Machine Learning for Financial Prediction Under Regime

Change Using Technical Analysis: A Systematic Review

Andrés L. Suárez-Cetrulo1, David Quintana2, Alejandro Cervantes3*
1
Ireland’s Centre for Applied AI (CeADAR), University College Dublin (Ireland)
2
Department of Computer Science and Engineering, Universidad Carlos III de Madrid, Avda. Universidad 30, 28911 Leganes
(Spain)
3
Escuela Superior de Ingeniería y Tecnología, Universidad Internacional de La Rioja (UNIR), Logroño
(Spain)

Received 26 May 2022 | Accepted 28 April 2023 | Early Access 23 June 2023

Abstract Keywords
Recent crises, recessions and bubbles have stressed the non-stationary nature and the presence of drastic Concept Drift, Finance,
structural changes in the financial domain. The most recent literature suggests the use of conventional Machine Learning,
machine learning and statistical approaches in this context. Unfortunately, several of these techniques are Meta Learning, Regime
unable or slow to adapt to changes in the price-generation process. This study aims to survey the relevant Change, Systematic
literature on Machine Learning for financial prediction under regime change employing a systematic approach. Literature Review.
It reviews key papers with a special emphasis on technical analysis. The study discusses the growing number
of contributions that are bridging the gap between two separate communities, one focused on data stream
learning and the other on economic research. However, it also makes apparent that we are still in an early stage
. The range of machine learning algorithms that have been tested in this domain is very wide, but the results of DOI: 10.9781/ijimai.2023.06.003
the study do not suggest that currently there is a specific technique that is clearly dominant.

I. Introduction covered by the marked efficiency hypothesis [25], we cannot observe

the individual behaviour of a trader or its intentions. Instead, we can

F inancial markets can be described as an evolutionary and nonlinear

dynamical complex system [1], [2]. Forecasting in the financial
domain has traditionally been performed under the assumption that
only observe changes in the price dynamics and macro or micro-
economic variables and extrapolate the changes that make them
modify their behaviour. The execution of these strategies is the actual
the underlying data has been created by a linear process [3]. Another generative process of the observed time series of prices or trends. The
line of work to make financial predictions is to use machine learning estimation of the hidden processes driving the market into different
(ML). These algorithms have surprised financial experts [4]–[6] regimes is often approached using regime-switching models, a type
because of their success in mapping nonlinear relationships without of time series model where parameters can have different values in
prior knowledge [7]. Deep learning algorithms (neural networks) and different cycles [26].
ensembles have been some of the techniques obtaining the best results Despite the fact that artificial intelligence has recently become a
for stock trend prediction [8]–[13]. trend and even a buzzword in many industries, this has not become
Different crises, recessions and bubbles, such as the COVID-19 the main trend yet for trading systems. This is mainly due to the high
pandemic, or volatile mid-term trends in crypto markets, have made complexity and hard explainability of these models [27], being the
apparent the non-stationary nature and the presence of drastic second a must for stakeholders and decision-makers in this sector [28].
structural changes in financial markets [14]. During these periods, Instead, traders tend to identify directional changes in the market state
mean returns, volatility and correlations among assets tend to change using different popular indicators tailored according to their needs.
quickly [15]. This has brought attention to the problem of concept drift Traditionally, the literature has used static methods to interpret
[16] in computational finance [17]. Many recent research works point patterns based on the meaning of these indicators and their historical
out that financial assets or companies present different states that may correlation to future prices. However, this correlation may vary over
repeat or not overtime or evolve due to inflation, deflation, or changes time. Behavioural shifts of investors changing continuously with
in supply and demand [18]–[24]. a hidden context can also be observed through the change in sell
In finance, a change in the collective behaviour of market versus buy volumes, in differences between local minima and local
participants and their reactions is called a regime change (RC). As maxima over time, and through different moving averages at different
time frames depending on the granular detail observed (frequencies)
at intraday, daily or weekly levels. Changes in financial markets
* Corresponding author. challenge traders and investors, as most of their models rely on
E-mail address: [email protected] previous patterns. Hence, a way to recognise these changes provides

Please cite this article in press as:

-1-
A. L. Suárez-Cetrulo, D. Quintana, A. Cervantes. Machine Learning for Financial Prediction Under Regime Change Using Technical Analysis: A Systematic
Review, International Journal of Interactive Multimedia and Artificial Intelligence, (2023), http://dx.doi.org/10.9781/ijimai.2023.06.003
 International Journal of Interactive Multimedia and Artificial Intelligence

a competitive advantage since it allows changes in trading strategies it is hard to keep track of the main contributions and instruments used
ahead of other investors [29]. Detecting concept shifts also helps lower to tackle the problem.
the risks of financial exposure in high-frequency trading (HFT). The literature presents a lack of studies on prediction under regime
The digitalisation of the financial industry has resulted in a growing changes based on technical analysis using machine learning. This is
amount of data that is available for decision-making. This, together unfortunate, as there is a lot to be gained in terms of efficiency and
with the increasing amount of computational resources, has accelerated performance. Within this field, we find very promising ideas. For
the adoption of a whole range of machine-learning-based solutions. instance, the problem can be framed using the data stream learning
Among the suite of instruments available to deal with regime changes, topic of concept drift. A significant number of contributions to this
online incremental ML algorithms seem especially appropriate. Among new concept have not been explicitly applied to finance yet, and they
the advantages that they offer, we can mention the fact they can handle are not widely known. They have not been widely present in machine
non-stationarities, shifts, and drifts in price generation processes. learning surveys outside the data stream learning niche area.
Another aspect that makes them a good fit for this context is the fact Exploring previous research showed that a comprehensive review
that they are scalable for continuous learning scenarios [30]. does not exist on these topics. Therefore, this study will help readers
One might consider two main scenarios regarding the nature of understand the current state of the art, bridge the gap among research
structural change. The first possibility is the existence of recurrence, fields, and address promising future lines of research in this domain.
that is, the idea that the system might transition back to a previous
price generation process. For instance, there might be a specific B. Research Method
market state for market openings at intraday frequencies and another In order to provide an overview of the state of the question, our
for financial bubbles that might be observable at lower frequencies. research has followed Kitchenham and Charters’ guidelines on
The alternative assumes that any drift results in a transition to a new Systematic Literature Review (SLR) [46], [47].
process. As we will discuss in detail, while there is a relevant number A systematic review is defined as an organised way to synthesise
of published studies on machine learning for data streams that pay existing work fairly. An SLR is a means to identify, evaluate and
attention to non-stationarity [31]–[35], the literature on financial interpret the available research works relevant to a definite research
applications of these algorithms, especially that focused on recurring question, topic area, or phenomenon of interest. After revising the
concept drifts, is more limited [17], [36]–[41]. literature for similar research objectives, it can be identified that there
We must also point out that the prediction of future financial is no previously published search on a topic.
trends can be tackled using fundamental or technical analysis. Despite
C. Planning
some controversy regarding its potential [25], [42], the latter is very
prevalent in short-term trading [43], hence the focus on this approach. The study aims to summarise the current status of predicting
Having said that, there are also relevant papers in the first category, financial time series in the financial literature during behavioural or
like the contributions of Geva and Zahavi [44], on the short-term, regime changes in markets. Kitchenham and Charters’ SLR protocol
intraday and high-frequency forecast using news data and the study was adapted to describe the plan for the review.
of Dogra et al. [4], analysing the impact of recent news on stock price The protocol comprises research background and questions, search
trends and challenges such as class imbalance. More recently, Chen strategy, study selection criteria and procedures, data extraction, and data
et al. [45] hybridised both approaches in a study that combines both synthesis strategies to guarantee that the investigation is undertaken as
sentiment analysis and technical indicators. intended and reduce the likelihood of bias in the study. In this protocol,
With this survey, we try to bring to the academic community’s the entire investigation plan was not decided from the beginning. Instead,
attention how ML is being used to deal with structural change in this and the results produced were recorded as the study progressed.
financial markets. Our goal is to identify directions on leveraging the
D. Research Questions
benefits of modern algorithms that work with different scenarios and
deal with any changes that may arise in real-time. This paper has the following two research questions:
The use of these approaches may help to find strategies to improve Q1 What are the different research areas for predicting under regime
prediction accuracy during times of change, limiting the need for changes in the financial literature? and
constant model retraining. Hence, some of these techniques efficiently Q2 What are the most commonly used machine learning techniques
increase the potential profits, avoiding the computational burden and applied to analysing regime changes?
benefiting from always having an up-to-date model. The results expected at the end of the systematic review were to
The rest of the document is structured as follows: Section II covers see what research or surveys had been applied or produced on the
the methodology and research questions used in this systematic topic so far and to identify the implications of using machine learning
review; Section IV will discuss the outcomes of each of the research to handle behavioural changes in financial markets in the scientific
questions. There, Subsection A will introduce the topic of regime literature.
change in financial series, and Section B will discuss the core literature
on machine learning for financial prediction under regime change. E. Search Strategy and Process
The final two sections will be reserved for a summary of results, main The search strategy included: i) search resources and ii) a search
conclusions, and future research lines. process. Each one of them is detailed in the following subsections.
1. Search Resources
II. Research Methodology
This study was planned to find all the literature available about
machine learning for forecasting under regime changes in finance.
A. Motivation and Objectives The sources used for the systematic review were:
Financial time series are often subject to structural change. Even
• IEEE Digital Library (http://ieeexplore.ieee.org);
though machine learning offers major advantages in this context, the
literature on the topic is limited and sparse. There seem to be different • ScienceDirect, on the subject of Computer Science (https://www.
research communities focused on different aspects of the problem, and sciencedirect.com/);

-2-
 Article in Press

• ACM Digital Library (http://portal.acm.org); of a random sample of the primary studies were made.
• Taylor & Francis Journals (http://www.tandfonline.com); Documents were kept when they satisfied at least one of the criteria
• Wiley Online Library (http://www.wiley.com/); below:
• SpringerLink (http://link.springer.com); and additionally • The work was explicitly related to regime changes or structural
breaks in non-stationary data.
• Google Scholar was explored as grey literature (https: //scholar.
google.com/). • The work was relevant to machine learning forecasting in
domains with complex dynamics and non-stationarities in the
2. Search Process financial field.
The overall search process is depicted in Fig. 1 and is explained in The authors reviewed all 223 research works and put them into
the following section. these different groups according to the previously mentioned criteria.
This list was reviewed to check for inconsistencies. The result of this
Identification of studies via databases stage was that 140 publications were classified as relevant.
Records removed before There is a risk that some relevant works have been missed.
screening: Therefore, this study cannot guarantee completeness. However, it can
Identification

Records identified from: Duplicate record removed

Databases (n = 7)
still be trusted to give a good overview of the relevant literature on
(n = 20)
Registers (n = 0) Records removed for not price forecasting in the financial domain under structural breaks.
being available or not having
a subscription (n = 8) 3. Data Extraction
The data extracted from each publication was documented and kept
in a reference manager. After the identification of the publications, the
Records screened Records excluded based on title following was extracted:
(n = 643) (n = 217)
• Source (journal, book, conference or strictly relevant technical or
white paper);
Reports not retrieved after • Title;
Reports sought for retreieval
reading abstract and keywords
(n = 426)
Screening

(n = 203) • Publication year;

• Authors;
Reports excluded based on full • Classification according to topics;
Reports assessed for eligibility
text:
(n = 223)
• Studies only focusing on ML • Summary of the research, including which questions were solved.
methods (n = 61)
• Financial regime change
works not focused in III. Summary of Results
techniques (n = 22)
In order to analyse the 223 works, we found the need to classify
Included

Studies included in review them in more ways than just according to the methodology defined in
(n = 140) Section II. When needed, the topics were updated or clarified during
the classification process. Results of the classification process with
regard to the research questions are detailed in Table I.
Fig. 1. Flow of information through the different phases of the review using
a PRISMA diagram [48]. TABLE I. Classification of Papers With Regard to the Research
Questions
The starting point was choosing a set of relevant keywords. They
Question Topic Relevant Studies Quantity
were: regime change, regime-switching model, machine learning, change
detection and stock trend forecasting. The search was then run on the Q1 Regime changes [14], [15], [18], [19], [23], 22
already mentioned databases in March 2022, returning 643 works [26], [29], [49]–[63]
in total in a time range, including the years 1970 to 2022. Irrelevant Q1 and Q2 ML in stock [1]–[13], [20]–[22], [24], 73
and duplicate publications were removed, and 223 unique research forecasting [25], [27], [28], [37], [38],
works remained. At that point, publications were reviewed based on [42]–[45], [64]–[110]
titles, abstracts, conclusions, references and keywords and then were Q2 Concept drift [16], [17], [30]–[36], 45
classified into three different types: and online ML [39]–[41], [111]–[143]
• Relevant works: these should satisfy one of the two inclusion
criteria covered later in this subsection; The data required for analysis were extracted by exploring the
• Process assessment works: if the publication is related to the full text of each research work. Table II presents the results of the
financial domain or concept drift literature and is relevant. search and the source of the documents. Table III presents the results
in the second stage. As mentioned before, the total number of papers
• Excluded works: works not relevant to the topic. remaining after the exclusion process was 140. Table I summarises
When there was doubt about the classification of a research their classification according to the knowledge area.
workpiece, it was included in the relevant group, leaving the possibility The relevance of regime changes or structural breaks in the literature
of discarding it during the next stage, when the full-text versions were of financial price forecasting leads to consider two major areas:
reviewed. Third, each full article was retrieved and read to verify its financial regime changes (related to majorly statistical approaches
inclusion or exclusion. The reason for exclusion or inclusion in this to detect change points or forecast under different regimes) and data
third stage was documented. Fourth, to check the consistency of the stream learning (where the problem of concept drift can be understood
inclusion/exclusion decisions, a test-retest approach and re-evaluation as a type of regime change in the ML literature).

-3-
 International Journal of Interactive Multimedia and Artificial Intelligence

TABLE II. Results Without Filtering 78.57%

Data Source Total Publications

ScienceDirect 76
Google Scholar 56

Frequency (Log scale)

Springerlink 33
IEEE Digital Library 27
10.00%
ACM Digital Library 14 7.86%
Wiley 9
Taylor & Francis 8

TABLE III. Second Stage Results 2.14%

Data Source Total Publications 1.43%

ScienceDirect 56
Journal Conference Book Technical White paper
Google Scholar 34 report
Springerlink 20
IEEE Digital Library 16 Fig. 3. Source distribution of research papers after filtering.
Taylor & Francis 5
Wiley 5 25.00%
ACM Digital Library 4

Fig. 2 shows how, out of a total of 140 relevant studies, the majority
of the works reviewed to correspond to ML techniques applied to stock
17.86%
forecasting. Some of these works overlap regime change research,
focusing primarily on probabilistic models to classify directional Frequency
changes and represent different regimes. The literature on online
12.50% 12.50%
learning does not tend to coincide with the one on regime changes.
However, studies of online ML tackle similar challenges as models
8.93% 8.93%
to handle regime changes, such as having up-to-date models and re-
7.14% 7.14%
training mechanisms. A deeper discussion on this matter will be held
in Section IV.

Regime changes
Concept drift
15.7% Ensembles Decision Evolving Neural Bayesian Evolving Fuzzy Neuro KNN
Trees (single) Fuzzy Networks and Markov Clustering Systems and SVM
32.1% Rules

Fig. 4. ML techniques found in Concept Drift studies, grouped by categories,

counting each different technique used in papers and assigning the
52.1% corresponding category separately. In this figure, a single paper comparing
several algorithms in the same category is counted as many times as algorithms.
ML in stock forecasting

15.56% 15.56% 15.56% 15.56%

Fig. 2. Topic distribution of research papers after filtering.

Fig. 3 shows the distribution of papers reviewed across various

sources. A majority of the research works have been retrieved from
11.11%
high-impact journals, followed by conferences and books. However,
since some of the topics reviewed, like online ML, are current research
Frequency

8.89% 8.89% 8.89%

areas, a remainder, close to « 4% of research works, belong to non-
peer-reviewed papers contained by open-access repositories.
Finally, we have extracted from the papers classified papers under
the topic "Concept Drift" the ML technique mainly used, either as
a new proposal or as a reference for comparison. We have grouped
these techniques into eight broad categories (Fig. 4 and 5). For this
task, we have excluded reviews. These results show that most of the
reviewed papers use techniques from four main categories: Evolving
systems (that include Evolving clustering, Evolving fuzzy rules and
Decision Ensembles Evolving Neural Bayesian Evolving Fuzzy Neuro KNN
Fuzzy neuro systems), Ensemble based systems (usually with tree- Trees (single) Fuzzy Networks and Markov Clustering Systems and SVM
Rules
based components), traditional systems adapted to concept change
(such as adaptive decision trees), and finally Neural Networks and Fig. 5. Types of ML found in Concept Drift studies, using the unique categories
Deep Learning. The latter are more recent in general, and therefore found in the same paper. In this figure, a single paper comparing six methods
this trend is likely to become more important in the near future. Fig. 6 in category A and one method in category B is counted as only two entries
(one for A and another for B).
shows the evolution of these categories.

-4-
 Article in Press

Bayesian and Markov

In the financial literature, changes in the price behaviour of financial
25 Decision Trees (adapted) markets that go beyond their normal price fluctuations receive the
Ensembles name of regime changes [19], [53],
ML technique in collection

Evolving Clustering
20
Evolving Fuzzy Rules [63] or business cycles shifts [80]. In order to model these regime
Fuzzy Neuro Systems changes, one of the most popular techniques is the regime-switching
15 KNN and SVM
Neural Networks
model [15], which was first applied by Hamilton [58] as a technique
to deal with cycles of different economic activities such as recessions
10
and market expansions.
5 In financial markets, there are periods of time with different
degrees of efficiency and predictability. There can be moments where,
0 due to the market-wide sentiment given by political or economic
<2005 2005-2015 >2015
circumstances, the behaviour of investors may change towards a bear,
Years
bull, lateral market, and periods or time frames with different levels of
Fig. 6. Categories per period of 10 years based on data used in Fig. 5. volatility [80].
At the macroeconomic level, RC are often related to abrupt breaks
in long-term cycles like the break of bubbles or economic crises
IV. Discussion [59]. Changes in market regimes could be driven as well by investor
This section describes the papers reviewed in this work. In this expectations [15]. The financial literature identifies two types of
discussion, we follow the schema in Fig. 7. regimes clear to recognise: steady and highly volatile regimes usually
linked to economic growth or deflation periods, respectively.
A. Regime Changes in Financial Series (Q1) This is illustrated in Fig. 8, which shows the breaks identified in
Early studies from the financial literature claim that financial [19] during the Great Recession.
markets are efficient [25] and, as a result, asset prices follow a random
walk [81]. Fama [25] claimed that markets cannot be consistently 0.45
beaten on a risk-adjusted basis and that their prices cannot be
anticipated has always been a source of controversy in the literature.
0.1
Many research works have pointed to different markets being
predictable using different sources of information [5], [77], [78], [85],
Log Daily Returns

[88], [104]. 0.05

Forecasting in the financial domain can be characterised by a
non-stationary and unstructured nature and by hidden relationships
[2], [74]. Economic, social and political factors within countries 0
and international impact add uncertainty to financial markets [66],
[70], [79], [93], [94], [100], [106]. Hence, markets can be considered
-0.05
an evolutionary and nonlinear complex system [1]. The financial
literature has covered different approaches to predicting market prices
using statistical and, more recently, AI-based methods. -0.1
07 08 09 10 11 12
In recent years, different events like the COVID-19 pandemic or n 20 n 20 n 20 n 20 n 20 n 20
1 Ja 1 Ja 1 Ja 1 Ja 1 Ja 1 Ja
the bankruptcy of Lehman Brothers in 2008 have led to periods with
changes in mean, volatility and correlations in stock market returns Fig. 8. Regime Changes in the DJIA Index (indicator of the United States
economy) identified by Tsang and Chen [19].
[15], stressing the non-stationary nature and the existence of drastic
structural changes in financial markets [18], [20]–[23].

Regime Change in Financial Prediction

Machine learning

Regime Change Basic concepts Techniques

Predictability Data preprocessing Traditional ML

External impact Dealing with change Deep Learning

Characterization of change
Classical techniques Predictability in presence of change Ensembles and meta-learning
Regime characterization using non supervised learning
Evolving Intelligent Systems
Fig. 7. Discussion on the research papers considered in this work.

-5-
 International Journal of Interactive Multimedia and Artificial Intelligence

Regime changes challenge investors, making them change their Authors like Patel et al. [13] discretise features based on the human
trading strategies as the collective trading behaviour of the market approach to investing and deriving the technical indicators using
changes. Different examples of RC have been covered in the recent assumptions from the stock market. This latter approach, though,
literature. Davies [53] analysed different cases and consequences of introduces human bias in the process. This is the opposite way to
regime changes in the Great Recession that impacted several asset approach the problem of trend prediction if we compare it to recent
classes such as equities, bonds, commodities and currencies at micro deep learning strategies that feed dozens of automatically generated
and macroeconomic levels. Hamilton [63] observed alternating indicators [109].
patterns between steady and turbulent periods since the Second World Overall, the above-mentioned literature reviews confirm that ML
War and subsequent recessions by looking at US unemployment rates. techniques can be used to predict price changes, but this entirely
Ang and Timmermann [15] identified cyclic changes in the behaviour depends on the time horizon and efficiency of the market in the period
of asset prices and mean, volatility and correlation patterns in stock predicted. Cavalcante et al. [3] provided another interesting review of
returns during the Great Recession and the 1973 oil crisis. Kritzman et pre-processing and clustering techniques used in the financial domain
al. [29] discovered that investors could benefit from having different to forecast future market movements. They highlighted the relevance
asset allocation strategies in different market regimes to minimise of concept drifts in financial markets and suggested that the data
losses. stream mining literature is of great importance in future research due
Many other studies consider these drastic changes an intrinsic to the non-stationarity and evolution of financial markets [91].
characteristic of financial data that might be caused by significant In computational finance, changes in the behaviour of the market
events, and thus, these will be observable not only in prices and are normally referred to as regime changes or switches [15], [61], [67],
economic variables but also in other kinds of public information structural breaks or changes [51], [54], [56], [56], volatility shifts [50],
[52], [58], [59], [62], [63], [92]. Hamilton [58] proposed a time-series switching processes [60] or market states [18]. In this kind of data,
based approach [26] to capture nonlinear effects like RC, identify long periods of stability might be interrupted by short episodes of
market breaks and hidden changes in economic cycles known as abrupt changes [61].
the regime-switching model [59]. This model, also known as the
These changes may or not be transitory since a newly adopted
Markov-switching model, is fitted to observations following different
behaviour in price dynamics, reflected as part of the mean returns,
patterns in different periods and is mainly applied to recognise low
their volatility or correlation among them may persist for several
volatility regimes with economic growth vs high volatility periods
periods. Timely recognition of these sudden behavioural changes in
with economic contractions [116]. Ang and Timmermann [15]
markets can significantly lower the risk of financial exposure. This has
applied these models to predict interest rates and equity and foreign
inspired the materialisation of techniques such as regime-switching
exchange returns. They discussed how to model RCs for these time
models in the financial literature, which work under the premise that
series models.
new dynamics of price returns and fundamentals persist for several
B. Approaches Based on Machine Learning (Q2) periods after a change. A key element in these models is identifying
Traditional statistical methods tend to model and predict future whether the exact market regimes reoccur over time (e.g. across
data based on the assumption that the time series under study is recessions or periods of economic growth) or if new regimes deviate
generated from a linear process with features normally distributed or have evolved from previous ones [15].
[144]. This presents challenges since financial data is characterised by The prediction of future values in financial markets and the
nonlinearity and non-stationarity besides a high level of uncertainty detection of regime changes in data streams with temporal dependence
and noise [82]. are common application areas for statistical methods and ML. Previous
A different approach to performing financial forecasts is the use of research reports high accuracy in forecasting price changes with
ML techniques. Several literature reviews show the benefits of these advanced techniques and the feasibility of making profits using these
techniques against traditional methods [5], [98], [99], [105], surprising predictions against the EMH, which points to unbeatable markets.
practitioners by contradicting early theories like the random walk, and An alternative theory is the adaptive market hypothesis (AMH)
efficient market hypothesis (EMH) [5], [25], [95]. Machine learning [57], introduced by Andrew Lo in 2004. This theory, with empirical
algorithms can handle nonlinear relationships without prior knowledge evidence in a increasing number of research works [68], combines
[7], outperforming traditional time series methods [10]–[13]. the EMH with principles from behavioural finance, allowing the ideas
of market efficiency and inefficiencies to co-exist. Under the AMH,
Different research works from the economic literature have the efficiency of a market evolves as market participants adapt to an
either used technical indicators or raw prices and returns. Technical environment that changes continuously. In this regard, participants
indicators are able to show behavioural patterns among traders and rely on heuristics to make their investment choice, leading to mostly
thus provide an extra level of signal to predictive models. These can be rational markets under those heuristics (like the EMH). The main
valuable to automate the behaviour of short-term traders [5]. In any difference is at the time of major behavioural shifts in the market
case, most of the economics literature has focused on linear processes participants, as in economic shocks or crises. In this case, the AMH
that may not have been able to extract relevant information nor infer considers a market that evolves, and the initially adaptive heuristics
complex relationships among technical indicators where some new may become static in certain market situations. Consequently, the
ML methods could [108]. EMH may not continue under periods of abnormal conditions, stress
Some of the literature reviews already cited describe common or abrupt changes in the market. Hence, financial markets may be
technical indicators used for stock market value and trend predictable in specific periods, as discussed by Lo [49]. Therefore,
prediction. Many of these papers, like [7], have shown that convergence to market efficiency is neither guaranteed nor likely to
different pre-processing steps, like the frequency level of the input occur. The level of efficiency depends on the market participants and
data, can impact its predictability. While a common approach in the market conditions at that time.
this regard is data normalisation, in the literature on data stream One of the few financial studies citing concept drift explicitly can
mining, data normalisation is not a usual practice since maximum be found in a recent work authored by Masegosa et al. [36]. They
and minimum values for each attribute in the data stream are analysed data from the Great Recession and claimed that economic
unknown beforehand [134].

-6-
 Article in Press

changes during this period manifested as concept drifts in their (HMM). Gomes et al. [122] also hypothesised about using Baum–
generative processes. An intermediate example of trying to predict Welch in conjunction with HMMs for continuous learning problems.
financial crises using ML methods can be found in [55], where the Baum–Welch has been used in the financial domain together with
authors studied possible contagion risks between financial markets other specific versions of EM and Gaussian mixture models (GMM)
that could trigger financial crises to signal warnings at an early stage. to forecast change direction in stock prices [72], [87] and to represent
More recently, Yang et al. [137] analysed the impact of concept drift market regimes [19], [29], [64], [86].
in business processes. More specifically, they modelled the response A set of relevant techniques from the ML literature in this regard are
to concept drift as a sequential decision-making problem by combing prototype generation techniques such as learning vector quantisation
a hierarchical Markov model and a Markov decision process. Martín [65], [96], [102], which have been proven to be useful for data
et al. [145] also dealt with structural changes introducing a trading partitioning and model selection in the financial domain. Choudhury
system based on grammatical evolution that commutes between an et al. [71], and Pavlidis et al. [69] use a combination of clustering and
active model and a candidate one to increase performance. forecasting algorithms to model the distribution of financial data.
Other two key research pieces in this regard are the works by Regarding the first step, the former authors use a two-layer abstraction
Tsang and Chen [19] and M "unnix et al. [18], which proposed that clusters stocks using self-organising maps (SOM) to then rely on
mechanisms to identify points of drastic changes in financial time K-means to obtain clusters of prototypes. The latter considers three
series. The former used statistical-based and traditional ML (e.g. naive different unsupervised algorithms to identify market states: growing
Bayes) approaches to classify normal versus abnormal regimes. They neural gas (GNG), density-based spatial clustering of applications with
proposed a framework based on the change speed of price returns and noise (DBSCAN), and Unsupervised k-Windows. Once the market
the degree of changes to visualise and discriminate between different states are identified, they use feed-forward neural networks to make
market regimes depending on the volatility of their price returns. The predictions.
latter [18] visualised differences in the correlation structure of the In the last years, several online incremental algorithms have used
price returns across assets in the S&P500 during the Great Recession. these techniques to adapt distinct learners to different cycles or
They extended the selection to a sample from 1992 to 2010, identifying seasonal behaviours in a data stream. In this regard, the use of online
eight market states repeating behavioural changes over time. ensembles, using non-supervised learning to represent different
Several approaches from the deep learning (DL) field (e.g. RNNs) behavioural patterns [135], [141] or supervised learning to train a pool
have also tried to face the problem of concept changes when learning of classifiers with high predictive accuracy under different conditions
continuously. [113], [126], [128], [130]. These advances have been [34], [117], have obtained state-of-the-art results adapting to different
successfully transferred to the financial domain, as discussed in recent states in the behaviour of data streams in many domains, including
surveys by Ozbayoglu et al. [9], and Li and Bastos [109]. finance [17].
Most of these mentioned research works focus on the likelihood Meta learners over data streams are a related subfield of ML of
of daily or monthly changes, where retraining a model is a feasible increasing popularity where a pool of former classifiers is managed
task. As of today, the amount of research devoted to seasonality and reused when the state or concept of the stream changes. This
and changes at the intraday level is significantly more limited [75], subfield is inspired by the human learning system that reuses
[89], [101], [103], [107]. The computational cost of ML and statistical previous knowledge to learn new tasks, not starting from zero every
methods, together with the inherent higher complexity derived from time. Although meta-learners have not been widely applied to the
the need to manage large amounts of data at these resolutions, makes financial domain yet, their logic resembles approaches applied to
keeping models up to date more challenging. finance as EM or Baum-Welch, but for continuous learning domains
While its application to high-frequency markets is still an open where models are always up-to-date and thus behave smoothly in
problem, recent research works are making progress in understanding case of structural breaks. For instance, Abad et al. [120] proposed
how to apply ML to intraday resolutions. Among them, we could a meta-learner that used hidden Markov models (HMM) to predict
mention the one presented by Sirignano and Cont [73], who claimed the sequence of change between discrete concepts. Their approach
that financial data at high frequencies exhibit stylised facts and may used fuzzy logic rules to compare classifiers to reuse former models.
hold learnable stationary patterns over long periods. Another relevant Maslov et al. [139] proposed a method to use patterns acquired
study is the one authored by Shintate and Pichl [76], who reviewed during previous changes and assumed a Gaussian distribution for
modern ML and DL approaches applied to high-frequency trading at the duration of the changes to predict the time of the next change
the minute level. point. Carta et al. in [83] recently combined the use of meta learners
with deep reinforcement learning to produce trading strategies and
Recently, several research works have approached the problem
maximise profits operating in Standard & Poor’s 500 future markets
of time-changing behaviours using non-supervised ML methods
and the J.P. Morgan and Microsoft stocks.
[127], [130], [135]. In these, micro-clusters or latent features may
be used to represent a summary of the incoming data and reduce Meta-learning approaches have inbuilt strategies to decide on when
the computational costs of correlating full data distributions. A to train, what model to replace and when, when to forget (prune) a
manner of doing this is using model-based clustering approaches. learner and when to create one [114], [136] by using the evaluation
These algorithms find models that fit input data and are also robust performance metrics of active and former models [149]. These, thus,
to the presence of noise [146], [147]. For instance, expectation are a closely related research area to evolving intelligent systems (EIS)
maximisation (EM) [148] fits a mixture of Gaussian distributions to [118], [129], [150] and evolving fuzzy
the data [133]. Chiu and Minku [141] used it in concept drift handling systems [125] [132] [143] [119]. EIS, which are also online and
based on clustering in the model space (CDCMS) to create concept incremental systems, can adapt themselves to concept drifts of
representations and keep a diverse ensemble learner. Zheng et al. [123] different natures on-the-fly through adaptive fuzzy rules [140]. EIS
relied on it to minimise intra-cluster dispersion and cluster impurity. have already demonstrated their ability to solve different kinds of
Tsang and Chen [19] applied the Baum–Welch algorithm, a special problems in various application domains like finance [90], [97], [151].
case of EM, to both detect the time of a change point and predict the These have achieved great results in classifying non-stationary time
next state (or concept) of financial data using a hidden Markov model series [24], [37], [38].

-7-
 International Journal of Interactive Multimedia and Artificial Intelligence

Recent EIS approaches can work as ensembles of rules [116] and solutions based on ensembles and evolving fuzzy systems. It is also
apply meta-cognitive scaffolding theory for tuning the learned model important to note how the relatively recent developments in deep
incrementally in what-to-learn, when-to-learn, and how-to-learn learning have fostered the popularity of approaches where artificial
[138]. In fact, ensembles are known for their good results in predicting neural networks play a key role.
both cyclic and non-stationary data such as stock prices [10], [13], Future research is likely to emphasise the application of data stream
[110]. These have also introduced the ability to deal with recurrent classification algorithms to financial streams. Online machine learning
concepts explicitly and have beaten other methods at predicting the has not been widely applied to the financial domain. However, as
S&P500 [37], [38], [111]. For instance, Pratama et al. employed an shown in this study, similar techniques like sequential and recurring
evolving type-2 recurrent fuzzy neural network to learn incrementally deep learning models are on the rise in finance. Applying the problem
and handle recurring drifts in both [124] and [38]. In any case, there is of concept drift to handling price change dynamics seems a natural
still a significant gap between EIS and the rest of the literature for data step forward on the research line of financial regime changes.
stream classification.
Having said that, better access to high-frequency data and
This line of work based on EIS is being complemented by other computational resources will also likely result in major progress in
studies that combine ensembles and other evolutionary algorithms to the near future.
tackle concept drift in financial applications like [84]. These authors
used ensembles of trading rules evolved using grammatical evolution
to manage structural change in the Standard & Poor’s 500 index. Acknowledgment
We would like to thank the editor and external reviewers for their
V. Conclusions and Future Work thoughtful and detailed comments on our paper. We would also like to
acknowledge the financial support of the Spanish Ministry of Science,
The application of AI to computational finance has been a very Innovation and Universities under grant PGC2018-096849-B-I00
active field of research for decades. Among the key difficulties (MCFin).
identified in the literature on financial prediction, we can mention
structural change. The price generation process of financial time
series is often affected by changes of different natures. Some of these References
changes can reoccur over time as seasonal patterns, while others do [1] Y. S. Abu-Mostafa, A. F. Atiya, “Introduction to financial forecasting,”
not repeat, being abrupt breaks in the non-stationary price dynamics. Applied Intelligence, vol. 6, pp. 205–213, 7 1996.
This study presented a systematic literature review of machine [2] W. Huang, Y. Nakamori, S.-Y. Wang, “Forecasting stock market movement
learning techniques for financial prediction under regime changes. A direction with support vector machine,” Computers & Operations Research,
variety of sources were inspected to perform an exhaustive search. vol. 32, pp. 2513–2522, 2005.
This review included: ScienceDirect, IEEE Digital Library, ACM [3] R. C. Cavalcante, R. C. Brasileiro, V. L. F. Souza, J. P. Nobrega, A. L. I.
Oliveira, “Computational Intelligence and Financial Markets: A Survey
Digital Library, Taylor & Francis, Wiley and SpringerLink. Out of a
and Future Directions,” Expert Systems with Applications, vol. 55, pp.
total of 140 relevant studies, these are distributed as follows: i) concept 194–211, 8 2016.
drift or online learning related (32.1%); ii) related to financial literacy [4] V. Dogra, S. Verma, Kavita, N. Z. Jhanjhi, U. Ghosh, D. N. Le, “A
for regime changes (15.7%), and iii) ML techniques applied to stock Comparative Analysis of Machine Learning Models for Banking News
forecasting (52.1%). Extraction by Multiclass Classification With Imbalanced Datasets of
The results of reviewed publications show that ML has proven to Financial News: Challenges and Solutions,” International Journal of
Interactive Multimedia and Artificial Intelligence, vol. 7, no. 3, pp. 35–52,
be a powerful tool to tackle the problem of financial prediction under
2022, doi: 10.9781/ijimai.2022.02.002.
concept drift, which we define as structural breaks, that can occur
[5] M. W. Hsu, S. Lessmann, M. C. Sung, T. Ma, J. E. Johnson, “Bridging the
at any frequency level. This includes solutions based on different divide in financial market forecasting: machine learners vs. financial
algorithms that adapt their prediction models to new circumstances economists,” Expert Systems with Applications, vol. 61, pp. 215–234, 2016.
either through new model generation or managing an archive [6] B. M. Henrique, V. A. Sobreiro, H. Kimura, “Literature review: Machine
of former successful models. In this regard, many meta-learning learning techniques applied to financial market prediction,” Expert
approaches in the ML literature rely on non-supervised algorithms Systems with Applications, vol. 124, pp. 226–251, 2019.
to try to identify the recurrence of a concept and retrieve previous [7] G. S. Atsalakis, K. P. Valavanis, “Surveying stock market forecasting
models or detect drifts. techniques - Part II: Soft computing methods,” Expert Systems with
Applications, vol. 36, pp. 5932-59-41, 4 2009.
In this context, the use of sequential DL models such as RNNs can [8] M.-Y. Chen, A. K. Sangaiah, T.-H. Chen, E. D. Lughofer, E. Egrioglu,
be insufficient to tackle abrupt changes since the previous market “Deep learning for financial engineering,” Computational Economics, pp.
dynamics are still in memory in the models impacting predictive 1–5, 2022.
accuracy. In contrast, the model learns the new regime [113], [115], [9] A. M. Ozbayoglu, M. U. Gudelek, O. B. Sezer, “Deep learning for financial
[121]. applications: A survey,” Applied Soft Computing, vol. 93, p. 106384, 2020,
doi: https://doi.org/10.1016/j.asoc.2020.106384.
Depending on the frequency involved, researchers suggest either
[10] M. Ballings, D. Van Den Poel, N. Hespeels, R. Gryp, “Evaluating multiple
solutions based on model retraining either at regular intervals or
classifiers for stock price direction prediction,” Expert Systems with
upon detection of changes or drifts or online incremental algorithms. Applications, vol. 42, no. 20, pp. 7046–7056, 2015.
This entails having up-to-date models with the use of forgetting [11] A. Booth, E. Gerding, F. McGroarty, “Automated trading with
mechanisms to avoid overfitting and adapting to new market performance weighted random forests and seasonality,” Expert Systems
behaviours. Regarding the latter, and despite the success of online with Applications, vol. 41, pp. 3651–3661, 6 2014.
ensembles dealing with complex systems and training base learners to [12] P. Ładyżyński, K. Żbikowski, P. Grzegorzewski, “Stock trading with
deal with different regimes, the use of these approaches from the data random forests, trend detection tests and force index volume indicators,”
stream learning literature is not as popular in financial forecasting yet. in Artificial Intelligence and Soft Computing: 12th International Conference,
ICAISC 2013, Proceedings, Part II, Berlin, Heidelberg, Jun. 2013, pp. 441–
Even though there doesn’t seem to be a clearly dominant ML 452, Springer Berlin Heidelberg.
technique in this space, it is worth mentioning the popularity of [13] J. Patel, S. Shah, P. Thakkar, K. Kotecha, “Predicting stock and stock price

-8-
 Article in Press

index movement using trend deterministic data preparation and machine [37] M. Pratama, E. Lughofer, J. Er, S. Anavatti, C.-P. Lim, “Data driven
learning techniques,” Expert Systems with Applications, vol. 42, no. 1, pp. modelling based on Recurrent Interval-Valued Metacognitive Scaffolding
259–268, 2015, doi: https://doi.org/10.1016/j.eswa.2014.07.040. Fuzzy Neural Network,” Neurocomputing, vol. 262, pp. 4–27, 2017.
[14] D. Ardia, K. Bluteau, M. Rüede, “Regime changes in Bitcoin GARCH [38] M. Pratama, J. Lu, E. Lughofer, G. Zhang, M. J. Er, “Incremental Learning
volatility dynamics,” Finance Research Letters, vol. 29, pp. 266–271, Jun. of Concept Drift Using Evolving Type-2 Recurrent Fuzzy Neural
2019, doi: 10.1016/J.FRL.2018.08.009. Network,” IEEE Transactions on Fuzzy Systems, pp. 1–1, 2016, doi: 10.1109/
[15] A. Ang, A. Timmermann, “Regime changes and financial markets,” TFUZZ.2016.2599855.
Annual Review of Financial Economics, vol. 4, no. 1, pp. 313–337, 2012. [39] C. Alippi, G. Boracchi, M. Roveri, “Just-in-time classifiers for recurrent
[16] A. Tsymbal, “The Problem of Concept Drift: Definitions and Related concepts,” IEEE Transactions on Neural Networks and Learning Systems,
Work,” Technical Report: TCD-CS-2004-15, Department of Computer vol. 24, pp. 620–634, 4 2013.
Science Trinity College, Dublin, 2004. [40] J. B. Gomes, M. M. Gaber, P. A. C. Sousa, E. Menasalvas, “Mining
[17] A. L. Suárez-Cetrulo, A. Cervantes, D. Quintana, “Incremental Market recurring concepts in a dynamic feature space,” IEEE Transactions on
Behavior Classification in Presence of Recurring Concepts,” Entropy, vol. Neural Networks and Learning Systems, vol. 25, pp. 95–110, 1 2014, doi:
21, p. 25, Jan. 2019, doi: 10.3390/e21010025. 10.1109/TNNLS.2013.2271915.
[18] M. C. Münnix, T. Shimada, R. Schäfer, F. Leyvraz, T. H. Seligman, T. Guhr, [41] P. M. Gonçalves Jr, R. Souto, M. De Barros, “RCD: A recurring concept
H. E. Stanley, “Identifying States of a Financial Market,” Scientific Reports, drift framework,” Pattern Recognition Letters, vol. 34, pp. 1018–1025, 2013.
vol. 2, p. 644, 12 2012. [42] A. W. Lo, H. Mamaysky, J. Wang, “Foundations of Technical Analysis:
[19] E. Tsang, J. Chen, Detecting regime change in computational finance: data Computational Algorithms, Statistical Inference, and Empirical
science, machine learning and algorithmic trading. CRC Press, 2020. Implementation,” The Journal of Finance, vol. 55, pp. 1705–1765, 8 2000.
[20] R. T. Das, K. K. Ang, C. Quek, “IeRSPOP: A novel incremental rough [43] F. E. Tay, L. Cao, “Application of support vector machines in financial
set-based pseudo outer-product with ensemble learning,” Applied Soft time series forecasting,” Omega, vol. 29, pp. 309–317, 8 2001.
Computing Journal, vol. 46, pp. 170–186, 9 2016. [44] T. Geva, J. Zahavi, “Empirical evaluation of an automated intraday stock
[21] V. Vella, W. L. Ng, “Enhancing risk-adjusted performance of stock market recommendation system incorporating both market data and textual
intraday trading with Neuro-Fuzzy systems,” Neurocomputing, vol. 141, news,” Decision Support Systems, vol. 57, pp. 212–223, Jan. 2014, doi:
pp. 170–187, 2014. 10.1016/J.DSS.2013.09.013.
[22] Y. Hu, K. Liu, X. Zhang, K. Xie, W. Chen, Y. Zeng, M. Liu, “Concept [45] C. H. Chen, P. Y. Chen, J. C. W. Lin, “An Ensemble Classifier for Stock
drift mining of portfolio selection factors in stock market,” Electronic Trend Prediction Using Sentence- Level Chinese News Sentiment and
Commerce Research and Applications, vol. 14, no. 6, pp. 444–455, 2015. Technical Indicators,” International Journal of Interactive Multimedia
[23] B. Silva, N. Marques, G. Panosso, “Applying neural networks for concept and Artificial Intelligence, vol. 7, no. 3, pp. 53–64, 2022, doi: 10.9781/
drift detection in financial markets,” in CEUR Workshop Proceedings, vol. ijimai.2022.02.004.
960, 2012, pp. 43–47. [46] B. Kitchenham, O. P. Brereton, D. Budgen, M. Turner, J. Bailey, S.
[24] X. Gu, P. P. Angelov, A. M. Ali, W. A. Gruver, G. Gaydadjiev, “Online Linkman, “Systematic literature reviews in software engineering–a
evolving fuzzy rule-based prediction model for high frequency trading systematic literature review,” Information and software technology, vol.
financial data stream,” in 2016 IEEE Conference on Evolving and Adaptive 51, no. 1, pp. 7– 15, 2009.
Intelligent Systems (EAIS), 5 2016, pp. 169–175, IEEE. [47] B. Kitchenham, “Procedures for performing systematic reviews,” Keele,
[25] E. F. Fama, “Efficient Capital Markets: A Review of Theory and UK, Keele University, vol. 33, no. 2004, pp. 1–26, 2004.
Empirical Work,” The Journal of Finance, vol. 25, no. 2, p. 383, 1970, doi: [48] M. J. Page, J. E. McKenzie, P. M. Bossuyt, I. Boutron, T. C. Hoffmann, C. D.
10.2307/2325486. Mulrow, L. Shamseer, J. M. Tetzlaff, E. A. Akl, S. E. Brennan, et al., “The
[26] J. Piger, Econometrics: Models of Regime Changes, pp. 2744–2757. New prisma 2020 statement: an updated guideline for reporting systematic
York, NY: Springer New York, 2009. reviews,” Systematic reviews, vol. 10, no. 1, pp. 1–11, 2021.
[27] A. G. Hoepner, D. McMillan, A. Vivian, Wese Simen, “Significance, [49] A. W. Lo, “Reconciling efficient markets with behavioral finance: the
relevance and explainability in the machine learning age: an econometrics adaptive markets hypothesis,” Journal of investment consulting, vol. 7, no.
and financial data science perspective,” The European Journal of Finance, 2, pp. 21–44, 2005.
vol. 27, no. 1-2, pp. 1–7, 2021. [50] S. Baek, S. K. Mohanty, M. Glambosky, “Covid-19 and stock market
[28] P. Bracke, A. Datta, C. Jung, S. Sen, “Machine learning explainability in volatility: An industry level analysis,” Finance Research Letters, vol. 37,
finance: an application to default risk analysis,” Bank of England, 2019. p. 101748, 2020.
[29] M. Kritzman, S. Page, D. Turkington, “Regime shifts: Implications for [51] M. L. De Prado, Advances in financial machine learning. John Wiley &
dynamic strategies (corrected),” Financial Analysts Journal, vol. 68, no. Sons, 2018.
3, pp. 22–39, 2012. [52] A. Ang, G. Bekaert, “How regimes affect asset allocation,” Financial
[30] R. Elwell, R. Polikar, “Incremental learning of concept drift in Analysts Journal, vol. 60, no. 2, pp. 86–99, 2004.
nonstationary environments.,” IEEE transactions on neural networks, vol. [53] G. Davies, “Regime changes in the financial markets,” 2016. [Online]. Available:
22, pp. 1517–31, 10 2011. https://www.ft.com/content/6556ec60-6aa3-3dfe- 8953-b94d6080c360.
[31] J. A. Gama, I. Žliobaitundefined, A. Bifet, M. Pechenizkiy, A. Bouchachia, [54] E. Andreou, E. Ghysels, Structural Breaks in Financial Time Series, pp.
“A survey on concept drift adaptation,” ACM Computing Surveys, vol. 46, 839–870. Berlin, Heidelberg: Springer Berlin Heidelberg, 2009.
Mar. 2014, doi: 10.1145/2523813. [55] A. Samitas, E. Kampouris, D. Kenourgios, “Machine learning as an
[32] G. Ditzler, M. Roveri, C. Alippi, R. Polikar, “Learning in nonstationary early warning system to predict financial crisis,” International Review of
environments: A survey,” IEEE Computational Intelligence Magazine, vol. Financial Analysis, vol. 71, p. 101507, 2020.
10, pp. 12–25, Nov. 2015. [56] D. Pettenuzzo, A. Timmermann, “Predictability of stock returns and asset
[33] G. I. Webb, R. Hyde, H. Cao, H. L. Nguyen, F. Petitjean, “Characterizing allocation under structural breaks,” Journal of Econometrics, vol. 164, pp.
concept drift,” Data Mining and Knowledge Discovery, vol. 30, pp. 964– 60–78, Sep. 2011, doi: 10.1016/j.jeconom.2011.02.019.
994, 7 2016. [57] A. W. Lo, “The Adaptive Markets Hypothesis: Market Efficiency from an
[34] H. M. Gomes, J. P. Barddal, F. Enembreck, A. Bifet, “A Survey on Ensemble Evolutionary Perspective,” MIT, 2004.
Learning for Data Stream Classification,” ACMComputing Surveys, vol. [58] J. D. Hamilton, “A new approach to the economic analysis of
50, no. 2, pp. 1–36, 2017, doi: 10.1145/3054925. nonstationary time series and the business cycle,” Econometrica: Journal
[35] S. Ramírez-Gallego, B. Krawczyk, S. García, M. Woźniak, F. Herrera, “A of the econometric society, pp. 357–384, 1989.
survey on data preprocessing for data stream mining: Current status and [59] J. D. Hamilton, “Regime switching models,” in Macroeconometrics and
future directions,” Neurocomputing, vol. 239, pp. 39–57, 2017. time series analysis, Springer, 2010, pp. 202–209.
[36] A. R. Masegosa, A. M. Martínez, D. Ramos-López, H. Langseth, T. D. [60] T. Preis, J. J. Schneider, H. E. Stanley, “Switching processes in financial
Nielsen, A. Salmerón, “Analyzing concept drift: A case study in the markets,” Proceedings of the National Academy of Sciences, vol. 108, no. 19,
financial sector,” Intelligent Data Analysis, vol. 24, no. 3, pp. 665–688, 2020. pp. 7674–7678, 2011, doi: 10.1073/pnas.1019484108/-/DCSupplemental.

-9-
 International Journal of Interactive Multimedia and Artificial Intelligence

[61] R. Hammerschmid, H. Lohre, “Regime shifts and stock return for algorithmic trading,” Applied Soft Computing, vol. 84, p. 105713, 2019,
predictability,” International Review of Economics and Finance, vol. 56, pp. doi: https://doi.org/10.1016/j.asoc.2019.105713.
138–160, Jul. 2018, doi: 10.1016/j.iref.2017.10.021. [85] C. Chen, W. Dongxing, H. Chunyan, Y. Xiaojie, “Exploiting social media
[62] Q. Dai, K. J. Singleton, W. Yang, “Is regime-shift risk priced in the us for stock market prediction with factorization machine,” in 2014 IEEE/
treasury market?,” Working paper, New York University and Stanford WIC/ACM International Joint Conferences on Web Intelligence (WI) and
University, 2003. Intelligent Agent Technologies (IAT), vol. 2, 2014, pp. 142–149, IEEE.
[63] J. D. Hamilton, “Macroeconomic regimes and regime shifts,” Handbook of [86] M. R. Hassan, K. Ramamohanarao, J. Kamruzzaman, M. Rahman, M.
macroeconomics, vol. 2, pp. 163– 201, 2016. M. Hossain, “A hmm-based adaptive fuzzy inference system for stock
[64] J. G. Dias, J. K. Vermunt, S. Ramos, “Clustering financial time series: New market forecasting,” Neurocomputing, vol. 104, pp. 10–25, 2013.
insights from an extended hidden markov model,” European Journal of [87] X. Zhang, Y. Li, S. Wang, B. Fang, S. Y. Philip, “Enhancing stock market
Operational Research, vol. 243, no. 3, pp. 852–864, 2015. prediction with extended coupled hidden Markov model over multi-
[65] K.-S. Kim, I. Han, “The cluster-indexing method for case-based reasoning sourced data,” Knowledge and Information Systems, vol. 61, no. 2, pp.
using self-organizing maps and learning vector quantization for bond 1071–1090, 2019.
rating cases,” Expert systems with applications, vol. 21, no. 3, pp. 147– 156, [88] N. Gârleanu, L. H. Pedersen, “Efficiently inefficient markets for assets
2001. and asset management,” The Journal of Finance, vol. 73, no. 4, pp. 1663–
[66] R. Bisoi, P. K. Dash, “A hybrid evolutionary dynamic neural network 1712, 2018.
for stock market trend analysis and prediction using unscented kalman [89] G. C. Friesen, P. A. Weller, L. M. Dunham, “Price trends and patterns
filter,” Applied Soft Computing, vol. 19, pp. 41–56, 2014. in technical analysis: A theoretical and empirical examination,” Journal
[67] M. Guidolin, A. Timmermann, “An econometric model of nonlinear of Banking and Finance, vol. 33, pp. 1089–1100, Jun. 2009, doi: 10.1016/j.
dynamics in the joint distribution of stock and bond returns,” Journal jbankfin.2008.12.010.
of Applied Econometrics, vol. 21, pp. 1–22, Jan. 2006, doi: 10.1002/jae.824. [90] C. F. Liu, C. Y. Yeh, S. J. Lee, “Application of type- 2 neuro-fuzzy modeling
[68] A. Urquhart, F. McGroarty, “Are stock markets really efficient? in stock price prediction,” Applied Soft Computing Journal, vol. 12, no. 4,
evidence of the adaptive market hypothesis,” International Review of pp. 1348– 1358, 2012, doi: 10.1016/j.asoc.2011.11.028.
Financial Analysis, vol. 47, pp. 39–49, 2016, doi: https://doi.org/10.1016/j. [91] R. C. Cavalcante, A. L. Oliveira, “An autonomous trader agent for the
irfa.2016.06.011. stock market based on online sequential extreme learning machine
[69] N. G. Pavlidis, V. P. Plagianakos, D. K. Tasoulis, M. N. Vrahatis, “Financial ensemble,” in 2014 International Joint Conference on Neural Networks
forecasting through unsupervised clustering and neural networks,” (IJCNN), 2014, pp. 1424–1431, IEEE.
Operational Research, vol. 6, no. 2, pp. 103–127, 2006. [92] R. Garcia, R. Luger, E. Renault, “Empirical assessment of an intertemporal
[70] A. A. Ariyo, A. O. Adewumi, C. K. Ayo, “Stock price prediction using the option pricing model with latent variables,” Journal of Econometrics, vol.
ARIMA model,” in 2014 UKSim-AMSS 16th International Conference on 116, no. 1-2, pp. 49–83, 2003.
Computer Modelling and Simulation, 2014, pp. 106–112, IEEE. [93] R. Dash, P. K. Dash, R. Bisoi, “A self adaptive differential harmony search
[71] S. Choudhury, S. Ghosh, A. Bhattacharya, K. J. Fernandes, M. K. Tiwari, based optimized extreme learning machine for financial time series
“A real time clustering and SVM based price-volatility prediction for prediction,” Swarm and Evolutionary Computation, vol. 19, pp. 25– 42,
optimal trading strategy,” Neurocomputing, vol. 131, pp. 419– 426, 5 2014. 2014.
[72] S.-H. Park, J.-H. Lee, J.-W. Song, T.-S. Park, “Forecasting change directions [94] A. A. Adebiyi, C. K. Ayo, M. O. Adebiyi, S. O. Otokiti, “Stock price
for financial time series using hidden Markov model,” in International prediction using neural network with hybridized market indicators,”
Conference on Rough Sets and Knowledge Technology, 2009, pp. 184–191, Journal of Emerging Trends in Computing and Information Sciences, vol.
Springer. 3, no. 1, pp. 1–9, 2012.
[73] J. Sirignano, R. Cont, “Universal features of price formation in financial [95] R. Tsaih, Y. Hsu, C. C. Lai, “Forecasting S&P 500 stock index futures with
markets: perspectives from deep learning,” Quantitative Finance, 2019, a hybrid AI system,” Decision Support Systems, vol. 23, pp. 161–174, 6 1998.
doi: 10.1080/14697688.2019.1622295. [96] J. C. Neves, A. Vieira, “Improving bankruptcy prediction with hidden
[74] G. Deboeck, Trading on the edge: neural, genetic, and fuzzy systems for layer learning vector quantization,” European Accounting Review, vol. 15,
chaotic financial markets. Wiley, 1994. no. 2, pp. 253–271, 2006.
[75] S. Schulmeister, “Profitability of technical stock trading: Has it moved [97] Y. L. Yong, Y. Lee, X. Gu, P. P. Angelov, D. C. L. Ngo, E. Shafipour, “Foreign
from daily to intraday data?,” Review of Financial Economics, vol. 18, no. 4, currency exchange rate prediction using neuro-fuzzy systems,” Procedia
pp. 190– 201, 2009, doi: 10.1016/j.rfe.2008.10.001. computer science, vol. 144, pp. 232–238, 2018.
[76] T. Shintate, L. Pichl, “Trend Prediction Classification for High Frequency [98] H. Ghoddusi, G. G. Creamer, N. Rafizadeh, “Machine learning in energy
Bitcoin Time Series with Deep Learning,” Journal of Risk and Financial economics and finance: A review,” Energy Economics, vol. 81, pp. 709–727,
Management, vol. 12, p. 17, Jan. 2019, doi: 10.3390/jrfm12010017. 2019, doi: https://doi.org/10.1016/j.eneco.2019.05.006.
[77] B. G. Malkiel, “The efficient market hypothesis and its critics,” Journal of [99] F. Rundo, F. Trenta, A. L. di Stallo, S. Battiato, “Machine learning for
economic perspectives, vol. 17, no. 1, pp. 59–82, 2003. quantitative finance applications: A survey,” Applied Sciences, vol. 9, no.
[78] A. W. Lo, A. C. MacKinlay, A Non-Random Walk Down Wall Street. 24, p. 5574, 2019.
Princeton University Press, Dec. 2011. [100] Z. Lin, “Modelling and forecasting the stock market volatility of SSE
[79] X. Ding, Y. Zhang, T. Liu, J. Duan, “Using structured events to predict Composite Index using GARCH models,” Future Generation Computer
stock price movement: An empirical investigation,” in Proceedings of the Systems, vol. 79, pp. 960–972, 2018.
2014 Conference on Empirical Methods in Natural Language Processing [101] M. Frömmel, K. Lampaert, “Does frequency matter for intraday technical
(EMNLP), 2014, pp. 1415–1425. trading?,” Finance Research Letters, vol. 18, pp. 177–183, 2016.
[80] F. Black, “Noise,” The Journal of Finance, vol. 41, 529–543, Jul. 1986, doi: [102] M. Schreyer, T. Sattarov, A. Gierbl, B. Reimer, Borth, “Learning sampling in
10.1111/j.1540- 6261.1986.tb04513.x. financial statement audits using vector quantised variational autoencoder
[81] B. Malkiel, A Random Walk Down Wall Street. WW Norton & Company, neural networks,” in Proceedings of the First ACM International Conference
1973. on AI in Finance, 2020, pp. 1– 8.
[82] B. Vanstone, G. Finnie, “An empirical methodology for developing [103] R. I. Webb, D. D. Ryu, D. D. Ryu, J. Han, “The price impact of futures
stockmarket trading systems using artificial neural networks,” Expert trades and their intraday seasonality,” Emerging Markets Review, vol. 26,
Systems with Applications, vol. 36, no. 3, Part 2, pp. 6668–6680, 2009, doi: pp. 80– 98, 2016, doi: 10.1016/j.ememar.2016.01.002.
https://doi.org/10.1016/j.eswa.2008.08.019. [104] P. Bacchetta, E. Mertens, E. Van Wincoop, “Predictability in financial
[83] S. Carta, A. Corriga, A. Ferreira, A. S. Podda, D. R. Recupero, “A multi- markets: What do survey expectations tell us?,” Journal of International
layer and multi-ensemble stock trader using deep learning and deep Money and Finance, vol. 28, no. 3, pp. 406–426, 2009.
reinforcement learning,” Applied Intelligence, vol. 51, no. 2, pp. 889–905, [105] L. Ryll, S. Seidens, “Evaluating the performance of machine learning
2021. algorithms in financial market forecasting: A comprehensive survey,”
[84] C. Martín, D. Quintana, P. Isasi, “Grammatical evolution-based ensembles arXiv preprint arXiv:1906.07786, 2019.

- 10 -
 Article in Press

[106] A. M. Rather, A. Agarwal, V. Sastry, “Recurrent neural network and [129] N. Kasabov, D. Filev, “Evolving intelligent systems: Methods, learning,
a hybrid model for prediction of stock returns,” Expert Systems with applications,” in 2006 International Symposium on Evolving Fuzzy Systems,
Applications, vol. 42, no. 6, pp. 3234–3241, 2015. Sept 2006, pp. 8– 18.
[107] J. P. Serbera, P. Paumard, “The fall of high-frequency trading: A survey of [130] S.-s. Zhang, J.-w. Liu, X. Zuo, “Adaptive online incremental learning for
competition and profits,” Research in International Business and Finance, evolving data streams,” Applied Soft Computing, vol. 105, p. 107255, 2021.
vol. 36, pp. 271–287, 2016, doi: 10.1016/j.ribaf.2015.09.021. [131] M. Mermillod, A. Bugaiska, P. Bonin, “The stability- plasticity dilemma:
[108] S. Alonso-Monsalve, A. L. Suárez-Cetrulo, A. Cervantes, D. Quintana, Investigating the continuum from catastrophic forgetting to age-limited
“Convolution on neural networks for high-frequency trend prediction learning effects,” Frontiers in psychology, vol. 4, p. 504, 2013.
of cryptocurrency exchange rates using technical indicators,” Expert [132] E. Lughofer, Evolving fuzzy systems-methodologies, advanced concepts and
Systems with Applications, 1 2020. applications, vol. 53. Springer, 2011.
[109] A. W. Li, G. S. Bastos, “Stock market forecasting using deep learning and [133] M. Carnein, H. Trautmann, “Optimizing data stream representation: An
technical analysis: a systematic review,” IEEE Access, vol. 8, pp. 185232– extensive survey on stream clustering algorithms,” Business & Information
185242, 2020. Systems Engineering, vol. 61, no. 3, pp. 277–297, 2019.
[110] J. Patel, S. Shah, P. Thakkar, K. Kotecha, “Predicting stock market index [134] A. Bifet, R. Gavaldà, G. Holmes, B. Pfahringer, Machine learning for data
using fusion of machine learning techniques,” Expert Systems with streams: with practical examples in MOA. MIT press, 2018.
Applications, vol. 42, pp. 2162–2172, 3 2015. [135] K. Namitha, G. Santhosh Kumar, “Learning in the presence of concept
[111] M. Pratama, S. G. Anavatti, P. P. Angelov, E. Lughofer, “PANFIS: A novel recurrence in data stream clustering,” Journal of Big Data, vol. 7, p. 75,
incremental learning machine,” IEEE Transactions on Neural Networks Dec. 2020, doi: 10.1186/s40537-020-00354-1.
and Learning Systems, vol. 25, pp. 55–68, 1 2014. [136] J. Gama, R. Sebastiao, P. P. Rodrigues, “On evaluating stream learning
[112] G. Widmer, M. Kubat, “Learning in the presence of concept drift and algorithms,” Machine learning, vol. 90, no. 3, pp. 317–346, 2013.
hidden contexts,” Machine Learning, vol. 23, no. 1, pp. 69–101, 1996, doi: [137] L. Yang, S. McClean, M. Donnelly, K. Burke, K. Khan, “Detecting and
10.1007/BF00116900. responding to concept drift in business processes,” Algorithms, vol. 15,
[113] D. Sahoo, Q. Pham, J. Lu, S. C. Hoi, “Online deep learning: Learning no. 5, 2022, doi: 10.3390/a15050174.
deep neural networks on the fly,” in Proceedings of the Twenty-Seventh [138] G. Sateesh Babu, S. Suresh, G.-B. Huang, “Meta- cognitive Neural
International Joint Conference on Artificial Intelligence IJCAI 2018, Jul. Network for classification problems in a sequential learning framework,”
2018, pp. 2660–2666. Neurocomputing, vol. 81, pp. 86–96, 2011.
[114] A. L. D. Rossi, B. F. De Souza, C. Soares, A. de Leon Ferreira de Carvalho, [139] A. Maslov, M. Pechenizkiy, I. Žliobaite˙, T. Kärkkäinen, “Modelling
C. Ponce, “A guidance of data stream characterization for meta-learning,” recurrent events for improving online change detection,” in Proceedings
Intelligent Data Analysis, vol. 21, no. 4, pp. 1015–1035, 2017. of the 2016 SIAM International Conference on Data Mining, 2016, pp. 549–
[115] C. Käding, E. Rodner, A. Freytag, J. Denzler, “Fine- tuning deep neural 557, SIAM.
networks in continuous learning scenarios,” in Computer Vision – [140] P. P. Angelov, D. P. Filev, “An approach to online identification of takagi-
ACCV 2016 Workshops, Cham, 2017, pp. 588–605, Springer International sugeno fuzzy models,” Trans. Sys. Man Cyber. Part B, vol. 34, pp. 484–498,
Publishing. Feb. 2004, doi: 10.1109/TSMCB.2003.817053.
[116] P. P. Angelov, Empirical Approach to Machine Learning. Springer, 2017. [141] C. W. Chiu, L. L. Minku, “A diversity framework for dealing with multiple
[117] B. Krawczyk, L. L. Minku, J. Gama, J. Stefanowski, M. Woźniak, M. types of concept drift based on clustering in the model space,” IEEE
Wó Zniak, “Ensemble learning for data stream analysis: A survey,” Transactions on Neural Networks and Learning Systems, 2020.
Information Fusion, vol. 37, pp. 132–156, 2017. [142] A. L. Suárez-Cetrulo, A. Cervantes, “An online classification algorithm
[118] R. D. Baruah, P. Angelov, “Evolving fuzzy systems for data streams: A for large scale data streams: igngsvm,” Neurocomputing, vol. 262, pp. 67–
survey,” Wiley Interdisciplinary Reviews: Data Mining and Knowledge 76, 2017, doi: https://doi.org/10.1016/j.neucom.2016.12.093.
Discovery, vol. 1, pp. 461–476, 11 2011. [143] E. Lughofer, C. Cernuda, S. Kindermann, M. Pratama, “Generalized smart
[119] M. Pratama, S. G. Anavatti, M. Joo, E. D. Lughofer, “pclass: An effective evolving fuzzy systems,” Evolving Systems, vol. 6, no. 4, pp. 269–292, 2015,
classifier for streaming examples,” IEEE Transactions on Fuzzy Systems, doi: 10.1007/s12530-015-9132-6.
vol. 23, pp. 369–386, Apr. 2015, doi: 10.1109/TFUZZ.2014.2312983. [144] P. E. Tsinaslanidis, D. Kugiumtzis, “A prediction scheme using
[120] M. Á. Abad, J. B. Gomes, E. Menasalvas, “Predicting recurring concepts perceptually important points and dynamic time warping,” Expert
on data-streams by means of a meta-model and a fuzzy similarity Systems with Applications, vol. 41, no. 15, pp. 6848–6860, 2014, doi:
function,” Expert Systems With Applications, vol. 46, pp. 87–105, 2015, doi: https://doi.org/10.1016/j.eswa.2014.04.028.
10.1016/j.eswa.2015.10.022. [145] C. Martin, D. Quintana, P. Isasi, “Dynamic generation of investment
[121] J. Read, “Concept-drifting data streams are time series; the case for recommendations using grammatical evolution.,” International Journal of
continuous adaptation,” arXiv preprint arXiv:1810.02266, 2018. Interactive Multimedia & Artificial Intelligence, vol. 6, no. 6, 2021.
[122] H. M. Gomes, J. Read, A. Bifet, J. P. Barddal, J. Gama, “Machine learning [146] B. Grün, “Model-Based Clustering,” Handbook of Mixture Analysis, pp.
for streaming data: state of the art, challenges, and opportunities,” 157–192, Feb. 2019, doi: 10.1201/9780429055911-8.
ACM SIGKDD Explorations Newsletter, vol. 21, no. 2, pp. 6–22, 2019, doi: [147] P. D. McNicholas, “Model-based clustering,” Journal of Classification, vol.
10.1145/3373464.3373470. 33, no. 3, pp. 331–373, 2016.
[123] X. Zheng, P. Li, X. Hu, K. Yu, “Semi-supervised classification on data [148] F. Dellaert, “The expectation maximization algorithm,” Georgia Institute
streams with recurring concept drift and concept evolution,” Knowledge- of Technology, 2002.
Based Systems, vol. 215, p. 106749, 2021. [149] J. Vanschoren, J. N. Van Rijn, B. Bischl, L. Torgo, “Openml: networked
[124] M. Pratama, J. Lu, E. Lughofer, G. Zhang, S. Anavatti, “Scaffolding type-2 science in machine learning,” ACM SIGKDD Explorations Newsletter, vol.
classifier for incremental learning under concept drifts,” Neurocomputing, 15, no. 2, pp. 49–60, 2014.
vol. 191, pp. 304–329, 2016. [150] P. P. Angelov, X. Zhou, “Evolving fuzzy-rule-based classifiers from data
[125] E. Lughofer, P. Angelov, “Handling drifts and shifts in on-line data streams,” IEEE Transactions on Fuzzy Systems, vol. 16, pp. 1462–1475, Dec.
streams with evolving fuzzy systems,” Appl. Soft Comput., vol. 11, pp. 2008, doi: 10.1109/TFUZZ.2008.925904.
2057–2068, Mar. 2011, doi: 10.1016/j.asoc.2010.07.003. [151] O. Cordón, F. Gomide, F. Herrera, F. Hoffmann, L. Magdalena, “Ten years
[126] Ł. Korycki, B. Krawczyk, “Streaming decision trees for lifelong learning,” of genetic fuzzy systems: Current framework and new trends,” in Fuzzy
in Joint European Conference on Machine Learning and Knowledge Sets and Systems, vol. 141, 2004, pp. 5–31.
Discovery in Databases, 2021, pp. 502–518, Springer.
[127] S. U. Din, J. Shao, “Exploiting evolving micro-clusters for data stream
classification with emerging class detection,” Information Sciences, vol.
507, pp. 404–420, 2020.
[128] R. Szadkowski, J. Drchal, J. Faigl, “Continually trained life-long
classification,” Neural Computing and Applications, pp. 1–18, 2021.

- 11 -
 International Journal of Interactive Multimedia and Artificial Intelligence

Andrés L. Suárez-Cetrulo
Andrés received his BSc and MSc in Computer Science at
Carlos III of Madrid (Spain) in 2013 and 2014, respectively.
He received his PhD in Computer Science at University
Carlos III of Madrid in 2022. He is currently a Data Science
Architect at Ireland’s National Centre for Applied AI, based
at University College Dublin. His current interests focus on
online machine learning for data streams, regime changes
in financial markets, deep learning, transformers and generative models.

David Quintana
David Quintana holds Bachelor’s degrees in Business
Administration and Computer Science. He has an M.S.
in Intelligent Systems from Universidad Carlos III de
Madrid and a PhD in Finance from Universidad Pontificia
Comillas (ICADE). He is currently Associate Professor at
the Computer Science Department at University Carlos III
of Madrid. There, he is part of the bio-inspired algorithms
group EVANNAI. His current research interests are mainly focused on
applications of Computational Intelligence in finance and economics.

Alejandro Cervantes
Graduated as Telecommunications Engineer at Universidad
Politecnica of Madrid (Spain), in 1993. He received his
PhD in Computer Science at University Carlos III of
Madrid in 2007. He is currently a professor at the Escuela
Superior de Ingeniería y Tecnología in UNIR (Universidad
Internacional de la Rioja). His interests include bio-inspired
algorithms for classification of non-stationary data, large
multi-objective optimization problems, swarm intelligence algorithms and deep
machine learning for meteorological forecasting, aeronautics and astrophysics.

- 12 -