, Spain.
3 Cátedra de Economía y Finanzas Sostenibles, Universidad de Málaga, Málaga, Spain. ✉email: [email protected]

HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | (2023)10:353 | https://doi.org/10.1057/s41599-023-01836-2 1

Content courtesy of Springer Nature, terms of use apply. Rights reserved.
ARTICLE HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | https://doi.org/10.1057/s41599-023-01836-2

F
Introduction
oreign exchange markets, commonly known as FOREX, of 0.72. Ince et al. (2016) used ordinary least squares (OLS) to
hold the distinction of being the largest financial market estimate theoretical models of the FOREX market for data on US
globally. They are renowned for their inherent complexity dollar-Swiss franc and US dollar–Japanese Yen exchange rates.
and volatility, characterized by unpredictable behavior and sig- They got an overall standard deviation superior to 1.06 in their
nificant fluctuations in exchange rates (Wei and Zhu, 2022). The estimates. Finally, Hauzenberger and Huber (2019) applied the
ability to forecast currency pair movements is of great importance Markov process of Time-Varying transition to forecast the
in the field of FOREX. Over the past years, researchers worldwide FOREX market for Japan, Norway, Australia, Switzerland,
have devoted considerable attention to the FOREX market, driven Canada, Sweden, South Korea, and the UK relative to the US
by its vulnerable characteristics (Islam et al., 2020; Ayitey et al., dollar. They reached an average deviation of 0.94. In their study,
2023). As a result, various types of research have been conducted Tigani et al. (2022) developed a model utilizing Gaussian kernel
to accurately predict future prices of FOREX currencies. density and Monte Carlo simulation to forecast the volatility
The models to estimate FOREX markets have been of great patterns of the EUR/USD currency pair on an hourly basis. The
importance in international finance in recent decades, occupying proposed model holds potential applications in the financial
an important effort by academics and professionals to forecast the market, particularly in algorithmic trading, where the Monte
future values of currencies, considered a difficult task (Cheung Carlo method is employed to estimate integrals within this fra-
and Erlandsson, 2005; Kamal et al., 2012; Hauzenberger and mework. The authors suggest that future research should include
Florian, 2019; Islam et al., 2020). These models are usually esti- a comparative analysis of the accuracy of various sampling
mated by statistical methodologies, both classic regression models methods, such as Metropolis-Hastings, Gibbs sampling, and
such as ordinary least squares (OLS) or factor models up to Markov Chain Monte Carlo, to further enhance the under-
Monte Carlo models, which have given great precision results and standing and effectiveness of the model.
are the most commonly used (Giacomini and Rossi, 2010; Therefore, many researchers have targeted their studies on the
Jaworski, 2018). These models have been used in different var- FOREX market by using methods ranging from statistical
iants, with Metropolis-Hastings and Sequential filtering being the methods to deep learning. However, because volatility behavior in
Monte Carlo designs most used in Economics and Finance. Such the foreign exchange market has important implications for
models are assessed based on the standard deviation result modeling and calculating risk in these markets, future researchers
achieved, the size of the sample, as well as the complexity. The will need to use innovative and improved approaches to analyze
literature shows the different fit results after applying different the impact and forecasting of this volatility behavior (Kamal et al.,
statistical techniques and Monte Carlo algorithms for these 2012; Islam et al., 2020; Chinthapalli, 2021). Hence, the Quantum
Currency Market models. For example, models constructed with Monte Carlo method emerges as a sophisticated quantum
linear statistical techniques, such as Ordinary Least Squares, lin- approach, offering a dependable solution to capture the men-
ear regression methods, and Factor models, have provided an tioned volatility. The aim is to simulate speculative attack models
adjustment of 0.93–1.84 standard deviation (Rossi, 2013; Park in order to provide further insights into potential events that may
and Park, 2013; Byrnea et al., 2016; Ince et al., 2016; Serjam and transpire within FOREX markets. Furthermore, Islam et al.
Sakurai, 2018), but their adjustments are between 1.26–2.11 when (2020) conclude that future research on FOREX market predic-
constructed with small samples (Rossi, 2013; Jacob and Uusküla, tion can be through robust and accurate approaches, like GRU,
2019). On the other hand, other more advanced statistical models, Monte Carlo methods, Kohonen’s self-organizing neural network,
especially non-linear models, such as Vector Autoregression, modular neural network, and numerous additional algorithms,
time-varying parameter models, and Error Correction models. which are not yet completely studied in this domain.
With large samples, these models have obtained a precision To fill this gap in the previous literature and to solve these
between 0.58–0.83 (Clements and Lan, 2010; Kavtaradze and problems of accuracy of the existing methods for estimating
Mokhtari, 2018; Taveeapiradeecharoen et al., 2019) while in the FOREX markets, our paper analyses USD/EUR and USD/JPY
case of small samples, the results have been 0.75–1.24 (Cheung exchange rates in the period 2013–2021. This work compares
et al., 2019; Colombo and Pelagatti, 2020; Rubaszek and Ca’ three Monte Carlo techniques, Markov Chain Monte Carlo,
Zorzi, 2020). Hence, it becomes apparent that econometric Sequential Monte Carlo, and Auxiliary-Field Quantum Monte
methods demonstrate a level of accuracy within the range of Carlo (AFQMC), with the AFQMC technique being the best
0.40–0.83 when performing simulations with sample sizes performer. This technique has already demonstrated its metho-
exceeding 100 observations. However, these methods do not yield dological superiority in other areas in carrying out accurate
a significantly low standard deviation when applied to small sampling with few observations and data distributions (Giacomini
samples (Park and Park, 2013; Beckmann and Schuessler, 2016). and Rossi, 2010; Rossi, 2013; Kolasa et al., 2017; Jaworski, 2018).
Giacomo and Rossi (2010) applied the Fluctuation test and the Our results show a more robust estimation according to accuracy,
One-Time Reversal test for the monthly period of 1973–2008 both in-sample, and out-of-sample estimations, a better behavior
with the data of the dollar/British pound exchange rate, showing a with small and irregular samples compared to the conventional
standard deviation higher than 0.49 for both econometric tech- regression, deep learning, and Monte Carlo methods. Also, these
niques. Park and Park (2013) conducted a comparative study Monte Carlo models need less time to make the estimates, espe-
including the Japanese Yens, Swiss francs, Canadian dollars, and cially in the case of the AFQMC method. Hence, these results also
UK pounds per the US dollar, for the quarterly period of show great computing of popular FOREX market models that
1973–2010, applying the time-varying cointegration coefficients previous literature showed as difficult to estimate accurately,
method. They showed a mean square error above 0.62. For its reducing the instability of the previous literature (Jaworski, 2018;
part, Beckmann and Schuessler (2016) demonstrate in a Monte Dash, 2018; Cheung et al., 2019; Hauzenberger and Huber, 2019;
Carlo simulation how a time-varying parameter model such as Colombo and Pelagatti, 2020). These results can be very useful in
ours, namely one that provides varying degrees of temporal their application in FOREX market models and in other models in
variation in the coefficients, can be very suitable for retrieving Financial Econometrics that help the valuation challenges of
samples in the database. Byrne et al. (2016) applied the Time- financial professionals and other related interest groups.
Varying Parameters method, for a sample of several OECD Our research has been motivated by the impact of recent
currencies per US dollar, obtaining an overall standard deviation economic events on FOREX markets, including the market

2 HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | (2023)10:353 | https://doi.org/10.1057/s41599-023-01836-2

Content courtesy of Springer Nature, terms of use apply. Rights reserved.
HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | https://doi.org/10.1057/s41599-023-01836-2 ARTICLE

collapse in 2020 following a decade of global economic growth more quickly to changes in the market and to exploit dealing
and recovery from the 2007 to 2008 financial crisis. The onset of occasions. Therefore, we apply the Quantum Monte Carlo
the COVID-19 pandemic in early 2020 further exacerbated the method in our study because most of the scenarios that define the
situation, resulting in significant economic downturns and financial field involve a great deal of computational difficulty, and
increased uncertainty. As a result, traders now require a thorough thus lend themselves to quantum computation. It is a widely used
understanding of market behavior before making trading deci- technique in science, with applications in physics, chemistry,
sions (Ilić and Digkoglou, 2022). Monte Carlo simulation is a engineering, and finance (Orus et al., 2019). Monte Carlo simu-
widely employed method in business and finance to assess lation facilitates visualizing all possible outcomes of decisions,
portfolios and investments by simulating various uncertain fac- inclusive of the actual probabilities of each occurring. This allows
tors that influence their value, thereby establishing the distribu- the impact of risk to be quantitatively assessed, enabling more
tion of their value across a range of possible outcomes. Extensive accurate forecasting and eventually better decision-making under
literature highlights the numerous applications of Monte Carlo uncertainty (Sikora et al., 2019).
methods in finance, encompassing security valuation, risk man- Secondly, our model helps to identify possible currency
agement, portfolio optimization, model calibration (Staum, 2009), movements in the foreign exchange market through Monte Carlo
derivatives valuation (Rebentrost et al., 2018), theoretical research simulation, thus avoiding a speculative attack. In this way, we can
(Creal, 2012), calculations (Asmussen, 2018), and stock market determine how a financial crisis could affect the price of cur-
forecasting (Parungrojrat and Kidsom, 2019). Moreover, the rencies. In the event that the current global virus situation wor-
Monte Carlo method is often combined with other techniques, sens or similar crises emerge in the future, it becomes increasingly
such as the Mali calculation (Fournié et al., 2001). In conclusion, important to mitigate the adverse effects of currency price vola-
as unexpected crises like the recent situation in Ukraine unfold, tility. In this regard, the insights gained from the implementation
predictive methods gain significance for investors and researchers of the Monte Carlo simulation hold great relevance. Conse-
(Ilić and Digkoglou, 2022). quently, our model serves as a valuable tool for the analysis of
The present research differs from others in that it compares complex problems that are difficult to tackle analytically, as well
various Monte Carlo techniques in FOREX markets prediction. as for their subsequent forecasting.
Most of the models in previous studies have been dominated by The paper is organized as follows: the section “Literature
statistical techniques such as ordinary least squares, quantile review” presents a literature review, discussing previous research
regression, and recently neural network techniques. Monte Carlo in the field. Section “FOREX markets fundamentals” introduces
simulation holds a significant position as one of the key algo- the exchange rate dynamics models used in the analysis. In the
rithms in finance and numerical computational science, playing a section “Estimation methods”, the methodologies employed for
crucial role in the realm of risk management being able to easily the estimations are explained. Section “Empirical methods” pre-
deal with high-dimensional problems such as volatility (Guo, sents the data utilized in the study and presents the results and
2022). Aşırım et al. (2023) concluded that the prediction of the findings. Finally, the section “Conclusion” concludes by sum-
foreign exchange market is quite difficult due to this volatility. marizing the key insights and conclusions derived from the
However, the Monte Carlo method could help risk management research.
in this prediction. In addition, Jarusek et al. (2022) recommend
applying the Monte Carlo method to manage market chaos and
volatility, since prediction in the foreign exchange market is Literature review
difficult to perform in a complex dynamic system. Applying the In recent years, there has been a strong focus on predicting the
Monte Carlo method to trading systems allows for better risk volatile FOREX market, leading researchers to explore various
analysis and risk management. The Monte Carlo method also methods. Among these, statistical analysis techniques have gained
helps to estimate when a system has stopped working, to know popularity, with researchers employing algorithms such as
the characteristics of the system, and therefore to understand regression, decision trees, trading rules, support vector regression
what we can expect from its performance (Aşırım et al., 2023). (SVR), and fuzzy systems (Dymova et al., 2016; Achchab et al.,
We provide at least two additional contributions to the lit- 2017). Raimundo and Okamoto (2018) contributed to this area by
erature. First, the application of AFQMC methodology has proposing a hybrid model for FOREX rate prediction. Their
increased the precisión of the FOREX market model. Our find- approach involved combining wavelet models with support vector
ings indicate stronger accuracy than previous studies, both in- regression (SVR). They leveraged the discrete wavelet transform
sample, and out-of-sample estimation, as enhanced efficiency (DWT) to extract relevant information from the FOREX dataset,
with short and irregular samples versus conventional regression, which was then used as input for SVR to forecast currency prices.
deep learning, and Monte Carlo methods. FOREX markets are To assess the accuracy of their hybrid model, they compared its
considered to be highly complex and volatile, with important performance to that of traditional autoregressive integrated
implications for modeling and calculating risk in these markets. moving average (ARIMA) and autoregressive fractionally inte-
The Quantum Monte Carlo approach delivers a sophisticated grated moving average (ARFIMA) models. Evaluation metrics
quantum analysis and is a credible way to measure such volatility, such as root mean square error (RMSE) and mean absolute error
the aim is to formulate speculative attack models to generate (MAE) were employed. The results demonstrated the superiority
further details regarding the potential developments in the of their hybrid system over the traditional models, underscoring
FOREX market. Despite the complexity of the new quantum its effectiveness in predicting FOREX rates.
method, we offer a new possibility that other authors and prac- In their respective studies, Serjam and Sakurai (2018),
titioners can develop in the future. So, our research has significant Taveeapiradeecharoen et al. (2019), and Thu and Xuan (2018)
implications for financial institutions and governments, given the explored different approaches for predicting foreign exchange
increasing relevance of advanced risk management. According to (FOREX) rates. Serjam and Sakurai utilized linear kernel support
Rebentrost et al. (2018), quantum computers reveal the hope of a vector regression (SVR) on historical data from major currency
substantial improvement in the speed of these computations. pairs, employing previous time frames as features. They dis-
One-day computations should be shortened to significantly covered a profitable rule, profitability inversion, applicable to
smaller time scales, making real-time risk analysis possible. This systems with unique characteristics. Taveeapiradeecharoen,
near real-time assessment could enable the entity to respond Chamnongthai, and Aunsri proposed a model based on

HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | (2023)10:353 | https://doi.org/10.1057/s41599-023-01836-2 3

Content courtesy of Springer Nature, terms of use apply. Rights reserved.
ARTICLE HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | https://doi.org/10.1057/s41599-023-01836-2

compressed vector autoregression, reducing FOREX data using a et al. (2019) introduced a hybrid model that combined the
random compression technique and employing Bayesian model GARCH model with a neural network to forecast foreign
averaging (BMA). Their model outperformed the existing exchange currency price volatility using the EUR/USD dataset,
benchmark for six currency pairs. Thu and Xuan introduced an leading to improved prediction accuracy. Finally, Fan et al. (2021)
SVM-based model for EUR/USD prediction, comparing different investigated the correlation between Taiwan Weighted Stock and
metrics and observing a significant disparity in performance Google Trends, demonstrating the superiority of neural networks
between the Gaussian radial basis function (RBF) and polynomial over support vector machines and decision trees in their machine
models. Additionally, they demonstrated a threefold increase in learning and trend search experiments.
profit rate using the SVM model compared to the conventional Recently, there has been a notable surge of interest in research
transaction method. focused on the FOREX markets, with researchers exploring
Das et al. (2019) proposed a hybrid system that combined an diverse methodologies to enhance market prediction capabilities.
online sequential model of extreme learning machine (ELM) with From statistical learning to deep learning, different models and
the krill herd (KH) optimization technique. Their system was hybrid approaches have emerged. However, in the field of eco-
compared with a recurrent backpropagation neural network nomics and finance, Monte Carlo techniques have been tradi-
(RBPNN) and the ELM algorithm using four currency pairs tionally applied and are considered more reliable. Therefore, in
(SGD/INR, YEN/INR, USD/EUR, and USD/INR). The evaluation our currency market prediction model, which focuses on data-
based on RMSE, MAPE, and Theil’s U metrics showed that their intensive currency quotes, we have employed Monte Carlo
system performed the best with error rates of 0.071646, 0.10375, simulation. This method allows us to consider all possible out-
and 0.00014516, respectively. However, in terms of MAE and comes of our decisions and provides the actual probabilities of
ARV performance evaluations, their proposed model did not each outcome occurring. By quantitatively assessing the impact of
provide the best results. For MAE evaluation, the online risk, we can achieve more accurate forecasting and make better
sequential model of the extreme learning machine performed the decisions under conditions of uncertainty (Sikora et al., 2019).
best, and for ARV evaluation, the krill herd with the extreme
learning machine (KH-ELM) achieved the best results. In another
study, Das et al. (2020) introduced a forecasting model that FOREX markets fundamentals
combined the Jaya optimization technique with extreme learning In this study, we estimated various exchange rate dynamics
machines for predicting currency exchange rates. They used two models, including uncovered interest rate parity, purchasing
currency pairs (USDEUR and USDINR) and evaluated their power parity, speculative pressure indexes (Cuthbertson and
model’s performance using MAPE, ARV, Theil’s U, and MAE. Nitzsche, 2004), sticky-price monetary, behavioral equilibrium
The performance comparison with other models based on ELM, exchange rate, and Taylor rule fundamentals models (Cheung
NN, and FLANN showed that ELM exhibited the best optimi- et al., 2019). To evaluate these models, we drew a sample of data
zation. Their evaluation data indicated that ELM DE provided the and constructed different-size designs. Uncovered interest rate
lowest error for MAPE evaluation, while ELM TLBO, ELM PSO, parity and purchasing power parity are the most commonly used
and ELM Jaya achieved the best results for MAE, ARV, and models for exchange rate dynamics estimation (Rossi, 2013). To
Theil’s U, respectively. Chou and Truong (2019) tested the estimate speculative attack scenarios, we used the model devel-
effectiveness of optimization using a benchmark function. They oped by Eichengreen et al. (1994), which has served as a basis for
examined daily CAN/USD rates and 4-h EUR/USD closing pri- subsequent models of speculative attacks (Braga de Macedo and
ces, which resulted in mean absolute percentage errors of 0.2532% Lempinen, 2013; Wang et al., 2020; Valchev, 2020). Braga de
and 0.169%, respectively. Their forecast system achieved an Macedo and Lempinen (2013) developed a general equilibrium
accuracy rate of 89.8–99.7%, demonstrating improved forecasting model to analyze and compare exchange rate adjustments to
accuracy compared to previous models for the CANUSD cur- those presented in Dornbusch (1976) and Kouri (1978). They
rency pair. The error rate of SMOF surpassed the baseline sliding concluded that only the aggregation of holdings of assets sus-
window model, ranging from 20.8% to 29.9%. ceptible to speculative expectations is relevant to their model, and
The utilization of neural networks has been instrumental in they studied three different cases.
time series prediction, particularly in the dynamic FOREX mar- To evaluate the efficacy of these models, we conducted Monte
ket, leveraging the power of hidden neurons for accurate fore- Carlo simulations, generating 10,000 continuous-time trajectories
casting. De Almeida et al. (2018) proposed an innovative FOREX spanning a 30-year period. We employed the 90% variance range
trading model that combined support vector machines (SVM) as a metric to assess performance. Our findings revealed that the
with the genetic algorithm (GA) to optimize trading rules, Dornbusch formulation exhibited a variance range of 200%
resulting in an impressive return on investment of 83% during around the average, which was reduced to 100% in the Kouri case
their testing. Colombo and Pelagatti (2020) conducted simula- and further narrowed down to 20% in the general equilibrium
tions using regularized regression splines, random forest (RF), scenario. Valchev (2020) proposed a new exchange rate deter-
and SVM on data from advanced economies, with SVM exhi- mination model that is robust to the evidence that the Uncovered
biting superior precision, boasting a mean error of ~0.4. Sun et al. Interest Parity reverses its direction at longer horizons. They
(2019) introduced a hybrid SVM method that incorporated a showed that the feedback to the momentum of equilibrium
neural network and applied it to forecast US dollar exchange rates output is not monotonic due to the interplay of fiscal and
against major currencies, achieving mean error ranges of 0.31–1.7 monetary policies, which fits well with the dynamics of exchange
for the period from 2011 to 2017. Ni et al. (2019) proposed a rates. In their study, Wang et al. (2020) developed an early
model for FOREX time series prediction utilizing the CRNN warning system to forecast turbulence in the Shanghai Stock
method, which combined a convolutional neural network and Exchange Composite Index. They employed a SWARCH model
recurrent neural network, demonstrating superior performance combined with a long short-term memory (LSTM) technique.
compared to LSTM and CNN models based on RMSE evalua- The researchers found that LSTM demonstrated robustness, as
tions. Cao et al. (2020) harnessed the power of deep learning, evidenced by its high accuracy of 96.4% and an average forecast
specifically deep long-short memory (LSTM), to forecast the horizon of 2.8 days. Furthermore, LSTM outperformed all base-
exchange rate between the US dollar and the Chinese yuan, line models consistently throughout the evaluation process,
achieving a remarkable precision level of up to 75%. Hajizadeh indicating its stability and effectiveness in predicting turbulence.

4 HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | (2023)10:353 | https://doi.org/10.1057/s41599-023-01836-2

Content courtesy of Springer Nature, terms of use apply. Rights reserved.
HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | https://doi.org/10.1057/s41599-023-01836-2 ARTICLE

They also suggested future directions for investigation, such as respectively. The PPP assumption can be represented by setting
expanding the number of explanatory variables and incorporating the parameters α and β to 0 and 1, respectively, similar to the
other deep-learning mechanisms. uncovered interest rate parity (UIRP) model (Ismailov and Rossi,
We also estimated the second-generation speculative attack 2018). Therefore, the PPP can be expressed as follows:
model of Flood and Marion (1997) using Monte Carlo estimation  
algorithms, including the Metropolis-Hastings, Sequential Monte st ¼ α þ β pt  p*t þ εt ; where α ¼ 0 and β ¼ 1 ð2Þ
Carlo (SMC) algorithms, and the novel AFQMC. In assessing the
Previous studies have raised concerns about the predictive
performance of our models, we utilized classification accuracy
power of the purchasing power parity (PPP) model, as its
measured in percentage using both in-sample and out-sample
performance is not significantly superior to that of the random
data. Additionally, we employed residual measurements such as
walk model in short-term forecasts. The random walk model
root-mean-square error (RMSE) and mean absolute percentage
demonstrates stronger forecasting ability in the short term, while
error (MAPE) to further evaluate the quality of our predictions.
the PPP model tends to outperform the random walk model in
This study extends the existing theoretical models in the lit-
longer time horizons. However, the gap between the two models
erature by incorporating well-established concepts such as
diminishes considerably when it comes to short-term predictions
uncovered interest rate parity, purchasing power parity, beha-
(Cheung and Erlandsson, 2005; Alquist and Chinn, 2008).
vioral equilibrium exchange rate, sticky price monetary, and
Taylor rule fundamentals. These theoretical foundations have
been previously discussed in works by Rossi (2013) and Ismailov Behavioral equilibrium exchange rate (BEER) model. The
and Rossi (2018), providing a strong theoretical basis for our behavioral equilibrium exchange rate (BEER) model is a theore-
research. Additionally, we evaluate various estimation techniques tical framework used to calculate the equilibrium exchange rate of
in the context of speculative attack models to simulate stress and a currency based on its economic fundamentals. The model
volatility scenarios, such as sudden currency value drops. operates on the principle that the exchange rate is determined by
the supply and demand of the currency in the foreign exchange
market. By taking into account various economic variables such
Uncovered interest rate parity (UIRP). Fisher (1962) introduced
as inflation, interest rates, productivity, and trade balance, the
the concept of uncovered interest rate parity (UIRP), which
BEER model aims to estimate an exchange rate that aligns with a
explains the relationship between interest rates and changes in the
country’s economic fundamentals. The underlying assumption is
relative value of currencies. According to UIRP, in an ideal
that the exchange rate will adjust to rectify any imbalances in the
situation where the nominal exchange rate is represented by St,
economy and achieve long-term equilibrium (Barbosa et al., 2018;
investors can buy 1 = St units of foreign bonds using one unit of
Demir and Razmi, 2022).
their domestic currency. Here, St denotes the price of the foreign
The BEER model is becoming increasingly popular among
currency in terms of the domestic currency. Under the assump-
policymakers and economists as it offers a framework for
tion that a foreign bond yields one unit plus the foreign interest
assessing the misalignment of a currency’s exchange rate relative
rate i*t+h between time t and t + h, the expected return of the
to its fundamentals. It has been utilized to identify cases of
foreign investment, converted to the home currency, should be
currency overvaluation or undervaluation and guide policy-
equal to the return of the home bond (1 + it+h), assuming no
makers in implementing appropriate policy measures to restore
transaction costs and no-arbitrage. This can be expressed as
the currency’s equilibrium exchange rate. A common standard
(1 + it+h) Et (St+h = St) = 1 + it+h, where Et(.) represents the
definition of this model is
expectation at time t. The parameters alpha and beta in this
equation have theoretical values of 0 and 1, respectively (Ismailov st ¼ β0 þ ^pt þ β1 ω
^ t þ β2^rt þ β3 gdebtt þ β4 tott þ β5 nfat þ ut
and Rossi, 2018). Thus, the equation represents the uncovered ð3Þ
interest rate parity (UIRP).


 In this framework, the logarithm of the price level (CPI),
1 þ i*tþh Et Stþh ¼ St ¼ 1 þ itþh ð1Þ represented as p, is a key factor. Other variables included are ω,
which represents the ratio of the price of non-tradable goods, r,
Using Eq. (1), Meese and Rogoff (1988) conducted a study to denoting the real interest rate, gdebt, indicating the ratio of public
predict out-of-sample real exchange rates by utilizing real interest debt to GDP, tot, representing the logarithm of the trade terms,
rate differentials. Their findings revealed that this approach and nfa, signifying foreign net assets. The model incorporates
outperformed a random walk model. Similarly, research by various elements such as the Balassa–Samuelson effect, which
(Cheung and Erlandsson, 2005) and Alquist and Chinn (2008) takes into account the relative price of non-tradable goods, the
supported the superiority of the uncovered interest rate parity real interest differential model, which considers the difference in
(UIRP) over a random walk model for extended time horizons, real interest rates, a foreign exchange premium associated with
although the improvement in yield was not statistically the public debt stock, and additional portfolio balance effects
significant. resulting from the net foreign asset exposure of the economy
(Barbosa et al., 2018).
Purchasing power parity (PPP). The concept of purchasing Models built upon this framework are frequently employed to
power parity (PPP) is based on the principle that the price of a ascertain the medium-term reference horizon at which foreign
basket of goods should be the same in different economies when currencies are projected to stabilize, particularly in the context of
expressed in a single currency. This implies that the purchasing analyzing monetary policies. Market professionals routinely
power of a unit of money should be equivalent across countries. utilize this technique to gauge the extent of divergence exhibited
Various interpretations of the PPP principle exist, but a common by currencies from their equilibrium values.
approach involves comparing price levels in one country with
those in a foreign country after converting the prices into a Sticky price monetary (SPM) model. The sticky price monetary
common currency. This concept has been discussed in prior (SPM) model is a theoretical framework employed in the field of
works by Giacomini and Rossi (2010) and Rossi (2013). In the FOREX to provide insights into the relationship between mone-
context of a commodity price (CP) index, the prices in the home tary policy changes and currency exchange rates. In this model,
country and the foreign country are denoted as pt and pt*, prices in the economy are characterized as “sticky,” meaning they

HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | (2023)10:353 | https://doi.org/10.1057/s41599-023-01836-2 5

Content courtesy of Springer Nature, terms of use apply. Rights reserved.
ARTICLE HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | https://doi.org/10.1057/s41599-023-01836-2

adjust slowly to changes in supply and demand. According to the central bank will lower interest rates to stimulate economic
SPM model, adjustments in monetary policy impact the nominal growth and raise the exchange rate (Rossi, 2013).
interest rate, which, in turn, influences the demand for the The Taylor rule fundamentals model offers insights into how
domestic currency. When the nominal interest rate increases, economic fundamentals, interest rates, and exchange rates are
there is upward pressure on the demand for the domestic cur- interrelated in the foreign exchange market. It suggests that
rency as investors seek higher returns. This increased demand central banks adjust their policy rates based on deviations in
leads to an appreciation of the domestic currency’s exchange rate inflation and output from their target levels. Moreover, the model
(Auray et al., 2019). takes into account the deviation of the exchange rate from its
In the short run, the sticky price phenomenon implies that equilibrium value, which is influenced by underlying fundamental
prices in the economy do not instantaneously respond to changes variables. Mathematically, the Taylor rule fundamentals model
in supply and demand. Consequently, even when the exchange can be represented as follows:
rate appreciates, the prices of goods and services do not
immediately decrease to reflect the currency appreciation. This stþk  st ¼ β0 þ β1b ^t þ ut
~yt þ β2 π ð5Þ
situation renders the country’s exports relatively more expensive being yet the output gap, st+k is the expected exchange rate at time
and less competitive in the global market, resulting in a decline in t + k, st represents the spot exchange rate at time t, β0 is the
demand for its goods and a reduced demand for domestic intercept or constant term of the model. β1 is the coefficient of the
currency. As a consequence, the exchange rate experiences output gap variable, which represents the sensitivity of the
depreciation. However, in the long run, prices gradually adjust exchange rate to changes in the output gap. β2 denotes the
to the new exchange rate, restoring the economy to its coefficient of the inflation variable, which represents the
equilibrium state (Chadwick et al., 2015). sensitivity of the exchange rate to changes in inflation. And ut
Over time, prices will eventually adapt to the new exchange is an error term.
rate, allowing the economy to reach its equilibrium state.
However, in the short run, according to the sticky price monetary Speculative attacks model. The identification of currency crises
(SPM) model, changes in monetary policy can cause fluctuations should not be limited to the exchange rate regime or changes in
in the exchange rate. This is primarily due to the sluggish the nominal exchange rate alone. It is possible that a regime
adjustment of prices in the economy, which do not immediately change does not reflect the true reasons behind a country’s
respond to changes in monetary conditions (Auray et al., 2019). decision to maintain the current level of its currency’s exchange
The sticky-price monetary model provides a fundamental rate. Economic development or political institutions’ decisions to
understanding of how flexible exchange rates behave. The model enter a monetary union may increase currency price volatility.
can be described as follows: Additionally, frustrated speculative attacks may occur, where an
excessive increase in demand for foreign exchange does not yield
^ t þ β2 ^yt þ β3^it þ β4 π
st ¼ β0 þ β1 m ^ t þ ut ð4Þ the expected benefit. Monetary authorities have different options
to cover this demand, including adjusting the exchange rate,
interest rates, and foreign exchange reserves (Alaminos et al.,
in which st represents the exchange rate at time t, m denotes the
2022a).
logarithm of money at time t, y states the logarithm of real GDP
Furthermore, speculative attacks that fail to achieve their
at time t, i and π show the interest rate and inflation rate at time t,
intended outcome can also contribute to currency crises. When
correspondingly, and ut is an error term.
there is an excessive increase in the demand for foreign exchange,
monetary authorities may use a variety of options to cover this
Taylor rule fundamentals. The Taylor rule fundamentals model demand, including adjusting the exchange rate, interest rates, and
expands on the concept of the Taylor rule, which is employed by foreign exchange reserves.
central banks to establish the target interest rate, by applying it to The speculative attacks model of Eichengreen et al. (1994) is a
foreign exchange rates in the currency market. This model takes framework used to identify currency crises by taking into account
into account various economic factors, including inflation, output more than just the exchange rate regime or nominal exchange
gap, and the exchange rate, to determine the equilibrium rate changes. While these factors are important, they may not
exchange rate. By incorporating these fundamental variables, the always accurately reflect the underlying reasons for a country to
Taylor rule fundamentals model aims to provide insights into the maintain its current exchange rate level. Instead, there are other
relationship between interest rate changes and exchange rate factors such as changes in economic development or a country’s
movements (Chen et al., 2017). decision to enter a monetary union that can also impact the
The underlying assumption of the model is that the central volatility of a currency’s price.
bank modifies interest rates based on variations in economic Therefore, Eichengreen et al. (1994) used three variables to
fundamentals in order to uphold price stability and achieve construct an index to capture currency crisis episodes: the series
optimal employment levels. The target interest rate is established of currency exchange rates, interest rate differentials between
through a rule that takes into account the difference between the national and foreign countries, and national reserve differentials
actual inflation rate and the target rate, as well as the output gap. between the two countries. The formula used to construct the
Furthermore, the model incorporates a component that captures speculative pressure index on the currency is as follows:
the divergence of the exchange rate from its equilibrium value, h       i
which is determined by the fundamental variables (Molodtsova EMPi;t ¼ α%Δei;t þ βΔ ii;t  ir;t  γ %Δri;t %Δrr;t
and Papell, 2009).
The Taylor rule fundamentals model is useful for predicting ð6Þ
future exchange rate movements based on changes in economic In this model, ei,t represents the exchange rate, which is the
fundamentals. If economic fundamentals suggest that the price of a foreign currency in terms of i’s currency at time t. The
exchange rate is overvalued, the model predicts that the central variables ii and ir represent the difference in short-term interest
bank will raise interest rates to bring the exchange rate back to its rates between i’s currency and a reference currency. The terms ri
equilibrium value. Conversely, if economic fundamentals suggest and rr indicate the percentage difference in the changes of ratios
that the exchange rate is undervalued, the model predicts that the between international reserves and narrow money (M1). The

6 HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | (2023)10:353 | https://doi.org/10.1057/s41599-023-01836-2

Content courtesy of Springer Nature, terms of use apply. Rights reserved.
HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | https://doi.org/10.1057/s41599-023-01836-2 ARTICLE

weights α, β, and γ assign importance to each of these factors. economic policy, and θ represents the weight assigned to
During a crisis, the occurrence is explained by an outlier value of deviations in inflation from the policy goal.
the EMP index that surpasses the sample mean by a significant As done with the previous model, a more detailed breakdown
standard deviation. of the meaning and effect of each part of the present model is
Explaining the meaning and effects of every large factor of this shown below for a better explanation to the reader:
model, we can detail the next breakdown: At time t, Lt represents the speculative pressure index,
α%Δei,t: The term mentioned represents the impact of indicating the level of pressure exerted by market participants
exchange rate fluctuations on the EMP. The coefficient α on the country’s currency to devalue. The value of 0.5 is a
determines the degree of sensitivity of the EMP to changes in constant weighting factor that equally considers the two
the exchange rate. A higher value of α indicates greater components of the index. The parameter θ represents the
responsiveness of the EMP to exchange rate movements. sensitivity of the index to changes in the exchange rate. A higher
βΔ(ii,t−ir,t): This component reflects the influence of interest θ value indicates that the index is more responsive to fluctuations
rate differentials on the EMP. The coefficient β represents the in the exchange rate. The variable pt refers to the current period’s
level of sensitivity of the EMP to variations in interest rate exchange rate, while pt−1 represents the exchange rate in the
differentials. A higher β value indicates greater responsiveness of previous period. The difference yt−y* signifies the gap between
the EMP to changes in interest rate differentials. the actual output (yt) and the potential output (y*) of the
γ(%Δri,t %Δrr,t): This term accounts for the impact of changes economy. This difference reflects the level of economic activity
in reserves on the EMP. The coefficient γ quantifies the degree of and growth in the country. The term 0.5θ(pt − pt−1) represents
sensitivity of the EMP to fluctuations in the reserves differential. the second component of the speculative pressure index,
A higher γ value signifies a greater sensitivity of the EMP to capturing the deviation of actual output from potential output.
changes in the reserves differential. When the economy is underperforming relative to its potential,
market participants may perceive a need for a currency
Speculative attacks’ second-generation model. The speculative devaluation to stimulate growth and enhance economic activity.
attacks model posits that the government’s choice to devalue its Consequently, this leads to increased pressure on the currency to
currency is contingent upon weighing the costs linked to aban- devalue.
doning the fixed exchange rate system against the loss of cred- In summary, the model indicates that the speculative pressure
ibility that ensues. If the cost of devaluation is deemed lower than index is influenced by expectations of exchange rate devaluation
the cost of maintaining the fixed exchange rate, the government and the deviation of the economy’s performance from its
may opt to devalue its currency. This decision-making process is potential. The index tends to increase when there are anticipa-
shaped by the anticipation of forthcoming economic policies, tions of devaluation or when the economy is underperforming.
which are depicted by different equilibrium levels. This creates a self-fulfilling cycle, as heightened pressure can
The second-generation models of currency crises diverge from trigger a devaluation, which in turn intensifies the pressure for
the first-generation models by incorporating multiple equilibria further devaluation.
that consider the interplay between the private sector and the
government. These models capture the interaction between the Estimation methods
private and public sectors, resulting in various potential outcomes Markov Chain Monte Carlo (Metropolis–Hasting Algorithm).
(Alaminos et al., 2022b). When international financial actors The Metropolis–Hasting (MH) algorithm is a well-known and
foresee a potential currency devaluation, it can trigger a financial complex sampling technique used in statistical analysis (Heratha
crisis as interest rates rise to incentivize domestic currency over and Herath, 2018; Haario et al., 2006). This algorithm is based on
foreign currencies. This situation may prompt the government to Markov chains and is related to rejection methods, which means
devalue its currency due to the high costs associated with that a proposed value is required, and the normalization of the
servicing its debt. Conversely, when private agents do not distribution function being sampled is not necessary. The MH
anticipate an exchange rate change, interest rates remain low, algorithm is inspired by the behavior of systems near equilibrium
reducing the likelihood of devaluation. in statistical mechanics. In this algorithm, the transition prob-
Flood and Marion (1997) introduced second-generation abilities between different states, X and Y, are utilized to describe
models to elucidate the self-fulfilling nature of shocks. According the evolution of the system. The concept of equilibrium is reached
to this framework, if economic agents anticipate a potential when, on average, the system has an equal probability of being in
currency devaluation, their expectations are factored into wage either state X or Y. Furthermore, the probability of transitioning
negotiations, leading to economic imbalances. These imbalances from state Y to X is equivalent to the probability of transitioning
subsequently result in an increase in the country’s price level. To from state X to Y. This equilibrium condition is mathematically
rectify such disequilibrium, the government may choose to adjust expressed through the detailed balance equation.
the exchange rate, which is fixed based on wage agreements.
When the government decides against devaluation, it addresses f ð X ÞPðYjX Þ ¼ f ðY ÞPð XjY Þ ð8Þ
the economic imbalances and prevents an inflationary surge by
reducing its influence on the variables that determine production where f(X)P(Y|X) represents the likelihood of finding the system
levels. Alternatively, if the government opts for a flexible in the vicinity of state X, denoted by f(X), multiplied by the
exchange rate regime, it contributes to a situation where both conditional probability of the system transitioning from state X to
wage levels and price levels in the country rise. Equation (7) state Y, denoted by P(Y|X) (Ayekple et al., 2018). The detailed
illustrates the costs associated with the exchange rate regime in balance condition plays a crucial role in maintaining the correct
both scenarios. distribution in the algorithm and is employed to determine the
acceptance probability of proposed moves in the

  2 Metropolis–Hastings algorithm. The acceptance probability is
Lt ¼ 0:5θ pt  pt1 þ 0:5 yt  y* ð7Þ
determined by comparing the probabilities on the left and right-
hand sides of the detailed balance equation. By adhering to this
being pt represents the national price level, yt represents the condition, the algorithm ensures that the samples obtained align
country’s output at time t, y* denotes the target output set by with the desired distribution. P(Y|X) is known and it is about

HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | (2023)10:353 | https://doi.org/10.1057/s41599-023-01836-2 7

Content courtesy of Springer Nature, terms of use apply. Rights reserved.
ARTICLE HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | https://doi.org/10.1057/s41599-023-01836-2

finding f(X). The MH algorithm claims the opposite: given an X) ≥ 0 ensures that the proposal probability ratio is non-negative,
f(X) is about finding the transition probability that brings the which is necessary for the acceptance probability to be well-
system to equilibrium. Transitions are proposed from X to Y defined (Betancourt, 2019).
following any test distribution T(Y|X). It compares f(Y) with f(X) The summary of the procedure explained so far to generate
and it accepts Y with likelihood A(Y|X). Markov chains from the Metropolis–Hastings algorithm has the
The acceptance probability determines whether a proposed following general outline:
state is accepted or rejected based on the comparison of
probability densities between the current state X and the
proposed state Y. If the probability density of Y is higher than Algorithm: Metropolis–Hastings
that of X, the acceptance probability A(Y|X) is set to 1, resulting Step 1: Initialize X0, t = 0.
in the acceptance of the move to Y. However, if the probability Step 2: Repeat {
density of Y is lower than that of X, the acceptance probability Generate a candidate Y ⁓ q(.|Xt)
A(Y|X) is <1. In this case, the move to Y is accepted with a Generate U ⁓ U(0, 1)
probability of A(Y|X) and rejected with a probability of 1−A If U ≤(Xt; Y), take Xt+1 = Y
(Y|X). The proposed state Y is randomly chosen from the otherwise, take Xt+1 = Xt
proposal distribution T(Y|X), which satisfies the detailed balance Increase t
condition. The detailed balance condition ensures that the ratio of }
transition probabilities, P(Y|X) to P(X|Y), is equal to the ratio of
acceptance probabilities, A(Y|X) to A(X|Y). By satisfying this For this derivation of the algorithm, we have the Metropolis
condition, the algorithm adheres to the principle of detailed random walks, where q(Y|X) = q(|X−Y|). In all cases, it is
balance, facilitating the convergence of the Markov chain to the necessary to note that the quantity α(X,Y) is fundamentally for
desired equilibrium distribution (Heratha and Herath, 2018). the construction of Markov chains. The choice of the form of
Therefore, α(X,Y), which is very simple, guarantees that π(·) satisfies the
balance condition, and therefore that π(·) is itself the
PðYjX Þ ¼ AðYjX ÞT ðYjX Þ ð9Þ
stationary distribution of the Markov chain (Neureiter et al.,
A Markov chain made up of the states is constructed X0, X1, 2022). The key quantity in the Metropolis algorithm is the
X2,…,XN which is reached in each transition, from an initial X0. acceptance probability α(X,Y), which determines whether the
Each Xn is a random variable that will satisfy the following proposed state Y is accepted or rejected. The choice of α(X,Y)
condition: is crucial for constructing Markov chains that have a desired
stationary distribution. The form of α(X,Y) is typically quite
lim ϕn ð X Þ ¼ f ð X Þ ð10Þ
n!1 simple and is designed to satisfy the balance condition. The
The mentioned condition implies that as the number of balance condition ensures that the desired distribution π(·) is
iterations increases indefinitely, the empirical distribution of the the stationary distribution of the Markov chain. In other
states Xn, represented by ∅n(X), approaches the target distribu- words, as the Markov chain converges, the distribution of
tion f(X). This means that the algorithm generates a sequence of states should approach the desired distribution π(·) (Ayekple
states, X0, X1, X2,…,XN, and as the number of iterations (N) grows et al., 2018). The balance condition is satisfied when the
larger, the distribution of the final state XN converges to the acceptance probability α(X,Y) is defined as α(X,Y) = min{1,
desired target distribution f(X) (Haario et al., 2006). During each π(Y)/π(X)}.
step of the random path, there is a transition probability T(Y|X) In this context, π(X) and π(Y) represent the probability
that ensures the transitions between states are properly normal- densities of states X and Y, respectively. The acceptance
ized, indicating the likelihood of moving from one state to probability is determined by comparing the relative likelihoods
another. of transitioning from state X to state Y. If the probability density
Z of the new state Y is higher than that of the current state X
dY T ðYjX Þ ¼ 1 ð11Þ (π(Y) > π(X)), the transition is always accepted with probability 1
(Heratha and Herath, 2018). If the new state Y has a lower
Assuming it is always possible to go from Y to X if it is possible probability density (π(Y) < π(X)), the transition is accepted with
to go from X to Y, and vice versa, we define, probability π(Y)/π(X). By incorporating this acceptance prob-
ability, the Markov chain tends to explore the state space in a way
T ðXjY Þf ðY Þ
qðYjX Þ ¼ ≥0 ð12Þ that is consistent with the desired distribution π(·).
T ðXjY Þf ð X Þ Following the implementation, this study used the
where T(Y|X) represents the probability of proposing state Y Gelman–Rubin Test as a criterion to analyze the convergence
given the current state X, which is known as the proposal of Markov chains (Nguyen and Jones, 2022), which consists of the
distribution. The target distribution is denoted by f(X) for the following steps:
current state X and f(Y) for the proposed state Y. The proposal
1. Generate M ≥ 2 strings, each with 2N iterations, starting
probability ratio q(Y|X) is used to calculate the acceptance
from different starting points.
probability of the proposed state Y given the current state X. This
2. Discard the first N iterations of each chain to allow for
acceptance probability determines whether the proposed state is
burn-in.
accepted or not. The value of q(Y|X) is treated as the probability
3. Calculate the within-chain variance of each chain s2j after
of acceptance in the algorithm, indicating the likelihood of
discarding the first N iterations. Then, calculate the average
transitioning to the proposed state Y.
  of the within-chain variances:
AðYjX Þ ¼ min qðYjX Þ; 1 ð13Þ 1 M 2
W¼ ∑s ð14Þ
The acceptance of the proposed state Y depends on the value of M j¼1 j
q(Y|X). If q(Y|X) is less than or equal to 1, the state Y is always
accepted. However, if q(Y|X) is greater than 1, the acceptance where s2 is the variance of each chain, calculated after
probability is determined by min{1, q(Y|X)}. The condition q(Y| discarding the first N iterations. Calculate the between-chain

8 HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | (2023)10:353 | https://doi.org/10.1057/s41599-023-01836-2

Content courtesy of Springer Nature, terms of use apply. Rights reserved.
HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | https://doi.org/10.1057/s41599-023-01836-2 ARTICLE

variance B by taking the variance of the means of the M are simulated from the system of formulas of state. Then, we
chains: allocate every particle m = 1, …, M a weight proportional to its
N M  2 probability:


B¼ ∑ θj  θ ð15Þ wt;m ¼ p yt xt ; ; ð21Þ
M  1 j¼1
where θ is the mean of the M strings. The weights can be normalized as
4. Estimate the variance of the target parameter θ as follows: wt;m
 e t;m ¼ M
w ð22Þ
1 1 ∑m¼1 wt;m
var ðθÞ ¼ 1  Wþ B ð16Þ
N N To generate the particles, the SMC filter uses Monte Carlo
This term represents a weighted average of the variance approximations. Starting from the
previous time step (t−1), the
within each chain and the variance between different filter creates M samples of p yt xt ; ; . These samples are
chains. simulated using the system of state equations. After generating
5. Compute the potential scalerffiffiffiffiffiffiffiffiffiffiffiffi
reduction factor, R: the particles, each particle is assigned a weight wt,m that is
varðθÞ ð17Þ proportional to the likelihood of the observation yt given the state
R¼
W value xt associated with that particle. This weight represents the
The chains are accepted to have converged when typically importance of that particle in approximating the target distribu-
0.97 < R < 1.03 (Neureiter et al., 2022). The Gelman–Rubin test is tion (Bloem-Reddy and Orbanz, 2018). Particles with a higher
used to evaluate convergence by comparing the within-chain likelihood of the observations are assigned higher weights.
variance (W) and the between-chain variance (B). Convergence is Therefore, the weights can be used to calculate Monte Carlo
indicated when the within-chain variance is similar to the integrals using importance sampling (Lux, 2018). Resampling
between-chain variance, suggesting that the chains have con- considers the number of offspring in proportion to the weight of
verged to the same distribution. The test calculates the statistic R, importance and is generated by simulating a set U from M of
which represents the ratio of the estimated variance of the target random variables uniformly distributed in [0; 1], employing the
parameter using all chains to the within-chain variance of each cumulative sum of the normalized weights
m
individual chain.
e t;j
qm ¼ ∑ w ð23Þ
j¼1
Sequential Monte Carlo (SMC). The particle filtering framework
and then setting Om equal to the number of points in U that are
is a strong state space model for inference (Bloem-Reddy and
between qm−1 and qm. When ∅ is fixed, the subsequent
Orbanz, 2018). They have also been commonly used for financial
distribution p(∅|y1:t) can be described as its Monte Carlo mean
and macroeconomic applications (Lux, 2018). The essence is to
map the state variable distribution using a Monte Carlo approach and its variance ;t and s2t , where s2t denotes the vector of
built through a vast amount of random samples, that evolve empirical variances of each element ∅ (Bloem‐Reddy and
according to a simulation-based updating schedule. As such, Orbanz, 2018). Instantly it can be noted that, for artificial
novel observations are rendered by the filter as they are available. evolution of the parameters, the Monte Carlo variance rises to
Every particle is given a weight and recursively updated. s2t þ ξ t . The Monte Carlo approximation can be stated as a

 softened kernel density of the particles:
The filtering problem lies in the determination of p xt y1:t ; ; .
This can be done in the projection (Eq. (19)) and update (Eqs.
  
M
ð jÞ ð jÞ
(20) and (21)) steps: p ;jy1:t  ∑ kt ;tþ1 j;t ; ξ t ð24Þ
Z j¼1




p xtþ1 y1:t ; ; ¼ p xtþ1 xt ; ; p xt y1:t ; ; dxt ð18Þ while the target variance s2t
can be expressed as




 s2t¼ s2t1 þ ξ t þ 2Cov ;t1 ; kt ð25Þ

 p xtþ1 y1:t ; ; p y1:tþ1 xtþ1 ; ;
 ξt
p xtþ1 y1:tþ1 ; ; ¼
 ð19Þ being Cov ;t1 ; kt ¼  2 .
p y1:tþ1 y1:t ; ; This Eq. (24) approximates the distribution of the unknown
Z variables (∅) given the observed data (y1:t) using a set of M




p ytþ1 y1:t ; ; ¼ p xtþ1 jy1:t ; ; p y1:t xtþ1 ; ; dxtþ1 ð20Þ particles. At every
 time  step, the particles are advanced from the
ðjÞ
previous state ;t to the current state (∅t+1) using a transition
In Eq. (18), the prediction step is performed, which involves
ðjÞ
estimating the probability of the next state. This estimation is kernel kt with a tuning parameter ξt. The importance weights of
done by integrating the current state using the transition function the particles are then calculated based on their ability to explain
and the current state probability. Equation (19) is the update step, the observed data at the current time step, as described by
where the probability of the next state given the observations is Eq. (21) (Lux, 2018).
estimated using Bayes’ rule and the likelihood function. Finally, Equation (25) provides the estimation of the target variance at
Eq. (20) is the likelihood computation step, where the likelihood time

2 t, taking into account the variance at the previous time step
of the next observation is estimated by integrating over the next st1 , the tuning parameter ξt, and the covariance between the
state using the transition function and the current observation particles at the previous time step (∅t−1) and the transition
probability (Bloem-Reddy and Orbanz, 2018). In summary, these kernel at the current time step (kt). This equation allows for the
equations outline the recursive estimation process employed in adaptation of the target variance, which helps in maintaining an
sequential Monte Carlo (SMC) to estimate the probability appropriate balance between exploration and exploitation in the
distribution of the system state at each time step. Sequential Monte Carlo algorithm.
The sequential Monte Carlo filter represents the distributions Finally, the bootstrap filter includes a resampling step to
of interest employing Monte Carlo approximations, that is, address particle degeneracy by eliminating particles with low-
employing a set of M particles.

We assume to be in time t−1 and importance weights. In the past, several authors used the
to create M extractions of p xt xt1 ; ; , in such a way that they importance-weighted empirical distribution, but this time, we

HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | (2023)10:353 | https://doi.org/10.1057/s41599-023-01836-2 9

Content courtesy of Springer Nature, terms of use apply. Rights reserved.
ARTICLE HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | https://doi.org/10.1057/s41599-023-01836-2

use a uniformly-weighted distribution as the way is being applied electrons with spin σ (where σ can be either up or down). The
to a few recent papers (Martino and Elvira, 2021). symbol φ represents an orbital of a single particle. The expansion


 φ ¼ ∑i φi xi i ¼ ∑i cyi φi j0i, based on the single-particle base
PN ¼ dx0:t y1:t ¼ N 1 ∑i NtðiÞ δ xðiÞ dx0:t ð26Þ
0:t states0 {|χi〉},1 can be represented as a vector of dimension
where NtðiÞis the number of a descendant of the particle x0:t which φ1
B φ2 C
will be calculated by the branching process during sampling. The B C
M : B . C. The state |φ〉 represents a many-body wave
most commonly used mechanism process is about resampling N @ .. A
times from P ^N (Gordon et al., 1993). During the sampling
φM
process, a branching mechanism is utilized to duplicate or function that can be expressed as a Slater determinant. Given N
eliminate each particle based on a branching process. The different single-particle orbitals, we can construct a many-body
decision to duplicate or eliminate a particle is generally governed wave function by taking their antisymmetric product, denoted as
by the assigned importance weights for each particle. This ensures
that particles with higher weights have a higher likelihood of ϕi  φ^ y1 φ
^ y2 ¼ φ
^ yN j0i. Here, the operator φ
^ ym is defined as the sum
y
being duplicated, while particles with lower weights may be ^ ym  ∑i ci φi;m . The matrix Φ represents a matrix of dimension
φ
eliminated. The process of eliminating particles and duplicating M × N, where M is the number of base states and N is the number
others helps to prevent particle depletion and ensure that the of orbitals. This matrix Φ contains the coefficients of the orbitals
particles adequately represent the posterior distribution. We required to form a Slater determinant, expressed as |∅〉:
ð jÞ 0 1
ðjÞ
require ∑i NtðiÞ ¼ N for all t. If Nt ¼ 0, the particle x0:t dies. We φ1;1 φ1;2    φ1;N
B φ2;1 φ2;2    φ2;N C
try to choose NtðiÞ such that ΦB .
B C
Z Z .. .. C. The resulting M × N matrix




 @ .. . . A
ht x0:t PN dx0:t y1:t  ht x0:t P bN dx0:t y1:t ð27Þ φM;1 φM;2    φM;N
corresponds to a Slater determinant. The state |Ψ〉 represents a
The surviving particles, indicated by N(i)t > 0, approximate the
many-body wave function that may not necessarily be a single
distribution pðx0:tjy1:tÞ. Importance weights are employed to
Slater determinant, with the initial state denoted as |Ψ0〉.
adjust the sample density and approximate the target distribution.
We can denote the superposition integral, which is a number,
Resampling is subsequently employed to eliminate particles with
between two non-orthogonal Slater determinants |φ〉 and |φ′〉 as
low weights and duplicate particles with high weights. This
follows:
process generates a new set of samples that closely resemble the

target distribution.The resampling step ensures that the samples ϕ ϕ0 ¼ det ϕy ϕ0 ð28Þ
remain diverse, avoiding the degeneracy of the particle filter.
being Φ† the conjugate transpose of the matrix Φ. The equation
given in the statement describes the superposition integral
Auxiliary-field quantum Monte Carlo (AFQMC). The CPMC
between two non-orthogonal

 Slater determinants, |φ〉 and |φ′〉.
algorithm consists of two main components. The first component
involves projecting the basic state as a random walk of open- The notation det ϕy ϕ0 represents the determinant of the matrix
importance samples in a deterministic Slater space. This random product of the conjugate transpose of the single-particle wave
walk utilizes an accurate restricted representation, which allows functions in |φ〉 and |φ′〉. A superposition integral serves as a
CPMC to scale polynomially. The algorithm also incorporates numerical measure of the overlap between two Slater determi-
importance sampling, enhancing its applicability in various scenar- nants, playing a vital role in evaluating transition probabilities
ios. The second component involves constraining the trajectories of and computing expectation values (Ceperley, 2010). Furthermore,
the random path. During this process, each created Slater determi- the operation of applying the exponential of an operator on a
nant maintains significant overlap with a reference test wave func- Slater determinant follows a specific procedure.
tion, represented as |ψT〉. This constraint eliminates the sign problem, M
resulting in CPMC scaling algebraically rather than exponentially. b
B ¼ exp ∑ cyi Uij cj ð29Þ
ij
However, it introduces a systematic error in the algorithm.
In the CPMC method, particular attention is given to a particle leads to another determinant of Slater:
basis that is specific to the problem at hand. The Hamiltonian ^0y ^0y
^ ϕi ¼ ϕ0y
B 0
1 ϕ2 ¼ ϕN j0i  ϕ i ð30Þ
employed in this method follows the Born–Oppenheimer
approximation and does not involve the mixing of spin states. ^0y ¼ ∑j cy ϕ0 and ϕ0  eU ϕ, being the matrix U formed
with ϕ m j jm
Throughout the discussion, it is assumed that the Hamiltonian
from elements Uij. The second property of a Slater determinant
conserves the total spin projection, ^Sz , and that the electron
states how it is affected by the exponential of an operator.
number remains fixed for each spin component. However, even if
Consider an operator B̂ that can be written as a sum of creation
the Hamiltonian does mix spin states, it can still be effectively
handled using the CPMC method, as demonstrated in a study by and annihilation operators, as shown in Eq. (29), where cyi and cyj
Nguyen et al. (2014). For clarity, the notation used in the are the creation and annihilation operators, respectively, and Uij
subsequent discussion will be explained in the following are the elements of a matrix U. When this operator acts on a
paragraph. Slater determinant |ϕ〉, it results in another Slater determinant
In the context of the Hubbard model lattice, M represents the |ϕ′〉, as shown in Eq. (30) (Nguyen et al., 2014). Here,
number of base states available for a single electron. The state |χi〉 ϕ0y ^0y ^0y
1 ϕ2 ¼ ϕN j0 are the rows of a new N × N matrix formed by
refers to the ith ground state of a single particle, with i ranging taking the transpose of a matrix ϕ′, which is given by multiplying
from 1 to M. The operators cyi and ci correspond to the creation the original matrix ϕ with the matrix exponential of U.
and annihilation operators, respectively, for an electron in the Given B ≡ eU indicates a square matrix M × M, the operation of
state |χi〉. The operator ni represents the number operator ^
B over simply involves multiplying eU, a matrix M × M, by Φ, a
associated with state |χi〉. The variable N represents the total matrix M × N (Ceperley, 2010). It is, therefore, appropriate to
number of electrons, while Nσ specifically indicates the number of consider every Slater determinant represented as two separate

10 HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | (2023)10:353 | https://doi.org/10.1057/s41599-023-01836-2

Content courtesy of Springer Nature, terms of use apply. Rights reserved.
HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | https://doi.org/10.1057/s41599-023-01836-2 ARTICLE

rotatable parts: function at the next time step can be expressed as:
Ψðnþ1Þ ¼ eΔτ ðHET Þ ΨðnÞi
" # ^
ϕi ¼ ϕ i  ϕ i ð31Þ ð37Þ
The appropriate matrix expression would be the following: The objective of the algorithm is to determine the ground state
ϕ¼ϕ ϕ " #
ð32Þ wave function |Ψ0〉, which represents the most stable state of the
currency exchange rate system. Equation (36) outlines the
where ϕ↑ and ϕ↓
have dimensions M × N↑ and M × N↓, iterative process of applying the ground state operator projection
accordingly. The overlap between two Slater determinants is to a test wave function |ΨT〉 in order to converge toward the
simply the product of the overlaps of individual turn determi- ground state wave function |Ψ0〉. The parameter Δτ controls the
nants (Ceperley, 2010): time step, while H ^ denotes the Hamiltonian operator and ET
Y h
y i h
y i
represents the estimated exchange rate. The algorithm gradually
hϕ ϕi0 ¼ hϕσ ϕ0σ i ¼ det ϕ" ϕ0"  det ϕ# ϕ0# ð33Þ evolves the wave function from an initial state (|ΨT〉) to the stable
σ¼";#
ground state (|Ψ0〉) by repeatedly applying the projection
Any operator B^ defined by Eq. (30) operates separately on the operator. Equation (37) illustrates the wave function at the
two rotating parts: (n + 1)th time step, obtained by applying the projection operator
^ ϕi ¼ B
^ " ϕ" i  B
^ # ϕ# i to the wave function at the previous time step. Through repeated
B ð34Þ
iterations, the algorithm approaches the ground state wave
Equation (32) expresses a Slater determinant as a product of function, which provides an estimation of the currency
two determinants corresponding to spin-up and spin-down exchange rate.
electrons, respectively. The notation ϕ′↑ and ϕ↓ denote the In the Hubbard model, the Hubbard–Stratonovich (HS)
^
submatrices corresponding to spin-up and spin-down electrons in transform is employed to convert the exponential term eΔτ V
the Slater determinant, respectively. The dimensions of these into the desired form. This transform is utilized to simplify the
submatrices are M × N↑ and M × N↓, respectively, where M is the mathematical representation of the system:
number of spatial orbitals, and N↑ and N↓ are the numbers of

spin-up and spin-down electrons, respectively. Equation (33) eΔτUni" ni# ¼ eΔτU ðni" þni# Þ=2 ∑p xi eγxi ðni" ni# Þ ð38Þ
shows how to calculate the overlap between two Slater where γ is given by cosh(γ) = exp(ΔτU/2). There is no
determinants, which is simply the product of the overlaps of information in the above procedure about the importance of
their spin-up and spin-down determinants. The overlap of each the deterministic result in the representation of |Ψ0〉 contained in
spin-determinant is expressed as a determinant of the corre- the sampling of ~ x.The following equation expresses the mixed
sponding submatrices. Finally, Eq. (34) states that any operator B ^ estimator of the ground state approximation. |
defined by Eq. (30) operates separately on the spin-up and spin-
down parts of the Slater determinant. This means that the ^ ψ0i
h;T H
Emixed  ð39Þ
operator acts on each submatrix independently. h;T ψ 0 i
Equation (32) provides a representation of a Slater determinant
as the product of two determinants, one for the spin-up electrons where it is required to estimate the denominator by ∑h;T ;k i,
(ϕ′↑) and another for the spin-down electrons (ϕ↓). These k
being |∅k〉 random walks after the balance (Nguyen et al., 2014;
determinants are submatrices with dimensions M × N↑ and Ceperley, 2010). We define the function estimates the super-
M × N↓, respectively, where M represents the number of spatial position of a Slater determinant ∅k:
orbitals, N↑ is the count of spin-up electrons, and N↓ is the count

of spin-down electrons. The overlap between two Slater OT ;k  h;T ;k i ð40Þ
determinants is computed in Eq. (33) by taking the product of
The mixed estimator (Eq. (39)) can be expressed as the
the overlaps of their spin-up and spin-down determinants, each
expectation value of the Hamiltonian operator, Ĥ , in the ground
expressed as a determinant of the corresponding submatrices.
^ defined by Eq. (30) acts state wave function, |ψ0〉, projected onto a trial wave function,
Equation (34) states that any operator B ∅T. The numerator represents this expectation value, while the
independently on the spin-up and spin-down parts of the Slater denominator corresponds to the overlap between the trial wave
determinant, operating on the respective submatrices separately. function and the ground state wave function.
The Hubbard model is a simple paradigm of an interacting We assign a weight wk = OT (∅k) to every random walk
electron system (Ceperley, 2010). Its Hamiltonian is given by the (Ceperley, 2010; Nguyen et al., 2014). The weight for each ride in
following equation: the set
  E is set to one, indicating equal importance for all rides
^ ¼K
H ^ þV ^ ¼ t∑ cyiσ cjσ þ cyjσ ciσ þ U∑ni" ni# ð35Þ ;ðk0Þ ¼ ;T i for all k. Then the form is iterated as follows:
E
e
;ðnþ1Þ ~ ð~
∑~x P x ÞB xÞ e
^ ð~ ;ðnÞ ð41Þ
where t represents an element of the jump matrix, and cyiσ and ciσ
are creation and destruction operators for electrons with spin σ at
where B^ ð~
xÞ ¼ B ^ V ð~
^ k=2 B ^ k=2 .
xÞB
position i (Nguyen et al., 2014). TheQHamiltonian is defined as a
network with dimensions M ¼ d Ld . The exchange rate The walkers, represented as e ;ðnÞ , are now sampled from a
calculations are performed in an approximation to the ground revised distribution. These walkers visually represent the wave
state in particle physics. The ground state wave function |Ψ0〉 can function of the fundamental state using a diagrammatic
be obtained by repeatedly applying the projection of the ground representation:
E
state operator to any test wave function |ΨT〉 that is not
;ðknÞ
orthogonal to |Ψ0〉. ψ ðnÞ / ∑k ωðnÞ   ð42Þ
P gs ¼ eΔτ ðHET Þ
^
ð36Þ OT ;ðknÞ

given ET as the best estimate of the exchange rate of the currency, Equation (42) shows how the wave function of the trial state at
if the wave function at the nth time step is |Ψ (n)〉, the wave iteration n, denoted by ψ ðnÞ , can be approximated using a sum

HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | (2023)10:353 | https://doi.org/10.1057/s41599-023-01836-2 11

Content courtesy of Springer Nature, terms of use apply. Rights reserved.
ARTICLE HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | https://doi.org/10.1057/s41599-023-01836-2

E
over different configurations (denoted by ;ðknÞ ) weighted by recorded: the auxiliary field parameters leading to the new walker
from its parent walker, and an integer tag describing the parent.
their corresponding weights (denoted by ω(n)). These configura- Backpropagation is then performed for m additional iterations.
tions are sampled from a new distribution, which is determined For each walker l in the (n + m)th population, a determinant
by the previous set of weights and the normalization factor OT. ψ T is initiated, and the propagators are applied in reverse order.
The equation also includes a normalization factor in the
The m successive propagators are constructedusing the  stored
denominator to ensure that the sum over all configurations is ΔτH^1
properly normalized to 1, which is a requirement for any valid elements between steps m and m + l, with exp  2 inserted
wave function. This normalization factor depends on the overlap as necessary, following Zhang’s, 2004 methodology. The resulting
D
between the trial state and each individual configuration ðmÞ
determinants ;l are combined with their parents from
(Ceperley, 2010). This iterative approach involves improving
the trial wave function by utilizing a set of walkers that represent iteration n to calculate the expectation value 〈O〉BP, similar to the
the system. The walkers are sampled from a distribution that is composite estimator. The weights are appropriately denoted as
updated at each iteration, taking into account the previous set of ωðl nþmÞ , accounting for significance sampling, as discussed by
weights. This systematic process allows for the gradual conver- Motta et al. (2019).
gence toward the true ground state of the system.
The modified function P eð~
xÞ is defined as the product of individual
e
probabilities Pð~
xÞ for sampling the auxiliary field at
each Empirical results
Q  site in the To estimate the models for USD/EUR and USD/JPY exchange
network. This can be expressed as ~pð~ xÞ ¼ M i
~
p xi , where M rates, daily close data from 2013 to 2021 have been utilized. The
represents the total number of sites in the network. data was obtained from Yahoo Finance historical data, providing
  a reliable source for exchange rate information. In addition to

 OT ;ðk;inÞ
 exchange rate data, macroeconomic indicators played a crucial
ep xi ¼   p xi ð43Þ role in this study. These indicators were sourced from various
OT ;ðk;i1
nÞ
reputable databases such as Federal Reserve Economic Data
E


 E (FRED) of St. Louis, Eurostat, World Bank Open Data, and Bank
where ;ðk;i1
nÞ
¼ ^bV xi1 ^bV xi2 ¼ ^bV xi ;ðknÞ , is the current of Japan’s statistics. By incorporating relevant macroeconomic
E data, a comprehensive analysis of the exchange rate dynamics and
state of the kth walker, ;ðknÞ , after your first (i−1) fields have been factors influencing them was conducted. The information used in
E
 nÞ E
sampled and updated, and ;ðk;1 nÞ
¼ ^bV xi ;ðk;11 is the next sub- this study is available at a monthly frequency, which is the
smallest frequency provided by international organizations and
step after
selecting
 the ith field (Ceperley, 2010; Nguyen et al., 2014). the databases used in this research. This choice of frequency
In each e p xi , xi can only take the
value of +1 or −1 and can be aligns with previous studies, such as Rossi (2013), which suggest
sampled after choosing e
xi del p
xi =Ni , where the normalization that the selected frequency for empirical exchange rate research


factor is N i  ~p xi ¼ þ1 þ ~p xi ¼ 1 , and carrying weight for does not significantly impact the reliability of the results. The
the random walk wk;i ðnÞ ðnÞ
¼ N i wk;11 . The inverse of the superposition selected economic indicators in this paper include short-term
h
 i1 interest rates, international reserves, narrow money (M1), com-
y ðnÞ
matrix ;T ;k is maintained and updated after each xi is modity prices, and 10-year government bond yields. These indi-
selected. cators have been chosen based on their relevance in previous
research on the models employed, as discussed in the section
For every walker ;ðnÞ k , a random walk step thus involves: “Literature review” of the paper. The inclusion of these indicators
6. One x sampling starting from the likelihood density enhances the understanding and prediction of financial market
 turbulence, including speculative attack models studied in this
function PeðxÞ=N ;ðnÞ . With discrete, Ising-like auxiliary
k research, following the suggestions of Wang et al. (2020). For the
fields for a Hubbard interaction, the sampling is achieved estimation process, two Intel Core I7-6500U quad-core pro-
by a heat bath-like algorithm, sweeping through each field cessors were used, and the code was implemented using the
xi. MATLAB package (version R2019b). To ensure robustness and
7. compute the ;ðknÞ walker to produce a new walker after accuracy, 500 computing runs were performed to estimate the
building the corresponding B(x).  FOREX fundamentals models, enhancing the reliability of the
8. allocate a weight ωðknþ1Þ ¼ ωðknÞ N ;ðknÞ to the new walker. results.

In the reorthogonalization procedure, no modification of the

weight of a walker is required. This is because the upper Results for USD/EUR exchange rate. Tables 1–7 show adjust-
triangular matrix R, which arises during reorthogonalization, only ment levels by means of accuracy, RMSE, and MAPE. In all
contributes to the overlap OT, and the weight of the walker Monte Carlo methods, the level of precision always surpasses
already accounts for this overlap. Therefore, after each reortho- 77.41% for out-of-sample data for the sample of 200 observations.
gonalization step, R can be disregarded, as explained by Motta For its part, the RMSE and MAPE levels are adequate, whereas
et al. (2019). To estimate the expected value of an observable that the Monte Carlo methods, especially the AFQMC quantum
does not commute with the Hamiltonian, the back-propagation technique, where improve both the precision results and the error
approach, proposed by Zhang in 2004, is employed. When two levels of the rest of the methodologies used. For the sample of 500
independent populations are decoupled, significant fluctuations observations, it shows this same trend, only that most meth-
in the estimator can occur after uncoupling the significance odologies improve their precision results. Therefore, the AFQMC
functions, since the population in the random walk is sampled. In technique obtains the best results with a precision range for data
the backpropagation outside the sample of 86.93–89.56% for the sample of 200
n method,
Eo an iteration n is selected, and the observations, while it reaches a range of 87.96–92.97% for the
entire population ;ðknÞ is stored. For each new walker as the sample of 500 observations. Overall, these findings yield a pre-
random walk progresses from n, two pieces of information are cision level significantly higher than that of prior investigations.

12 HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | (2023)10:353 | https://doi.org/10.1057/s41599-023-01836-2

Content courtesy of Springer Nature, terms of use apply. Rights reserved.
HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | https://doi.org/10.1057/s41599-023-01836-2 ARTICLE

Table 1 Results of the UIRP model.

N Model In-sample Out-of-sample

Classification (%) RMSE MAPE Classification (%) RMSE MAPE

200 MCMC 85.65 0.62 0.37 81.76 0.64 0.39
SMC 86.71 0.61 0.36 82.77 0.64 0.38
AFQMC 93.18 0.37 0.22 89.56 0.39 0.23
500 MCMC 89.04 0.33 0.19 85.61 0.35 0.20
SMC 87.77 0.38 0.22 84.39 0.39 0.23
AFQMC 95.21 0.23 0.14 92.97 0.26 0.15

N is the number of observations.

Table 2 Results of the PPP model.

N Model In-sample Out-of-sample

Classification (%) RMSE MAPE Classification (%) RMSE MAPE

200 MCMC 84.20 0.72 0.53 80.37 0.75 0.56
SMC 85.18 0.78 0.44 81.30 0.81 0.46
AFQMC 92.17 0.61 0.39 88.39 0.64 0.41
500 MCMC 87.30 0.40 0.18 83.33 0.42 0.19
SMC 86.38 0.32 0.21 82.44 0.34 0.22
AFQMC 93.53 0.26 0.15 90.06 0.28 0.17

N is the number of observations.

Table 3 Results of the BEER model.

N Model In-sample Out-of-sample

Classification (%) RMSE MAPE Classification (%) RMSE MAPE
200 MCMC 84.83 0.66 0.39 80.14 0.69 0.41
SMC 85.88 0.65 0.38 81.13 0.68 0.40
AFQMC 93.24 0.39 0.24 87.39 0.41 0.25
500 MCMC 89.09 0.35 0.21 83.92 0.37 0.22
SMC 87.82 0.40 0.24 82.72 0.43 0.25
AFQMC 95.27 0.25 0.16 90.71 0.28 0.17

N is the number of observations.

Table 4 Results of the SPM model.

N Model In-sample Out-of-sample

Classification (%) RMSE MAPE Classification (%) RMSE MAPE

200 MCMC 84.25 0.78 0.57 79.60 0.81 0.60
SMC 85.23 0.84 0.48 80.53 0.87 0.50
AFQMC 92.23 0.65 0.43 88.03 0.68 0.45
500 MCMC 87.35 0.44 0.20 82.53 0.46 0.21
SMC 86.43 0.34 0.23 81.66 0.36 0.24
AFQMC 93.59 0.28 0.17 89.69 0.30 0.19

N is the number of observations.

For example, in the study by Hauzenberger and Huber (2019), an In light of our results, the reliability and stability of the
accuracy of around 77% is revealed, in the work of Cheung et al. AFQMC method can be verified, if we observe the average result
(2019) it is close to 74%, and in the study of Beckmann and of RMSE and MAPE after 500 iterations compared to the rest of
Schüssler (2016), it approaches 75.94%. Other works such as Park the applied methodologies. Therefore, the AFQMC method
and Park (2013), and Rubaszek and Ca’ Zorzi (2020) achieve an improves the precision and error values shown by the rest of
accuracy lower than even 70%. Therefore, the difference shown the Monte Carlo methodologies and the regression techniques
by the Monte Carlo methodologies, and especially the quantum (Lee, 2011; Jaworski, 2018; Hauzenberger and Huber, 2019), both
variant, used in this research surpasses by far the accuracy with a sample of 200 observations out of 500 observations. This
reported in the earlier literature. novel quantum method demonstrates a high level of precision for

HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | (2023)10:353 | https://doi.org/10.1057/s41599-023-01836-2 13

Content courtesy of Springer Nature, terms of use apply. Rights reserved.
ARTICLE HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | https://doi.org/10.1057/s41599-023-01836-2

Table 5 Results of the Taylor model.

N Model In-sample Out-of-sample

Classification (%) RMSE MAPE Classification (%) RMSE MAPE

200 MCMC 85.67 0.74 0.55 81.40 0.78 0.57
SMC 86.53 0.80 0.46 82.22 0.84 0.48
AFQMC 92.23 0.63 0.40 88.54 0.65 0.43
500 MCMC 88.45 0.41 0.19 84.05 0.44 0.20
SMC 87.27 0.33 0.22 82.93 0.34 0.23
AFQMC 93.83 0.28 0.17 90.47 0.30 0.18

N is the number of observations.

Table 6 Results of the speculative attacks model.

N Model In-sample Out-of-sample

Classification (%) RMSE MAPE Classification (%) RMSE MAPE
200 MCMC 85.62 0.68 0.51 82.19 0.72 0.53
SMC 86.47 0.74 0.42 83.01 0.78 0.44
AFQMC 92.17 0.59 0.38 88.90 0.61 0.39
500 MCMC 88.40 0.39 0.17 84.86 0.40 0.18
SMC 87.22 0.31 0.20 83.73 0.32 0.21
AFQMC 93.77 0.26 0.15 90.84 0.28 0.16

N is the number of observations.

Table 7 Results of the speculative attacks’ second-generation model.

N Model In-sample Out-of-sample

Classification (%) RMSE MAPE Classification (%) RMSE MAPE
200 MCMC 78.53 0.64 0.47 77.41 0.70 0.50
SMC 83.35 0.71 0.39 81.28 0.77 0.43
AFQMC 89.25 0.57 0.35 86.93 0.58 0.38
500 MCMC 82.73 0.38 0.16 78.31 0.39 0.18
SMC 81.79 0.29 0.19 83.23 0.31 0.20
AFQMC 93.33 0.24 0.14 87.96 0.27 0.15

N is the number of observations.

the FOREX market estimation, different from the methods used (Rossi, 2013; Jacob and Uusküla, 2019; Rubaszek and Ca’ Zorzi,
in the previous literature. 2020; Adegboye et al., 2021). In the same way, it even improves
In order to demonstrate the superiority of the quantum Monte the precision results shown in works related to research lines such
Carlo AFQMC methodology in estimating dynamic exchange rate as the FOREX market prediction with Machine Learning
models and to provide robustness tests, the Diebold–Mariano methodologies, such as with neural networks, where they have
(DM) test (Diebold and Mariano, 1995) was employed. The shown accuracies in the range of 73.92–86.48% (Ni et al., 2019;
comparison between the techniques used also considered the time Wei et al., 2019; Zheng et al., 2019; Parot et al., 2019; Cao et al.,
required for estimation. Table 8 presents the results of the DM 2020; Wang et al., 2020), the precision displayed by Support
analysis, showing that AFQMC outperforms other methods as the Vector Machines with a range of 71.58–88.26% (Fu et al., 2019;
results fall within the range of −1.96 to 1.96, indicating that the Sun et al., 2019), or Random Forest with a precision of
observed differences in estimation between the methods are 68.31–84.19% (Wei et al., 2019; Colombo and Pelagatti, 2020).
statistically significant at a 5% significance level. Negative values
indicate that the second choice in the comparison is superior.
These results hold for both the sample sizes of 200 and 500 Results for USD/JPY exchange rate. Tables 10–16 show the
observations. Additionally, Table 9 displays the mean execution precision levels as the results of error levels measured by RMSE
time for the estimation approaches, revealing that Monte Carlo and MAPE. The precision values of the Monte Carlo methodol-
approaches, particularly AFQMC, require less time for estima- ogies used exceed 90.48% for out-of-sample data for the sample of
tion, both for in-sample and out-of-sample data. The estimated 200 observations. After analyzing the error data, the RMSE and
time ranges from 0.6 to 0.20, considering both sample and out-of- MAPE values show stable levels, where the Monte Carlo methods
sample data, as well as the sample sizes of 200 and 500 and the AFQMC quantum technique improve both the precision
observations. This improves the time elapsed to obtain results results and the error levels of the rest of the methodologies used.
from the FOREX market models carried out in previous works The same conclusions are shown for the sample of 500

14 HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | (2023)10:353 | https://doi.org/10.1057/s41599-023-01836-2

Content courtesy of Springer Nature, terms of use apply. Rights reserved.
 Table 8 Comparison of Diebold–Mariano test results.

N UIRP model PPP model BEER model SPM model Taylor model Speculative attacks Speculative attacks’ SG
model model

In-sample Out-of In-sample Out-of In-sample Out-of In-sample Out-of In-sample Out-of In-sample Out-of In-sample Out-of
sample Sample sample sample sample sample sample
200 MCMC −2.41 −2.46 −2.42 −2.23 −2.48 −2.54 −2.49 −2.30 −2.56 −2.62 −2.33 −2.33 −2.35 −2.36
vs. SMC
MCMC −2.59 −2.65 −2.65 −2.42 −2.68 −2.74 −2.74 −2.49 −2.76 −2.83 −2.52 −2.51 −2.55 −2.61
vs.
AFQMC
SMC vs. −2.55 −2.60 −2.74 −2.49 −2.63 −2.69 −2.82 −2.57 −2.72 −2.77 −2.45 −2.31 −2.53 −2.34
AFQMC
500 MCMC −2.35 −2.41 −2.54 −2.22 −2.43 −2.48 −2.62 −2.29 −2.51 −2.56 −2.53 −2.50 −2.59 −2.51
vs. SMC
MCMC −2.65 −2.72 −2.68 −2.55 −2.74 −2.80 −2.77 −2.63 −2.83 −2.89 −2.54 −2.53 −2.63 −2.56
vs.
AFQMC
SMC vs. −2.56 −2.61 −2.75 −2.75 −2.64 −2.70 −2.83 −2.83 −2.73 −2.78 −2.70 −2.63 −2.80 −2.65
AFQMC

SG means second-generation.

Table 9 Time-lapse for estimating results by methods.

No. UIRP model PPP model BEER model SPM model Taylor Model Speculative attacks Speculative attacks’ SG
model model

In-sample Out-of In-sample Out-of In-sample Out-of In-sample Out-of In-sample Out-of In-sample Out-of In-sample Out-of
sample sample sample sample sample sample sample
200 MCMC 0.15 0.13 0.10 0.09 0.16 0.14 0.11 0.10 0.17 0.14 0.22 0.19 0.32 0.19
SMCMC 0.13 0.11 0.08 0.07 0.14 0.12 0.09 0.07 0.14 0.12 0.19 0.15 0.27 0.21
AFQMC 0.12 0.10 0.08 0.07 0.13 0.11 0.09 0.07 0.13 0.11 0.17 0.14 0.17 0.21

HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | (2023)10:353 | https://doi.org/10.1057/s41599-023-01836-2

500 MCMC 0.13 0.11 0.08 0.07 0.14 0.12 0.09 0.07 0.14 0.12 0.20 0.17 0.22 0.21
SMC 0.12 0.10 0.10 0.08 0.13 0.11 0.11 0.09 0.13 0.11 0.17 0.14 0.22 0.20
HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | https://doi.org/10.1057/s41599-023-01836-2

AFQMC 0.10 0.09 0.07 0.06 0.11 0.10 0.07 0.06 0.11 0.10 0.14 0.12 0.19 0.12

SG means second-generation.

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

15
ARTICLE
ARTICLE HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | https://doi.org/10.1057/s41599-023-01836-2

Table 10 Results of the UIRP model.

N Method In-sample Out-of-sample

Classification (%) RMSE MAPE Classification (%) RMSE MAPE

200 MCMC 88.20 0.57 0.35 82.56 0.61 0.35
SMC 89.30 0.54 0.33 83.42 0.59 0.34
AFQMC 95.97 0.31 0.19 90.48 0.36 0.21
500 MCMC 90.01 0.31 0.17 86.40 0.35 0.18
SMC 88.72 0.33 0.18 85.01 0.37 0.21
AFQMC 96.26 0.20 0.11 93.91 0.22 0.13

N is the number of observations.

Table 11 Results of the PPP model.

N Method In-sample Out-of-sample

Classification (%) RMSE MAPE Classification (%) RMSE MAPE

200 MCMC 87.42 0.68 0.50 84.34 0.72 0.53
SMC 88.42 0.73 0.41 85.32 0.78 0.44
AFQMC 94.84 0.58 0.38 92.72 0.61 0.39
500 MCMC 90.63 0.39 0.17 87.45 0.40 0.18
SMC 89.67 0.30 0.20 86.52 0.32 0.21
AFQMC 95.89 0.23 0.14 94.43 0.27 0.14

N is the number of observations.

Table 12 Results of the BEER model.

N Method In-sample Out-of-sample

Classification (%) RMSE MAPE Classification (%) RMSE MAPE
200 MCMC 87.36 0.61 0.37 87.11 0.65 0.37
SMC 88.45 0.58 0.35 88.01 0.63 0.36
AFQMC 96.02 0.33 0.21 95.02 0.38 0.23
500 MCMC 90.07 0.33 0.19 91.16 0.37 0.20
SMC 88.78 0.35 0.20 89.69 0.39 0.23
AFQMC 96.32 0.22 0.11 95.62 0.24 0.15

N is the number of observations.

Table 13 Results of the SPM model.

N Method In-sample Out-of-sample

Classification (%) RMSE MAPE Classification (%) RMSE MAPE

200 MCMC 87.47 0.74 0.54 88.98 0.78 0.57
SMC 88.47 0.79 0.45 90.02 0.84 0.48
AFQMC 94.89 0.62 0.40 97.37 0.65 0.43
500 MCMC 90.69 0.41 0.19 92.26 0.44 0.20
SMC 89.72 0.32 0.22 91.29 0.34 0.23
AFQMC 95.95 0.25 0.16 95.16 0.29 0.16

N is the number of observations.

observations since most of the methodologies slightly improve 7580% (Park and Park, 2013; Ince et al., 2016; Byrne et al., 2016;
their results. This means that the AFQMC method achieves the Kavtaradze and Mokhtari, 2018; Serjam and Sakurai, 2018;
best results with a precision range for out-of-sample data of Taveeapiradeecharoen et al., 2019; Cheung et al., 2019; Hau-
90.48–97.37% for the sample of 200 observations, while it reaches zenberger and Huber, 2019; Rubaszek and Ca’ Zorzi, 2020), being
a range of 90.09–95.62% for the sample of 500 observations. difficult to achieve high levels of success. Therefore, the difference
These results conclude an improvement of the precisions shown shown by the AFQMC Monte Carlo quantum specification sur-
by other works, where regression and Monte Carlo methods were passes the accuracy displayed by prior literature. The precision
used. Thus, previous work for the estimation of FOREX markets and error levels achieved for the USD/JPY exchange rate surpass
with the usual statistical methodology has an accuracy of around the results obtained in previous studies using Machine Learning

16 HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | (2023)10:353 | https://doi.org/10.1057/s41599-023-01836-2

Content courtesy of Springer Nature, terms of use apply. Rights reserved.
HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | https://doi.org/10.1057/s41599-023-01836-2 ARTICLE

Table 14 Results of the Taylor model.

N Method In-sample Out-of-sample

Classification (%) RMSE MAPE Classification (%) RMSE MAPE

200 MCMC 87.81 0.70 0.52 89.76 0.75 0.55
SMC 88.69 0.77 0.44 90.66 0.81 0.46
AFQMC 94.53 0.60 0.38 96.64 0.63 0.40
500 MCMC 90.67 0.39 0.18 92.69 0.41 0.19
SMC 89.46 0.31 0.21 91.45 0.33 0.22
AFQMC 96.17 0.25 0.13 95.31 0.26 0.17

N is the number of observations.

Table 15 Results of the speculative attacks model.

N Method In-sample Out-of-sample

Classification (%) RMSE MAPE Classification (%) RMSE MAPE

200 MCMC 87.76 0.65 0.48 85.08 0.69 0.51
SMC 88.64 0.71 0.40 85.93 0.75 0.42
AFQMC 94.47 0.56 0.36 92.02 0.59 0.38
500 MCMC 90.61 0.37 0.16 87.85 0.39 0.17
SMC 89.40 0.29 0.19 86.68 0.31 0.20
AFQMC 96.11 0.23 0.13 93.62 0.24 0.15

N is the number of observations.

Table 16 Results of the speculative attacks’ second-generation model.

N Method In-sample Out-of-sample

Classification (%) RMSE MAPE Classification (%) RMSE MAPE
200 MCMC 86.50 0.63 0.48 79.53 0.67 0.48
SMC 85.72 0.70 0.39 85.11 0.73 0.39
AFQMC 98.13 0.52 0.34 90.90 0.55 0.37
500 MCMC 85.73 0.35 0.15 84.17 0.39 0.16
SMC 85.58 0.27 0.18 83.98 0.29 0.20
AFQMC 93.87 0.21 0.12 90.09 0.23 0.14

N is the number of observations.

techniques, which have reported accuracies ranging from 68.31 to FOREX market models, with elapsed time ranging from 0.5 to
88.26% (Contreras et al., 2018; Wei et al., 2019; Zheng et al., 2019; 0.17 min for both the samples of 200 and 500 observations, as
Parot et al., 2019; Fu et al., 2019; Cao et al., 2020; Wang et al., well as estimates within and outside the sample data.
2020; Sun et al., 2019; Colombo and Pelagatti, 2020). These In conclusion, our estimation model for the USD/EUR and
improvements align with the findings observed for the USD/EUR USD/JPY exchange rates has demonstrated the reliability and
exchange rate. stability of the AFQMC method. This approach outperforms
Therefore, these levels of precision and errors of the AFQMC other Monte Carlo methodologies in terms of accuracy and error
quantum method demonstrate the difference and superiority over values, regardless of whether the sample consists of 200 or 500
the rest of the techniques, whether they are regression techniques observations. Thus, the AFQMC technique offers a significant
or other Monte Carlo variants. This concludes that there is an improvement in the accuracy of FOREX market models
important margin of greater precision of this new technique for compared to previous simulations in the literature. Specifically,
the estimation of FOREX market models compared to other for the USD/EUR exchange rate, the UIRP model exhibits high
simulations carried out by the previous literature. precision in AFQMC estimations, while for the USD/JPY
In line with the analysis conducted for the USD/EUR exchange exchange rate, the SPM model performs well. Furthermore, the
rate in section “Results for USD/EUR exchange rate”, we also DM analysis confirms that AFQMC remains the preferred option
utilize the Diebold–Mariano (DM) test (Diebold and Mariano, over other methods, as results exceeding the threshold of 1.96/
1995) to compare different techniques and measure estimation −1.96 do not lead to the rejection of the null hypothesis at a 5%
time for the USD/JPY exchange rate. The results in Table 17 significance level, indicating significant differences in estimation
affirm the continued superiority of the AFQMC methodology, as between the methods.
the DM test confirms the non-rejection of the null hypothesis at a
5% significance level, with results surpassing the threshold of
−1.96. This indicates that the observed differences in estimation Conclusions
methods are statistically significant. Table 18 further reveals that In this study, we present new simulations of FOREX market
the AFQMC methodology requires less time for executing models using the Auxiliary-Field Quantum Monte Carlo

HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | (2023)10:353 | https://doi.org/10.1057/s41599-023-01836-2 17

Content courtesy of Springer Nature, terms of use apply. Rights reserved.
 18
Table 17 Comparison of Diebold–Mariano test results.

N UIRP model PPP model BEER model SPM model Taylor model Speculative attacks Speculative attacks’ SG
model model
ARTICLE

In-sample Out-of In-sample Out-of In-sample Out-of In-sample Out-of In-sample Out-of In-sample Out-of In-sample Out-of
sample sample sample sample sample sample sample
200 MCMC −2.23 −2.23 −2.40 −2.34 −2.30 −2.30 −2.47 −2.42 −2.38 −2.38 −2.40 −2.52 −2.42 −2.57
vs. SMC
MCMC −2.41 −2.49 −2.47 −2.59 −2.48 −2.57 −2.55 −2.68 −2.57 −2.65 −2.60 −2.80 −2.66 −2.88
vs.
AFQMC
SMC vs. −2.53 −2.36 −2.73 −2.36 −2.61 −2.44 −2.81 −2.44 −2.70 −2.52 −2.86 −2.43 −2.87 −2.46
AFQMC
500 MCMC −2.30 −2.24 −2.46 −2.44 −2.38 −2.31 −2.54 −2.52 −2.46 −2.39 −2.61 −2.58 −2.61 −2.66
vs. SMC
MCMC −2.41 −2.55 −2.49 −2.57 −2.48 −2.63 −2.57 −2.65 −2.57 −2.72 −2.65 −2.70 −2.68 −2.77
vs.
AFQMC
SMC vs. −2.47 −2.50 −2.49 −2.59 −2.55 −2.58 −2.57 −2.68 −2.63 −2.67 −2.67 −2.75 −2.74 −2.76
AFQMC

SG means second-generation.

Table 18 Time-lapse for estimating results by methods.

N UIRP model PPP model BEER model SPM model Taylor model Speculative attacks Speculative attacks’ SG
model model

In-sample Out-of In-sample Out-of In-sample Out-of In-sample Out-of In-sample Out-of In-sample Out-of In-sample Out-of
sample sample sample sample sample sample sample
200 MCMC 0.13 0.11 0.09 0.07 0.13 0.12 0.10 0.08 0.14 0.12 0.18 0.16 0.23 0.23
SMCMC 0.11 0.09 0.06 0.05 0.12 0.10 0.07 0.05 0.12 0.10 0.16 0.14 0.17 0.17
AFQMC 0.10 0.08 0.06 0.05 0.11 0.09 0.07 0.05 0.11 0.09 0.14 0.11 0.18 0.18
500 MCMC 0.11 0.09 0.07 0.06 0.12 0.10 0.08 0.07 0.12 0.10 0.17 0.15 0.19 0.17
SMC 0.10 0.08 0.08 0.07 0.11 0.09 0.09 0.08 0.11 0.09 0.14 0.11 0.14 0.13
AFQMC 0.09 0.07 0.06 0.05 0.10 0.08 0.07 0.05 0.10 0.08 0.13 0.10 0.13 0.16

SG means second-generation.

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | (2023)10:353 | https://doi.org/10.1057/s41599-023-01836-2
HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | https://doi.org/10.1057/s41599-023-01836-2
HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | https://doi.org/10.1057/s41599-023-01836-2 ARTICLE

(AFQMC) technique, with a particular focus on speculative attack banks to the possibility of substantial currency depreciation
models. Our objective is to gain deeper insights into potential resulting from imbalances in the balance of payments. By pro-
extreme events that may occur in FOREX markets. We utilize viding valuable insights, this work helps stakeholders take
exchange rate data spanning the period 2015–2021, specifically proactive measures to prevent adverse effects on the economy. In
examining the US dollar against the Euro and Japanese yen. Two the same way, the results obtained in this work can help FOREX
different Monte Carlo methods, namely Markov Chain Monte traders and investors to make better decisions in any financial
Carlo and Sequential Monte Carlo, are employed in the analysis. market with an interest in FOREX markets. However, given the
Our findings indicate that the majority of the proposed Monte complexity of this novel quantum method, our research provides
Carlo approaches exhibit a low level of error and demonstrate a new possibility that other authors and professionals may
stability in estimating FOREX market models. Importantly, the develop in the future. Also, further research in this field includes
novel AFQMC technique outperforms the other methods, deli- estimating other kinds of economic and financial models that are
vering the highest level of accuracy and reliability in our usually estimated through Monte Carlo methods.
estimations.
In addition, the study aims to enhance the precision of earlier
studies by employing various deep learning and statistical Data availability
methods. The results achieved in this research are superior to the The datasets used and/or analyzed during the current study are
existing literature, with an average precision range of included in the supplementary information.
87.96–95.62% for out-of-sample using the AFQMC method.
While other Monte Carlo methods were used, it has only reached Received: 14 August 2022; Accepted: 6 June 2023;
an accuracy range of 74.60–90.39%, also for out-of-sample.
Unlike prior investigations, the present research proved to extend
the estimation of FOREX markets addressing the precision and
error results, both small sample and larger sample. This study
makes a significant contribution to the fields of International
Finance and Computational Finance. References
Achchab S, Bencharef O, Ouaarab A (2017) A combination of regression techni-
The findings of this study hold great significance for various ques and cuckoo search algorithm for FOREX speculation. In: Rocha Á,
stakeholders in the FOREX markets, including public managers, Correia AM, Adeli H, Reis LP, Costanzo S (eds), Recent advances in infor-
finance analysts, central bankers, and other interested parties. mation systems and technologies, vol. 15. Springer International Publishing,
These findings offer valuable insights into which indicators can pp. 226–235
Adegboye A, Kampouridis M, Otero F (2021) Improving trend reversal estimation
provide reliable, accurate, credible, and promising forecasts of in forex markets under a directional changes paradigm with classification
future market performance. The proposed method proves to be a algorithms. Int J Intell Syst 36(12):7609–7640
valuable tool for both financial and macroeconomic decision- Alaminos D, Salas MB, Fernández-Gámez MA (2022a) Forecasting stock market
making. Corporate decision-makers can utilize the model for risk crashes via real-time recession probabilities: a quantum computing approach.
management, portfolio optimization, and derivative asset pricing, Fractals-Complex Geom Patterns Scaling Nat Soc 30(5):2240162
Alaminos D, Salas MB, Fernández-Gámez MA (2022b) Deep neural networks
while political decision-makers can leverage it to monitor eco- methods for estimating market microstructure and speculative attacks
nomic factors, assess their impact on exchange rates, and inform models: the case of government bond market. Singap Econ Rev, Online
fiscal and monetary policy development. Moreover, given the Ready
inherent volatility of the FOREX market, our research provides a Alquist R, Chinn MD (2008) Conventional and unconventional approaches to
reliable model that enables investors to anticipate market beha- exchange rate modelling and assessment. Int J Finance Econ 13(1):2–13
Aşırım A, Aşırım ÖE, Salepçioğlu MA (2023) Analysis of local system behavior in
vior prior to executing their operations. This study makes a the foreign exchange-market using neural networks and Monte-Carlo
valuable contribution to the existing knowledge in economics and method for predictıon and risk assessment. SN Appl Sci 5(3):78
finance by introducing a novel estimation analysis that incorpo- Asmussen S (2018) Conditional Monte Carlo for sums, with applications to
rates three Monte Carlo techniques. Among these techniques, insurance and finance. Ann Actuar Sci 12(2):455–478
AFQMC stands out for its superior precision. Auray S, Devereux, MB, Eyquem, A (2019) Endogenous trade protection and
exchange rate adjustment. NBER Working Papers, 25517
Accurate currency forecasting offers numerous advantages to Ayekple YE, Tetteh CK, Fefemwole PK (2018) Markov chain Monte Carlo method
traders, enabling them to anticipate future events and make for estimating implied volatility in option pricing. J Math Res 10(6):108–116
informed decisions in FOREX markets. These models are parti- Ayitey Jr M, Appiahene P, Appiah O, Bombie CN (2023) Forex market forecasting
cularly valuable for long-term investors as they provide insights using machine learning: Systematic Literature Review and meta-analysis. J
into the market’s direction over months and years. For spec- Big Data 10(1):9
Barbosa LOS, Jayme FG, Missio FJ (2018) Determinants of the real exchange rate in
ulators trading on shorter time horizons, these models offer the long-run for developing and emerging countries: a theoretical and
estimates of the overall trend in FOREX markets, allowing them empirical approach. Int Rev Appl Econ 32(1):62–83
to make strategic moves based on this information. Beckmann J, Schüssler R (2016) Forecasting exchange rates under parameter and
This study employs well-established theoretical models of model uncertainty. J Int Money Finance 60(C):267–288
FOREX fundamentals, which have been widely recognized in Betancourt M (2019) The convergence of Markov chain Monte Carlo methods:
from the metropolis method to Hamiltonian Monte Carlo. Ann Phys
previous literature for their ability to estimate the theoretical 531(3):1700214
value of currency exchange rates. But one of the limitations of this Bloem‐Reddy B, Orbanz P (2018) Random‐walk models of network formation and
study could be not having added simulations of these Monte sequential Monte Carlo methods for graphs. J R Stat Soc Ser B (Stat Meth-
Carlo models to trading models in a possible market strategies odol) 80(5):871–898
simulation. This would be an interesting line of future research as Braga de Macedo J, Lempinen U (2013) Exchange rate dynamics revisited
(December 2013). NBER Working Paper No. w19718
an extension of the present work. Byrne J, Korobilis D, Ribeiro PJ (2016) Exchange rate predictability in a changing
To conclude, this research is a significant contribution to the world. J Int Money Finance 62:1–24
analysis of foreign exchange market trends. The obtained results Cao C, Zhu W, Wang W, Demazeau Y, Zhang C (2020) A deep coupled LSTM
have important implications for the decision-making of public approach for USD/CNY exchange rate forecasting. IEEE Intell Syst
and financial institutions, allowing them to mitigate the risks 35(2):43–53
Clements KW, Lan Y (2010) A new approach to forecasting exchange rates. J Int
associated with large volatility in exchange rate trends. Addi- Money Finance 29:1424–1437
tionally, the research assists in alerting governments and central

HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | (2023)10:353 | https://doi.org/10.1057/s41599-023-01836-2 19

Content courtesy of Springer Nature, terms of use apply. Rights reserved.
ARTICLE HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | https://doi.org/10.1057/s41599-023-01836-2

Ceperley DM (2010) An overview of quantum Monte Carlo methods. Rev Mineral Heratha SB, Herath TC (2018) Post-audits for managing cyber security invest-
Geochem 71(1):129–135 ments: Bayesian post-audit using Markov Chain Monte Carlo (MCMC)
Chadwick MG, Fazilet F, Tekatli N (2015) Understanding the common dynamics simulation. J Account Public Policy 37(6):545–563
of the emerging market currencies. Econ Model 49(C):120–136 Ilić M, Digkoglou P (2022) The volatility of stock market returns: application of
Chen C, Yao S, Ou J (2017) Exchange rate dynamics in a Taylor rule framework. J Monte Carlo simulation. Econ Sustain Dev 6(2):17–30
Int Financ Mark Inst Money 46:158–173 Ince O, Molodtsova T, Papell DH (2016) Taylor Rule deviations and out-of-sample
Cheung Y, Chinn MD, Garcia PA, Zhang Y (2019) Exchange rate prediction exchange rate predictability. J Int Money Finance 69:22–44
redux: new models, new data, new currencies. J Int Money Finance Islam MS, Hossain E, Rahman A, Hossain MS, Andersson K (2020) A review on
95:332–362 recent advancements in FOREX currency prediction. Algorithms 13(8):186
Cheung Y, Erlandsson UG (2005) Exchange rates and Markov switching dynamics. Ismailov A, Rossi B (2018) Uncertainty and deviations from uncovered interest rate
J Bus Econ Stat 23(3):314–320 parity. J Int Money Finance 88(C):242–259
Chinthapalli U (2021) A comparative analysis on probability of volatility clusters Jacob P, Uusküla L (2019) Deep habits and exchange rate pass-through. J Econ Dyn
on cryptocurrencies, and FOREX currencies. J Risk Financ Manag Control 105(C):67–89
14(7):308 Jarusek R, Volna E, Kotyrba M (2022) FOREX rate prediction improved by Elliott
Chou JS, Truong TTH (2019) Sliding-window metaheuristic optimization- waves patterns based on neural networks. Neural Netw 145:342–355
based forecast system for foreign exchange analysis. Soft Comput Jaworski K (2018) Density forecasts of emerging markets’ exchange rates using
23(10):3545–3561 Monte Carlo simulation with regime switching. In: Jajuga K, Locarek-Junge
Colombo E, Pelagatti M (2020) Statistical learning and exchange rate forecasting. H, Orlowski L (eds) Contemporary trends and challenges in finance. Springer
Int J Forecast 36(4):1260–1289 proceedings in business and economics. Springer, Cham
Contreras AV, Llanes A, Pérez-Bernabeu A, Navarro S, Pérez-Sánchez A, López- Kamal Y, Haq HIU, Ghani U, Khan MM (2012) Modeling the exchange rate volatility,
Espín JJ, Cecilia JM (2018) ENMX: an elastic network model to predict the using generalized autoregressive conditionally heteroscedastic (GARCH) type
FOREX market evolution. Simul Model Pract Theory 86:1–10 models: evidence from Pakistan. Afr J Bus Manag 6(8):2830–2838
Creal D (2012) A survey of sequential Monte Carlo methods for economics and Kavtaradze L, Mokhtari M (2018) Factor models and time-varying parameter
finance. Econom Rev 31(3):245–296 framework for forecasting exchange rates and inflation: a survey. J Econ Surv
Cuthbertson K, Nitzsche D (2004) Quantitative financial economics: stocks, bonds 32(2):302–334
and foreign exchange financial economics and quantitative analysis series, Kolasa M, Rubaszek M, Ca’ Zorzi M (2017) Exchange rate forecasting with DSGE
2nd edn. John Wiley & Sons Inc models. NBP Working Papers 260. Narodowy Bank Polski, Economic
Dash R (2018) Performance analysis of a higher order neural network with an Research Department
improved shuffled frog leaping algorithm for currency exchange rate pre- Kouri P (1978) The balance of payments and the foreign exchange market: a
diction. Appl Soft Comput 67:215–231 dynamic partial equilibrium model, a revised version of Kouri (1976b),
Das SR, Mishra D, Rout M (2020) A hybridized ELM-Jaya forecasting model for Cowles Foundation discussion paper no. 510. In: Bhandari J, Levin J (eds)
currency exchange prediction. J King Saud Univ-Comput Inf Sci Economic interdependence and flexible exchange rates. MIT Press, pp
32(3):345–366 329–370 (1983 reprinted in de Macedo and Lempinen, 2011)
Das SR, Mishra D, Rout M (2019) An optimized feature reduction based currency Lee H (2011) Nonlinear exchange rate dynamics under stochastic official inter-
forecasting model exploring the online sequential extreme learning machine vention. Econ Model 28(4):1510–1518
and krill herd strategies. Physica A 513:339–370 Lux T (2018) Estimation of agent-based models using sequential Monte Carlo
De Almeida BJ, Neves RF, Horta N (2018) Combining support vector machine methods. J Econ Dyn Control 91:391–408
with genetic algorithms to optimize investments in Forex markets with high Martino L, Elvira V (2021) Compressed Monte Carlo with application in particle
leverage. Appl Soft Comput 64:596–613 filtering. Inf Sci 553:331–352
Demir F, Razmi A (2022) The real exchange rate and development theory, evi- Meese R, Rogoff K (1988) Was it real? The exchange rate-interest differential
dence, issues and challenges. J Econ Surv 36(2):386–428 relation over the modern floating-rate. J Finance 43(4):933–948
Diebold FX, Mariano R (1995) Comparing predictive accuracy. J Bus Econ Stat Molodtsova T, Papell DH (2009) Out-of-sample exchange rate predictability with
13:253–265 Taylor rule fundamentals. J Int Econ 77(2):167–180
Dornbusch R (1976) Expectations and exchange rate dynamics. J Political Econ Motta M, Zhang S, Chan GK (2019) Hamiltonian symmetries in auxiliary-field
84:1161–1176 quantum Monte Carlo calculations for electronic structure. Phys Rev B
Dymova L, Sevastjanov P, Kaczmarek K (2016) A Forex trading expert system 100:045127
based on a new approach to the rule-base evidential reasoning. Expert Syst Neureiter N, Ranacher P, Efrat-Kowalsky N et al. (2022) Detecting contact in
Appl 51:1–13 language trees: a Bayesian phylogenetic model with horizontal transfer.
Eichengreen B, Rose AK, Wyplosz C (1994) Speculative attacks on pegged Humanit Soc Sci Commun 9:205
exchange rates: an empirical exploration with special reference to the Eur- Nguyen MH, Jones TE (2022) Building eco-surplus culture among urban residents
opean monetary system. NBER Working Papers 4898. National Bureau of as a novel strategy to improve finance for conservation in protected areas.
Economic Research, Inc Humanit Soc Sci Commun 9:426
Fan MH, Chen MY, Liao EC (2021) A deep learning approach for financial market Nguyen H, Shi H, Xu J, Hang S (2014) CPMC-Lab: a Matlab package for Constrained
prediction: utilization of Google trends and keywords. Granul Comput Path Monte Carlo calculations. Comput Phys Commun 185(12):3344–3357
6:207–216 Ni L, Li Y, Wang X, Zhang J, Yu J, Qi C (2019) Forecasting of forex time series data
Fisher I (1962) The Rate of Interest. Garland Publishing, New York based on deep learning. Procedia Comput Sci 147:647–652
Flood R, Marion N (1997) The size and timing of devaluation in capital controlled Orus R, Mugel S, Lizaso E (2019) Quantum computing for finance: overview and
economies. J Dev Econ 54(1):123–147 prospects. Rev Phys 4:100028
Fournié E, Lasry JM, Lebuchoux J, Lions PL (2001) Applications of Malliavin Park C, Park S (2013) Exchange rate predictability and a monetary model with
calculus to Monte-Carlo methods in finance II. Finance Stoch 5:201–236 time-varying cointegration coefficients. J Int Money Finance 37:394–410
Fu S, Li Y, Sun S, Li H (2019) Evolutionary support vector machine for RMB Parot A, Michell K, Kristjanpoller WD (2019) Using artificial neural networks to
exchange rate forecasting. Physica A 521(1):692–704 forecast exchange rate, including VAR‐VECM residual analysis and predic-
Giacomini R, Rossi B (2010) Forecast comparisons in unstable environments. J tion linear combination. Intell Syst Account Finance Manag 26(1):3–15
Appl Econom 25(4):595–620 Parungrojrat N, Kidsom A (2019) Stock price forecasting: geometric Brownian motion
Gordon NJ, Salmond DJ, Smith AF (1993) Novel approach to nonlinear/non- and Monte Carlo simulation techniques. MUT J Bus Adm 16(1):90–103
Gaussian Bayesian state estimation. In: IEE proceedings F (radar and signal Raimundo MS, Okamoto J (2018) SVR-wavelet adaptive model for forecasting
processing) pp. 107–113 financial time series. In: 2018 International Conference on Information and
Guo T (2022) Black-Scholes process and Monte Carlo simulation-based options Computer Technologies (ICICT). IEEE, pp. 111–114
pricing. In: 2022 International Conference on Artificial Intelligence, Internet Rebentrost P, Gupt B, Bromley TR (2018) Quantum computational finance: Monte
and Digital Economy (ICAID 2022). Atlantis Press, pp. 733–741 Carlo pricing of financial derivatives. Phys Rev A 98(2):022321
Haario H, Laine M, Mira A, Saksman E (2006) DRAM: efficient adaptive MCMC. Rossi B (2013) Exchange rate predictability. J Econ Lit 51(4):1063–1119
Stat Comput 16:339–354 Rubaszek M, Ca’ Zorzi M (2020) Exchange rate forecasting on a napkin. J Int
Hajizadeh E, Mahootchi M, Esfahanipour A, Kh MM (2019) A new NN-PSO Money Finance 104:102168
hybrid model for forecasting Euro/Dollar exchange rate volatility. Neural Serjam C, Sakurai A (2018) Analyzing predictive performance of linear models on
Comput Appl 31:2063–2071 high-frequency currency exchange rates. Vietnam J Comput Sci 5:123–132
Hauzenberger N, Huber F (2019) Model instability in predictive exchange rate Sikora G, Michalak A, Bielak Ł, Miśta P, Wyłomańska A (2019) Stochastic mod-
regressions. J Forecast 39(2):168–186 eling of currency exchange rates with novel validation techniques. Physica A
523:1202–1215

20 HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | (2023)10:353 | https://doi.org/10.1057/s41599-023-01836-2

Content courtesy of Springer Nature, terms of use apply. Rights reserved.
HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | https://doi.org/10.1057/s41599-023-01836-2 ARTICLE

Sun S, Wang S, Wei Y (2019) A new multiscale decomposition ensemble approach Competing interests
for forecasting exchange rates. Econ Model 81(C):49–58 The authors declare no competing interests.
Staum J (2009) Monte Carlo computation in finance. In: L’ Ecuyer P, Owen AB
(eds), Monte Carlo and Quasi-Monte Carlo methods 2008. Springer, Berlin,
Heidelberg, pp. 19–42
Ethical approval
The authors declare that they did not involve human or animal participants (and their
Taveeapiradeecharoen P, Chamnongthai K, Aunsri N (2019) Bayesian compressed
data) in the present study.
vector autoregression for financial time-series analysis and forecasting. IEEE
Access 7:16777–16786
Thu TNT, Xuan VD (2018) Using support vector machine in FoRex predicting. In: Informed consent
Joyo MK (ed), 2018 IEEE International Conference on Innovative Research The authors declare that they did not involve human participants in the study, and
and Development (ICIRD). IEEE, pp. 1–5 therefore it was not necessary to obtain any consent about this kind of data.
Tigani S, Tadist K, Saadane R, Chehri A, Chaibi H (2022) Deep learning based
currency exchange volatility classifier for best trading time recommendation.
Procedia Comput Sci 207:1591–1597 Additional information
Valchev R (2020) Bond convenience yields and exchange rate dynamics. American. Supplementary information The online version contains supplementary material
Am Econ J Macroecon 12(2):124–166 available at https://doi.org/10.1057/s41599-023-01836-2.
Wang P, Zong L, Ma Y (2020) An integrated early warning system for stock market
turbulence. Expert Syst Appl 153:113463 Correspondence and requests for materials should be addressed to David Alaminos.
Wei Y, Sun S, Ma J, Wang S, Lai KK (2019) A decomposition clustering ensemble
learning approach for forecasting foreign exchange rates. J Manag Sci Eng Reprints and permission information is available at http://www.nature.com/reprints
4(1):45–54. 2019
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in
Wei W, Zhu D (2022) Generic improvements to least squares monte carlo methods
published maps and institutional affiliations.
with applications to optimal stopping problems. Eur J Oper Res
298(3):1132–1144
Zhang S (2004) Quantum Monte Carlo Methods for Strongly Correlated Electron
Systems. In: Sénéchal D, Tremblay AM, Bourbonnais C (eds) Theoretical Open Access This article is licensed under a Creative Commons
methods for strongly correlated electrons. CRM series in mathematical Attribution 4.0 International License, which permits use, sharing,
physics. Springer, New York, NY adaptation, distribution and reproduction in any medium or format, as long as you give
Zheng J, Fu X, Zhang G (2019) Research on exchange rate forecasting based on appropriate credit to the original author(s) and the source, provide a link to the Creative
deep belief network. Neural Comput Appl 31:573–582 Commons license, and indicate if changes were made. The images or other third party
material in this article are included in the article’s Creative Commons license, unless
indicated otherwise in a credit line to the material. If material is not included in the
article’s Creative Commons license and your intended use is not permitted by statutory
Acknowledgements regulation or exceeds the permitted use, you will need to obtain permission directly from
This research was funded by the Universidad de Málaga. We would also like to thank the
the copyright holder. To view a copy of this license, visit http://creativecommons.org/
Universitat de Barcelona, UB-AE-AS017634, for funding this research.
licenses/by/4.0/.

Author contributions © The Author(s) 2023

All the authors contributed equally to completing the manuscript.

HUMANITIES AND SOCIAL SCIENCES COMMUNICATIONS | (2023)10:353 | https://doi.org/10.1057/s41599-023-01836-2 21

Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Terms and Conditions
Springer Nature journal content, brought to you courtesy of Springer Nature Customer Service Center GmbH (“Springer Nature”).
Springer Nature supports a reasonable amount of sharing of research papers by authors, subscribers and authorised users (“Users”), for small-
scale personal, non-commercial use provided that all copyright, trade and service marks and other proprietary notices are maintained. By
accessing, sharing, receiving or otherwise using the Springer Nature journal content you agree to these terms of use (“Terms”). For these
purposes, Springer Nature considers academic use (by researchers and students) to be non-commercial.
These Terms are supplementary and will apply in addition to any applicable website terms and conditions, a relevant site licence or a personal
subscription. These Terms will prevail over any conflict or ambiguity with regards to the relevant terms, a site licence or a personal subscription
(to the extent of the conflict or ambiguity only). For Creative Commons-licensed articles, the terms of the Creative Commons license used will
apply.
We collect and use personal data to provide access to the Springer Nature journal content. We may also use these personal data internally within
ResearchGate and Springer Nature and as agreed share it, in an anonymised way, for purposes of tracking, analysis and reporting. We will not
otherwise disclose your personal data outside the ResearchGate or the Springer Nature group of companies unless we have your permission as
detailed in the Privacy Policy.
While Users may use the Springer Nature journal content for small scale, personal non-commercial use, it is important to note that Users may
not:

1. use such content for the purpose of providing other users with access on a regular or large scale basis or as a means to circumvent access
control;
2. use such content where to do so would be considered a criminal or statutory offence in any jurisdiction, or gives rise to civil liability, or is
otherwise unlawful;
3. falsely or misleadingly imply or suggest endorsement, approval , sponsorship, or association unless explicitly agreed to by Springer Nature in
writing;
4. use bots or other automated methods to access the content or redirect messages
5. override any security feature or exclusionary protocol; or
6. share the content in order to create substitute for Springer Nature products or services or a systematic database of Springer Nature journal
content.
In line with the restriction against commercial use, Springer Nature does not permit the creation of a product or service that creates revenue,
royalties, rent or income from our content or its inclusion as part of a paid for service or for other commercial gain. Springer Nature journal
content cannot be used for inter-library loans and librarians may not upload Springer Nature journal content on a large scale into their, or any
other, institutional repository.
These terms of use are reviewed regularly and may be amended at any time. Springer Nature is not obligated to publish any information or
content on this website and may remove it or features or functionality at our sole discretion, at any time with or without notice. Springer Nature
may revoke this licence to you at any time and remove access to any copies of the Springer Nature journal content which have been saved.
To the fullest extent permitted by law, Springer Nature makes no warranties, representations or guarantees to Users, either express or implied
with respect to the Springer nature journal content and all parties disclaim and waive any implied warranties or warranties imposed by law,
including merchantability or fitness for any particular purpose.
Please note that these rights do not automatically extend to content, data or other material published by Springer Nature that may be licensed
from third parties.
If you would like to use or distribute our Springer Nature journal content to a wider audience or on a regular basis or in any other manner not
expressly permitted by these Terms, please contact Springer Nature at

[email protected]