time series momentum trading strategy and autocorrelation amplification

PLoS One. 2022; 15(3): e0230124.

Stock price prediction using principal components

Mahsa Ghorbani, Conceptualization, Data curation, Formal analysis, Support acquisition, Probe, Methodological analysis, Project administration, Resources, Software system, Validation, Visualization, Committal to writing – primary draft, Writing – review danamp; redaction ^1, ^* and Edwin K. P. Chong, Conceptualization, Formal analysis, Methodology, Supervision ²

Mahsa Ghorbani

¹ PhD Student, Department of Systems Engineering, CO State University, Fort Collins, Colorado, United States of America of America,

Edwin K. P. Chong

² Professor, Department of Electrical and Computing device Engineering, CO State University, Fort Collins, Colorado, United States of America,

Stefan Cristian Gherghina, Editor

Standard 2022 Nov 6; Accepted 2022 Feb 21.

Abstract

The literature provides strong certify that stock toll values can be predicted from past price data. Principal part analysis (PCA) identifies a young number of principle components that explicate nearly of the fluctuation in a data set. This method acting is often used for dimensionality reduction and analysis of the data. In that paper, we develop a general method for stock price prediction using time-varying covariance information. To address the time-varying nature of financial time series, we assign exponential weights to the price data so that recent information points are weighted more intemperately. Our proposed method involves a property-reduction surgery constructed based happening principle components. Projecting the noisy observation onto a rule subspace results in a well-conditioned problem. We exemplify our results supported real day by day price data for 150 companies from different market-capitalization categories. We compare the performance of our method to cardinal other methods: Karl Friedrich Gauss-Bayes, which is numerically demanding, and moving norm, a simple method often used aside technical traders and researchers. We investigate the results based on mean squared misplay and directing change statistic of prediction, as measures of performance, and volatility of prediction as a measure of risk.

Introduction

Predicting future timeworn price values is a very provocative task. There is a big body of literature on different methods and different predictors to incorporate into those methods to predict the future values as nearly as possible. The lit provides tough tell apart that chivalric price/return data derriere be used to predict succeeding stock prices. Some studied have found significant auto-correlation coefficient for returns over a short period. Gallic and Roll find disinclined correlation for man-to-man securities for daily returns [1]. Other studies show there is a sensationalism correlation for returns all over the period of weeks or months [2]. Studies likewise exhibit stock return correlation over the period of multiple months operating theater years. Fama and French report that the car-coefficient of correlation is stronger for longer periods, trio to five years, compared to regular or time period periods [3]. Cutler et Heart of Dixie. report positive auto-correlation over the horizon of several months and negative machine-correlation all over the horizon of three to basketball team years [4]. Thither are some other studies that also bear witness correlation in stock returns over a multiple year interval [5, 6] which all corroborate that price/return values are predictable from then price/return values.

Bogousslavsky shows that trading by investors with heterogeneous rebalancing horizons butt produce to autocorrelation in the returns at varied frequencies [7]. Chowdhury et al. investigate the autocorrelation construction of seven Gulf Cooperation Council (GCC) stock markets. Completely the markets except for Dubai and Kuwait show significant first-order autocorrelation of returns. They likewise find that autocorrelation between weekdays is usually larger than that between the above all trading years of the week [8]. Li et al. study the nonlinear autoregressive dynamics of stock index returns in vii major advanced economies (G7) and China using the quantile autoregression model. For the stock markets in the seven developed economies, the autoregressive parameters broadly follow a decreasing rule across the quantiles with significant portions outdoorsy the ordinary least squares estimate intervals [9]. Another study investigates the autocorrelation social organization of stock and portfolio returns in the unique securities industry setting of Saudi Arabia [10]. Their results indicate that in that location is significantly empiricist philosophy autocorrelation in individual stock and food market returns. Some other study applies the threshold quantile autoregressive model to study stock return autocorrelations in the Chinese stock market [11]. They reputation negative autocorrelations in the lour regimen and positive autocorrelations in the higher regime.

Other fundamental or macroeconomic factors can besides equal used in predicting future stock price values. Macroeconomic factors so much as interest rates, expected inflation, and dividend can be used in stock devolve predictions models [3, 12]. Besides fundamental variables so much as earnings fruit, cash in flow yield, size and al-Qur'an to market equity [13, 14] receive been found to have estimation power in predicting future price/return values.

Silvennoinen and Teräsvirta report correlation between individual U.S. stocks and the totality U.S. grocery [15]. Dennis et alibi. study the dynamic coitus between time unit stock returns and daily volatility innovations, and they study negative correlations [16]. Another study investigates the effectuate of common factors on the relationship among stocks and connected the distribution of the investment weights for stocks [17]. They report that commercialise plays a possessive role in both structuring the relationship among stocks and in constructing a advisable-wide-ranging portfolio. Dimic et alii. examine the wallop of international financial market incertitude and domestic macroeconomic factors connected stock–bond correlation in emerging markets [18]. In another study, the focussing is analyzing the impact of oil Mary Leontyne Pric shocks connected the interactions of oil-stock prices [19]. The results show that negative changes in oil prices have a significant impact on the shopworn market.

Therein paper, we describe a general method acting for predicting future bloodline price values settled along historical price data, victimisation time-varying covariance information. When the number of observations is large compared to the number of predictors, the maximum-likelihood covariance estimate [20] or straight-grained the a posteriori covariance is a good judge of the covariance of the data, but that is non e'er the case. When the number of observations is smaller than the matrix attribute, the problem is smooth worse because the matrix is not positive definite [21]. This problem, which happens quite an often in finance, gives rise to a new class of estimators such as shrinkage estimators. For example Ledoit and Brute, shrink the sample covariance towards a armoured unit matrix using a shrinkage coefficient that minimizes the hateful squared erroneous belief of the prediction [22]. Some other studies in this field let in [23–25]. In our numerical evaluations therein paper we have decent empirical data to reliably track the covariance matrix over time.

Impulse-settled forecasting relies on prices favourable a trend, either upward or downwards. Supported the assumption that trends like this exist and can be exploited, momentum is used Eastern Samoa a heuristic program for forecasting and is probably the most favourite technical index used by traders; particularly, the method of Focal point Bm Index (DMI), due to Billy Wilder [26]. This kind of heuristic is a special pillow slip of pattern-based forecasting, where, in the case of momentum, the pattern is simply the upward or downward trend. Our method is a systematic method to capture capricious patterns, non just upward OR downward trends. So, we figure out prevalent patterns in the form of eigenvectors (or "eigen-patterns") of the local covariance intercellular substance. As such, we are able to exploit Thomas More imprecise patterns that are prevalent (but non necessary known in advance) in price time series.

The mean squared erroneousness (MSE) measures the distance between predicted and real values and is a precise common metric unit to evaluate the performance of predictive methods [27]. Variable conditional mean minimizes the mean squared error [28] and is a good estimator for future price values. However, numerical results using this method cannot always be trusted because of associated feverous-conditioning issues. In this paper we introduce a method acting with siamese estimation efficiency that does non stick out from this issue.

Principal constituent depth psychology (PCA), which is a method for dimensionality reduction of the data, is misused in different fields such as applied math variables analysis [29], pattern recognition, feature extraction, data compression, and visualisation of graduate dimensional data [30]. It also has individual application in exploring financial time serial publication [31], dynamic trading strategies [32], financial risk computations [32, 33], and statistical arbitrage [34]. In this work, we implement PCA in estimating future stock monetary value values.

Yu et al. introduce a machine-learning method to construct a stock-selection model, which can execute nonlinear classification of stocks. They use PCA to extract the low-dimensional and efficient information [35]. In some other subject, three mature dimensionality reducing techniques, PCA, fuzzy robust principal constituent analysis, and kernel-settled PCA, are applied to the whole data set to simplify and rearrange the original data structure [36]. Wang et al. immediate a stochastic function based on PCA developed for fiscal time-series prediction [37]. In another survey, PCA is applied to three subgroups of stocks of the Cut down Jones Developed (DJI) forefinger to optimise portfolios [38]. Narayan et al. apply PCA to test for predictability of excess stock returns for 18 emerging markets victimization a range of economics and institutional factors [39].

Factor psychoanalysis is a technique to identify the variability of observed data through few factors and is in some sense similar to PCA. In that respect is a long debate in the literature along which method is greatest [40, 41]. Factor analysis begins with the assumption that the data comes from a specific model where underlying factors satisfy reliable assumptions [42]. If the first exemplar formulation is not done properly, then the method will not perform well. PCA on the other handwriting involves no assumption on the form of the covariance matrix. In this paper, we focus on developing an algorithm that can ultimately exist used in different fields without prior knowledge of the system, and therefore PCA is the method acting of choice. In the case study presented in the following section, although only price data is used, it would have been also affirmable to include multiple predictors to estimate futures values of sprout prices.

Our method acting bears some similarity with subspace filtering methods. Such methods assume a low-down-rank model for the data [43]. The noisy data is decomposed onto a signal subspace and noise based happening a modified singular value decay (SVD) of data matrices [44]. The irrelevant decomposition can be finished by an SVD of the noisy notice matrix or equivalently by an eigenvalue of a matrix decomposition of the noisy signal covariance matrix [43].

We compare the public presentation of our proposed methods in footing of MSE and directional modify statistic. Stock-price direction prediction is an important result in the financial earthly concern. Justified small improvements in predictive performance tin be very profitable [45]. Directional change statistic calculates whether our method acting can predict the correct focus of change in price values [46]. It is an important rating measure of the performance because predicting the direction of toll movement is very important in some grocery store strategies.

Another important parameter that we are interested in is standard deviation, one of the distinguish fundamental risk measures in portfolio management [47]. The authoritative deviation is a statistical measure of volatility, often used by investors to measure the risk of a stock operating theatre portfolio.

As mentioned above, in this paper we cente forecasting stock prices from daily historical terms data. In Section, we introduce our technical methodology, and in peculiar estimation techniques using covariance information. In Section, we distinguish our method for processing the data and estimating the time-varying covariance intercellular substance from existential data, including data normalisatio. We also demonstrate the performance of our method acting.

Theoretical methodological analysis

Estimation techniques

In this section we introduce a unaccustomed computationally appealing method for estimating future stock price values using covariance information. The empirical covariance can personify used as an estimate of the covariance matrix if adequate empiric data is available, operating theater we can use techniques like to the ones introduced in the previous section, though the time-varying nature of the covariance must make up addressed.

Suppose that we are given the stock price values for M days. Our finish is to betoken company stock prices for M + 1 to N trading days, using the observed values of the previous consecutive M days. The reason for introducing N will make up clean below.

Gauss-Bayes or conditional estimation of z given y

Guess that x is a random vector of length N. Let M ≤ N and suppose that the first M data points of vector x represent the end-of-daytime prices of a company stock complete the past M consecutive trading days. The variable hit-or-miss vector x and put up be partitioned in the anatomy

Let random vector y represent the first M information points and z the price of the next N − M days in the early. We wish to estimate z from y.

The covariance matrix for the unselected vector x can be written A

where Σ_yy is the covariance of y and Σ_zz is the covariance of z. Assuming that y and z are jointly normally distributed, deliberate the prior distribution of x = [y, z], the Bayesian keister distribution of z given y is given away

$\begin{matrix} {\hat{z}}_{z ∣ y} & = Σ_{z y} Σ_{y y}^{- 1} y \\ {\hat{Σ}}_{z ∣ y} & = Σ_{z z} - Σ_{z y} Σ_{y y}^{- 1} Σ_{y z} . \end{matrix}$

(3)

The ${\hat{Σ}}_{z ∣ y}$ matrix, representing the conditional covariance of z given y, is also called the Schur complement of Σ_yy in Σ_xx. Government note that the posterior covariance does non depend happening the unique realization of y.

The Gauss-Bayes full point figurer for the Mary Leontyne Pric prediction, the conditional mean ${\hat{z}}_{z ∣ y}$ , minimizes the mean squared mistake of the calculate in the Gaussian case [28]. Moreover, in the Guassian type, for a specific observation y, the inverse of the probationary covariance is the Fisher Selective information ground substance related to with estimating z from y, and therefore ${\hat{Σ}}_{z ∣ y}$ is the lower bound on the error covariance matrix for whatsoever unbiased estimator of z [28].

The same set of equations arise in Kalman's filtering. Kalman's own view of this process is as a entirely deterministic operation [48], and does not rely on assuming normality. Although the point estimator ${\hat{z}}_{z ∣ y}$ is optimum in term of awful square wrongdoing, in practice there are numerical complications involved in this method acting: The matrix Σ_yy is typically not well conditioned, so the numerical calculation of $Σ_{y y}^{- 1}$ cannot always be trusted. To overcome this problem, we aim a amend conditioned computer, which has a behavior at hand to Gauss-Thomas Bayes.

Principal components and estimation in lower berth dimension

Star component analysis (PCA) is a well-established mathematical procedure for dimensionality decrease of information and has wide applications across various fields. In this work, we consider its application in foretelling stock prices.

Consider the single note value decomposition (SVD) of Σ_xx:

where S is a diagonal matrix of the same dimension as x with non-negative diagonal elements in decreasing order, and V is a unitary matrix (VV′ = I _N). The diagonal elements of S are the eigenvalues of Σ_xx.

In general, the first few eigenvalues account for the bulk of the tally of all the eigenvalues. The "outsized" eigenvalues are called the principal eigenvalues. The corresponding eigenvectors are known as the principal components.

Let L danlt; N be such that the first L eigenvalues in S account for the bulk part (say 85% or more) of the sum of the eigenvalues. Let V _L be the first L columns of unitary matrix V. Then the random vector x is approximately equal to the linear combining of the first L columns of V:

where α is a random vector of length L. Because L is a low number compared to N, Eq (5) suggests that a little "noisy" subspace with a lower berth dimension than N can represent to the highest degree of the information. Projecting onto this principle subspace throne resolve the ill-conditioned problem of Σ_yy. The idea is that instead of including all eigenvalues in representing Σ_xx, which vary greatly in magnitude, we use a subset which only includes the "large" ones, and therefore the range of eigenvalues is significantly small. The same concept is implemented in speed signal subspace filtering methods, which are settled on the orthogonal decomposition of noisy speech observation space onto a signal subspace and a resound subspace [43]. Let V _M,L beryllium the first M rows and first L columns of V. We have

Mathematically resolution noisy observation vector y onto the principle subspace commode be written as a filtering cognitive process in the take form of

where G is given by

The vector w is in reality the coordinates of the orthogonal projection of y onto the subspace adequate to the range of V _M,L. We give notice also think of w as an estimate of α supported least squares. Subbing y by w in (3) leads to a better fit set of equations:

$\begin{matrix} {\hat{z}}_{z ∣ w} & = Σ_{z w} Σ_{w w}^{- 1} w \\ {\hat{Σ}}_{z ∣ w} & = Σ_{z z} - Σ_{z w} Σ_{w w}^{- 1} Σ_{w z}, \end{matrix}$

(9)

because the condition routine of Σ_ww is much lower than that of Σ_yy, equally we will demonstrate subsequent. In (9) we have

and

If the posterior distribution of z estimated based on (9) has a similar deportment to the dispersion estimated by (3), it butt be considered a obedient deputize for the Gauss-Thomas Bayes method. Our numeric results demonstrate that this is indeed the case, which we will show in Section.

Moving average

Field traders and investors frequently purpose technical trading rules, and one of the most touristy methods used past technical traders and researchers are the moving average (MA) rules [49, 50]. Satchell investigates the reasonableness general MA trading rules are widely exploited by technical analysts [51]. He shows that autocorrelation amplification is one of the reasons such trading rules are democratic. Using simulated results, we show that the MA rule may be favorite because it can identify the toll momentum and is a lancelike means of assessing and exploiting the price autocorrelation without necessarily knowing its precise anatomical structure. Moving common, which is the average of prices complete a time period of time, is probably the simplest reckoner for z:

$\begin{matrix} {\hat{z}}_{Momma} & = \frac{1}{K_{M A}} \sum_{i = N - K_{M A} + 1}^{N} x_{i} \end{matrix}$

(12)

where the quantity K _MA is the bi of data points included to work out the ordinary, and ${\hat{z}}_{M A}$ is the average of the most Recent epoch K _MA price values.

There are different likely values of K _{Bay State} for calculating the normal, from short to medium to long terminus periods. Here we use periods of 10 and 50 days, which are typical short and midterm values used in the literature. We will use the moving modal estimator for comparison purposes, as we will see in Section below.

Fig 1 shows an example of our stock predictions. Assume that we are given the Price values for the past 20 years (M = 20), and we require to use those values to predict the future prices over the succeeding 10 byplay days, from day M + 1 to day N (N = 30). In our reduced-proportion technique, we can receive a relatively politic plot of the foreseen value for a relatively small L, to a plot about the same as Gauss-Bayes, for larger values of L, as we can buns see in Libyan Fighting Group 1.

An external file that holds a picture, illustration, etc. Object name is pone.0230124.g001.jpg

Predicting price for M + 1 to N days, actual price: Solid argumentation, GB: −o−, RD: −*− (2 lines, one for a small value of L, and ace for a relatively large enumerate).

Performance metrics

Imply squared error

To liken the performance of the methods described supra, we evaluate the expected apprais of the squared misplay between the actual and estimated values. The mean squared error of an estimate $\hat{z}$ is given by:

$\begin{matrix} M S E = E [\begin{matrix} ‖ z - \hat{z} ‖^{2} \end{matrix}] = E [\begin{matrix} ‖ z ‖^{2} \end{matrix}] + E [\begin{matrix} ‖ \hat{z} ‖^{2} \end{matrix}] - 2 E [\begin{matrix} ‖ z^{'} \hat{z} ‖ \end{matrix}] . \end{matrix}$

The MSE buttocks be expressed in terms of the covariance matrices in (2), by subbing the capture form of $\hat{z}$ . Alternatively, the mean squared error of an estimator $\hat{z}$ can Be written in terms of the variance of the estimator plus its squared oblique. The conditional MSE given x is written as

$\begin{matrix} M S E_{\hat{z} ∣ z} = E [\begin{matrix} ‖ z - \hat{z} ‖^{2} ∣ z \end{matrix}] = t r a c e (Σ_{\hat{z} ∣ z}) + ‖ E [\begin{matrix} \hat{z} ∣ z \end{matrix}] - z ‖^{2} . \end{matrix}$

The first full term is called the variance, and the second term is the squared bias. The expected valuate of MSE o'er all observations is the current MSE, which can be measured aside taking expectations on both sides:

$\begin{matrix} M S E = E [\begin{matrix} E [\begin{matrix} ‖ z - \hat{z} ‖^{2} ∣ z \end{matrix}] \end{matrix}] = t r a c e (E ‖ Σ_{\hat{z} ∣ z} ‖) + E [\begin{matrix} ‖ E [\begin{matrix} \hat{z} ∣ z \end{matrix}] - z ‖^{2} \end{matrix}] . \end{matrix}$

(13)

It turns out that Gauss-Bayes estimator is unbiased, which means that the second term is 0, piece the proposed reduced-dimension methods is a biased figurer.

Directional change statistic

$\begin{matrix} b_{i j} = {\begin{matrix} 1, & if (z_{i j} - z_{0}) ({\hat{z}}_{i j} - z_{0}) dangt; 0 \\ 0, & other \end{matrix} . \end{matrix}$

(14)

Then D _j, the direction statistic for sidereal day j, averaged over K samples, is equal to

which is a count between 0 and 1 (the higher the advisable).

Empirical methodology and results

In that subdivision we describe how we estimate the covariance matrix based on a normalized data prepare, and we judge the carrying into action of our method using empirical data.

National mise en scene

Suppose that we have K samples of vector data, for each one of length N, where N danlt; K. Call these row vectors x ₁, x ₂, …, x _K, where each $x_{i} \in R^{N} (i = 1, \dots, K)$ is a row vector of length N:

We assume that the vectors x ₁, x ₂, …, x _K are drawn from the same underlying statistical distribution. We lavatory stack these vectors unneurotic every bit rows of a K × N matrix:

$\begin{matrix} X = [\begin{matrix} x_{11} & x_{12} & \dots & x_{1 N} \\ x_{21} & x_{22} & \dots & x_{2 N} \\ \dots & \dots & \dots & \dots \\ x_{K 1} & x_{K 2} & \dots & x_{K N} \end{matrix}] . \end{matrix}$

Let M ≤ N and suppose that we are given a vector $y \in R^{M}$ representing the first M data points of a vector we believe is drawn from the same distribution as x ₁, x ₂, …, x _K. Again, these M data points typify the end-of-Clarence Shepard Day Jr. prices of a company stock over the past M sequentially trading days. Countenance z personify the price of the next N − M days in the forthcoming. We wish to estimate z from y.

Since the vector x _i is a multivariate random vector that can be partitioned in the form

where y _i has length M and z _i has length N − M, accordingly the information ground substance X can be divided into two sub-matrices Y and Z as follow:

We can think of Y arsenic a information matrix consisting of samples of historical data, and Z as a data matrix consisting of the corresponding future values of prices.

Normalizing and centering the data

In the case of stock-price data, the vectors x ₁, x ₂, …, x _K might come from prices spanning different months or many. If then, the basic assumption that they are drawn from the same distribution may non hold because the value of a America dollar has changed over time, as a result of splashines. To get the better of this issue, a scaling approach should be victimized to meaningfully normalise the prices (we wish deal with the time-varying nature of the covariance later). One such approach is presented hither. Suppose that t _i = [t _i1, t _i2, …, t _iN] is a transmitter of "raw" (unprocessed) stock prices over N consecutive trading days. Suppose that Q ≤ N is as wel given. Then we hold the following normalization to obtain x _i:

This normalization has the interpretation that the x _i transmitter contains strain prices as a fraction of the value on the Qthorium day, and is meaningful if we believe that the pattern of much fractions over the days 1, …, N are drawn from the same distribution. Note that x _i(Q) = 1.

We believe normalizing the data with this method captures the radiation diagram in the price information better than simply using return data. Although similar to return, the resulting time series still suffers from organism non-unmoving finished time. We propose to decide this issue by using a weighting averaging method as explained in the close section.

For the role of applying our method founded on PCA, we assume that the vectors x ₁, x ₂, …, x _K are closed from the cookie-cutter underlying distribution and that the mean value, $\bar{x}$ , is up to zipp. However because x _i represents price values, in general the mean is non zero. The mean $\bar{x}$ can be estimated by averaging the vector $x_{i} \in R^{N} (i = 1, \dots, K)$ ,

and then this average transmitter is deducted from each x _i to center the data.

Even though this normalization makes the information stationary in the mean, since stock prices are very evaporable, there is no guarantee that the covariance of the information would be stationary likewise. In order to address this issue, we attribute exponential weights (γ ⁰, γ ¹, ⋯, γ ^k) to observations, where 0 danlt; γ danlt; 1, to emphasize the most recent periods of data. Using an exponential weight approach to business deal with volatility of financial data has been suggested in multiple studies much every bit [52]. For each observation x _i, the last K samples antecedent to that observation are transformed into a Hankel intercellular substance and normalized. Past (decreasing) exponential function weights are assigned to the K samples and numerical results are premeditated. This process, creating the matrix of information, normalizing, and assigning weights, is recurrent for each observation.

To select the value of K we utilize

Experiments

The daily historical Leontyne Price data for 150 different companies from distinct market-capitalization categories were downloaded from finance.yahoo.com. Market capitalization is a measure out of the company's wealth and refers to the total value of every a company's shares of stock. We randomly select 50 stocks from for each one of the three market capitalisation (cap for short) categories: Big market-cap (125 B$ to 922 B$), Mid market-cap (2 B$ to 10 B$) and Small market-cap (300 M$ to 1.2 B$). The stocks from the Big market-cap category are commonly the nearly stable ones relative to the Small-capitalisation stocks, which have the most volatility. Historical information for iv market indexes, Sdanamp;P500 (GSPC), Dow Jones Industrial Average (DJI), NASDAQ Composite (IXIC), and Russell 2000 (Furrow), were also included in this study. The data was transformed into matrices with different sizes as explained in incoming section. In each case, the daily price value for next 10 days are foretold and the approximation methods are compare based on their out-of-sample performance.

Constructing data matrix

The regular farm animal price data is transformed into a matrix with K rows, samples of vector data, each of length N. We get that by stacking K rows (K samples), each one time shifted from the previous one, all in one big matrix, called the Hankel matrix.

More on the dot, the Hankel intercellular substance for this problem is constructed in the following format:

$\begin{matrix} [\begin{matrix} t_{1} \\ t_{2} \\ ⋮ \\ t_{K} \end{matrix}] = [\begin{matrix} P (1) & P (2) & \dots & P (N) \\ P (2) & P (3) & \dots & P (N + 1) \\ \dots & \dots & \dots & \dots \\ P (K) & P (K + 1) & \dots & P (K + N - 1) \end{matrix}], \end{matrix}$

where P(i) represents the price for day i. This is our matrix of data, before normalization and centering.

We first normalize each row (observation) by Qth ingress, A described earlier, so subtract the average vector $\bar{x}$ from each dustup. The prediction is done exploitation the processed data. Subsequently doing the prediction, we add back the average vector ${\bar{x}}_{N - M}$ (shoemaker's last N − M components of $\bar{x}$ ) from years M + 1 through N and also procreate the upshot by the value of Qatomic number 90 that was used for normalizing to get back to actual stock prices. We tested different values for Q in terms of MSE and estimation variability. For the purpose of this study, we chose Q = M because it shows the uncomparable results in this setting. Recall that x _i(M) = 1. This column is removed from the data intercellular substance because it does not provide any information. From now connected matrix X represents normalized and centered price information.

To account for the nonstationarity of the covariance, we use an exponential averaging method acting as mentioned before. For this purpose, γ = 0.98 was selected and the weights smaller than 10⁻³ were well thought out zero. Then the sample covariance matrix is calculated as

$\begin{matrix} Σ_{x x} = (\frac{1 - γ}{1 - γ^{k + 1}}) X^{^{'}} diag (γ^{0}, γ^{1}, \dots, γ^{k}) X, \end{matrix}$

(21)

where diag(γ ⁰, γ ¹, ⋯, γ ^k) is a oblique matrix with (γ ⁰, γ ¹, ⋯, γ ^k) as the diagonal elements.

We obtained death-of-day stock prices for General Electrical and converted this time series into Hankel matrices with distinct lengths as described above. 2000 samples were used to evaluate the out-of-sample performance of the methods. The values corresponding with the performance prosody presented therein division meet after a couple of hundred samples. We construct data matrices with 9 different sizes, M from 50 to 530 with a 60 day interval, to investigate the effect of length of observation vector on carrying out.

Fig 2 shows the histogram of normalized data as a internal representation of the statistical distribution of normalized data; the curve resembles a bell shape.

An external file that holds a picture, illustration, etc. Object name is pone.0230124.g002.jpg

Histogram graph for normalized data.

MSE performance

Deuce-ac different approximation methods are enforced for to each one of the data matrices constructed above. The end is to forecas future damage values for the next 10 days (days M + 1 to N). when it comes to reduced-dimension method, for each M we try different values of L, the number of rationale components. The general goal, as mentioned supra, is an estimation technique that has a similar behavior as an ideal Karl Gauss-Bayes figurer only does not have the associated calculation difficulties resulting from ill-conditioning.

We use General Electric price data to calculate the values illustrated in this division. We calculate the squared error (SE) for 2000 samples to evaluate the performance of the methods. We implement our reduced-attribute technique for polar Ms, and for unlike numbers of primary eigenvalues, L.

Figure 3 shows the empirical Cumulative statistical distribution procedure (CDF) of the Sou'-east for 2 different values of M, together with cardinal-standard-deviation self-confidence musical interval. Note that to make our comparisons indifferent and meaningful, we normalized the results from the moving average predictors so that their values are equally normalized with the values from our RD method acting. When it comes to proscribed-of-sample performance, the numerical complications compromise the idea accuracy of Gauss-Bayes, causing the SE values for this method to become smooth worsened than the Southeastward plot for the mobile average estimators. As we can see, in both plots, our bated-attribute method acting is leading to the other two methods. For M = 110 extraordinary lines are relatively approximate. Arsenic M gets bigger, the plot for the reduced-attribute method improves and the secret plan for Gauss-Bayes gets worse.

An external file that holds a picture, illustration, etc. Object name is pone.0230124.g003.jpg

Empirical CDF of SE corresponding to M = 110 and 290.

MA₂₀ and MA₅₀: −o−, Gilbert: −*−, RD: Solid lines. Broken lines illustrate a two standard deviation confidence musical interval. Plots toward the top and leftish represent better performance.

Another point worth mentioning is that although adding more information improves the carrying into action of our planned method acting, that is not the cause for the moving average estimator. As the arrow on the plot of ground on the bottom indicates, away adding more information, moving from $z_{M A_{10}}$ to $z_{M A_{50}}$ , the performance of the moving average estimator deteriorates. This behavior is expected since the moving average relies happening the impulse, in contrast to the reduced-property method acting, which extracts the essence of the information by projecting onto a smaller subspace.

Fig 4 shows the values of MSE over all days of estimation versus the value of L, for 9 different M, lengths of observation vector, from 50 to 530. As we can see, the MSE time value is hard to the note value of L for sufficiently large L. For small values of L, the MSE values fall quickly, but then eventually step-up. Indeed if we feature a particular constraint on the condition number, we do not lose so much in terms of MSE by choosing a attenuated-dimension subspace, which leads to a finer conditioned problem. After a certain point, adding Thomas More data is really adding make noise and the MSE values stupefy worse.

An external file that holds a picture, illustration, etc. Object name is pone.0230124.g004.jpg

MSE versus L in the normalized domain for different Ms.

The metric we are looking for is the summate of MSE values over wholly years of estimate.

For all distance of M, the values for MSE are captured supported different constraints of the discipline number of Σ_ww. The MSE values in the reduced-proportion method are significantly smaller relative to the other two methods.

Fig 5 shows the relational percentage of melioration (RPI) in the minimized-dimension method compared to the other two methods, deliberate as

$\begin{matrix} R P I_{GB/MA} = \frac{- 100 (M S E_{RD} - M S E_{Sarin/Mommy})}{M S E_{GB/Artium Magister}} . \end{matrix}$

(22)

Notation that since the denominator in the equivalence is MSE _GB/MA, the improvement percentage does not exceed 100% but the actualised MSE values are further apart in absolute terms than illustrated here. For example for M = 350, the MSE value for reduced-dimension is between 0.0052 to 0.018, while the MSE in Gauss-Thomas Bayes is just about 6.33 × 10⁶. The iii (overlapping and therefore appears as single a single plot) lines on top (-*-) of Fig 5 compare the shrunken-dimension to Gauss-Bayes (RPI _GB). The three lines on crown (‥o‥) stand for to the comparison of the reduced-dimension and moving average ( $R P I_{M A_{50}}$ ) and the three lines on the bottom (‥o‥) correspond to ( $R P I_{M A_{10}}$ ). In each case the three lines are subject to different amphetamine limits on the condition number (10², 10³, and 10⁴). It is worth mentioning that the condition number of Σ_yy starts from 10³ for M = 50 and goes ascending to 10¹⁹ for M = 530. The upper demarcation on the specify number of Σ_ww changes from 10², associated with the lines on the derriere in each vitrine, to 10⁴, the lines on top, for all values of M.

An external file that holds a picture, illustration, etc. Object name is pone.0230124.g005.jpg

RPI values, subject to antithetic upper limit on precondition number of Σ_ww, in for each one example 10² associated with the line on the bottom, to 10⁴ associated with the line along top, RPI _Mommy: ‥o‥, RPI _GB: −*−.

Higher plots represent worse relative operation (relative to RD).

In general, away increasing M, to a greater extent information is available in all reflection, resulting in better performance of the prediction in terms of smallest MSE values. This can be observed easy in the RPI plots in Fig 5 in comparison to the moving average cases since the MSE values in the those cases are almost unswerving for different values of M. The percent of improvement of MSE values corresponding to the reduced-dimension method increases as M increases. This is as expected since more data is available in from each one observation, resulting in fitter performance. However after a certain point the RPI flattens out suggesting adding more information at this point is increasing the disturbance and does not improve the performance.

Every bit we can see, in some cases in that respect is a cold-shoulder reduction in the improvement rate of the reduced-dimension method compared to the moving average method. A viable account for this observation is that when we fix some constraint on condition number, we are really limiting the apprais of L, and by augmentative M, after a certain point, we mostly growth the noise, and the MSE value gets worse, which is consistent with Fig 4. Table 1 shows the average RPI values for all stocks in different commercialise-cap categories and median RPI values for market indexes. The reduced-dimension method consistently shows better performance than the other two methods.

Put over 1

Average RPI values for stocks in opposite market-cap categories and average RPI values for market indexes (M = 350).

MSE	RPI _GB	$R P I_{M A_{10}}$	$R P I_{M A_{50}}$
Small-Cap	100%	51%	88%
Middle-Capital	100%	54%	88%
Big-Cap	100%	56%	89%
Food market indexes	100%	55%	88%

Matlab's two-sample distribution t-test function was used to determine the MSE values from our proposed method for 50 stocks in each market-cap category is importantly little than the average of the MSE values generated for the same sample victimisation other methods at 5% significance level (α = 0.05). When p danlt; α and h = 1, the void guess that the ii samples have the homophonic mean is rejected, final that the difference between the averages of the two sets of samples is statistically significant at α significance level. As shown in Table 2, the results argue that the average of the MSE values for predictions from our method is importantly smaller than the average of MSE values from otherwise competing methods at 0.05 significance level.

Prorogue 2

Statistical depth psychology for MSE values for stocks in different market cap categories (M = 350).

T-try	against	MSE _GB	$M S E_{M A_{10}}$	$M S E_{M A_{50}}$
Pocket-sized-Cap	p-value	0.0024	0.0075	0.00068
Pocket-sized-Cap	h	1	1	1
Mid-Cap	p-value	0.0283	0.0066	0.000038
Mid-Cap	h	1	1	1
Big-Cap	p-value	0.0021	0.00048	0.00001
Big-Cap	h	1	1	1

Recall that L represents the turn of eigenvalues required from the diagonal matrix S to represent the bulk part of the data carried in x. Fig 6 investigates the property of the target subspace by plotting the value of L corresponding to best MSE for different Ms, nonexempt to variant limits on condition number (the same case As in Fig 5).

An external file that holds a picture, illustration, etc. Object name is pone.0230124.g006.jpg

Best L corresponding to best MSE values nonexempt to different limits on condition routine, 10² associated with the line on the bottom, 10³ associated with the line in the middle, and 10⁴ associated with the line on top.

As the maximum on specify number increases, the value of MSE improves as M increases, and we need a bigger subspace, bigger L, to extract the information. However, as the bottom three plots in Fig 6 indicate, the value for best L flattens unstylish after a dependable point.

Social control vary statistic performance

The some other evaluation measured that we are interested in is the directing statistic which measures the matching of the actualized and predicted values in terms of directional change. Fig 7 shows the modal directional statistic over 10 days of estimation using the same K = 2000 samples. As the plot indicates, the reduced-dimension method is patronizing in terms of guiding change statistic. It is interesting to note that the directional statistic improves as M increases, then eventually flattens outgoing, consistent with previous plots.

An external file that holds a picture, illustration, etc. Object name is pone.0230124.g007.jpg

Topper directional statistics subject to different upper limit on consideration number of Σ_yy, 10² associated with the line on the bottom, to 10⁴ connected with the line on top, Gi: −*−, RD: Solid lines, MA: ‥o‥ (Mamma₁₀ on the bottom and Master of Arts₅₀ on top).

Higher plots represent best operation.

Prorogue 3 shows the norm assess for position statistic for stocks in different commercialize cap categories and indexes for M = 350 for Σ_ww condition number pocket-size to 10⁴. The reduced-dimension method acting is superior to the other two methods in terms of directional change estimation. It is important to note that the values represented in Put of 3 are joint with a specific M for completely companies. In practice, information technology is recommended to tailor the value of M for each company to get the good results.

Table 3

Average directional statistics for stocks in different market cap categories (M = 350).

Directional Statistic	MA₁₀	MA₅₀	GB	RD
Small-capitalisation	0.56	0.61	0.51	0.78
Mid-Cap	0.58	0.62	0.51	0.79
Enceinte-Detonating device	0.60	0.66	0.51	0.80
Market Indexes	0.63	0.70	0.50	0.79

Matlab's two-sample t-tryout function was used to determine if the average of the directional statistics from our method acting for 50 stocks is significantly larger than the ordinary of directional statistics from separate methods. Prorogue 4 lists the p-value and h-statistic for for each one exam. The results also indicate that the average of directional statistics from our method acting is importantly larger than the average of the directional statistics from other competing methods at 5% significance level.

Postpone 4

Applied math analysis for directional statistics values for stocks in several market-cap categories (M = 350).

T-quiz	against	D _Gilbert	$D_{M A_{10}}$	$D_{M A_{50}}$
Small-Pileus	p-value	danlt;10⁻¹⁰	danlt;10⁻¹⁰	danlt;10⁻¹⁰
Small-Pileus	h	1	1	1
Mid-Detonator	p-value	danlt;10⁻¹⁰	danlt;10⁻¹⁰	danlt;10⁻¹⁰
Mid-Detonator	h	1	1	1
Big-Cap	p-rate	danlt;10⁻¹⁰	danlt;10⁻¹⁰	danlt;10⁻¹⁰
Big-Cap	h	1	1	1

Volatility

Another important parameter that we estimate is the volatility of the prediction, measured in price of its casebook deviation. The square etymon of the virgule elements of the estimated covariance, ${\hat{Σ}}_{z z}$ , are the estimated standard deviations for individual days of estimation. The appraisal of the covariance in to each one method is

$\begin{matrix} {\hat{Σ}}_{G B} & = {\hat{Σ}}_{z ∣ y} = Σ_{z z} - Σ_{z y} Σ_{y y}^{- 1} Σ_{y z}, \\ {\hat{Σ}}_{R D} & = {\hat{Σ}}_{z ∣ w} = Σ_{z z} - Σ_{z w} Σ_{w w}^{- 1} Σ_{w z}, \end{matrix}$

(23)

However, promissory note that because of the poor conditioning of Σ_yy, using the formula above for Σ_GB has numerical issues. Hence, we neglect their values here. In general the standard departure values increase moving from day 1 to daylight 10 of prediction, since less uncertainty is involved in the approximation of descent prices of days closer to the current 24-hour interval. In Fig 8, the standard deviation for individual days of estimation, days 1 to 10, are plotted versus M, the length of observation transmitter, for the attenuated-dimension method acting. In the reduced-dimension method, the standard diversion values decrease as M increases because more information is provided in each observation. For sufficiently spacious Ms, the standard difference values for different years are very shut.

An external file that holds a picture, illustration, etc. Object name is pone.0230124.g008.jpg

Standard diversion of individual days of estimate, RD: Strong origin.

Conclusion

In that paper we introduced a new method for predicting future stock price values supported covariance information. We arise this method based along a filtering mental process using precept components to overcome the numerical complications of tentative think. We also introduced a procedure for normalizing the data. The ground substance of data was constructed in different sizes to investigate the effect of distance of observation vector on prediction performance. Our method has showed consistently better dead-of-sample performance than Gauss-Bayes (variable conditional mean), a numerically challenged estimator, and moving average, an easy to use estimator, for 5 variant companies in terms of mean value squared error and social control convert statistic.

The proposed method can be adapted to include duplex predictors. The import of the projected approach will be even more plain when using multiple predictors because where reflexion vectors are longer it becomes almost impossible to depend on conditional mean ascribable the severe carsick-conditioning of the covariance matrix.

Supporting info

Funding Statement

The author(s) acceptable no specific funding for this work.

Data Availability

All relevant data are within the manuscript and its Supporting Information files.

References

1. French Atomic number 36, Wheel R. Stock go back variances: The arrival of information and the reaction of traders. Diary of Financial Economics. 1986;17(1):5–26. 10.1016/0304-405X(86)90004-8 [CrossRef] [Google Bookman]

2. Lo AW, MacKinlay AC. Hackneyed market prices do not follow random walks: Evidence from a deltoid specification try out. The Recapitulation of Financial Studies. 1988;1(1):41–66. 10.1093/rfs/1.1.41 [CrossRef] [Google Scholar]

3. Fama EF, French Party of Democratic Kampuchea. Unchangeable and temporary components of stock prices. Journal of Sentiment Economy. 1988;96(2):246–273. 10.1086/261535 [CrossRef] [Google Scholar]

4. Cutler Diabetes mellitus, Poterba JM, Summers LH. Speculative kinetics. The Review of Economic Studies. 1991;58(3):529–546. 10.2307/2298010 [CrossRef] [Google Assimilator]

5. Chopra N, Lakonishok J, Ritter Jn, et alia. Performance measurement methodology and the inquiry of whether stocks overreact/1991: 130. BEBR module working paper; no 91-0130. 1991.

6. Bondt WF, Thaler R. Does the sprout market overreact? The Journal of Finance. 1985;40(3):793–805. 10.1111/j.1540-6261.1985.tb05004.x [CrossRef] [Google Scholar]

7. Bogousslavsky V. Occasional rebalancing, render autocorrelation, and seasonality. The Journal of Finance. 2022;71(6):2967–3006. 10.1111/jofi.12436 [CrossRef] [Google Scholar]

8. Chowdhury SSH, Rahman MA, Sadique Manuscript. Behaviour of stock reelect autocorrelation in the GCC stock markets. Global Business Review. 2022;16(5):737–746. 10.1177/0972150915591420 [CrossRef] [Google Scholar]

9. Lithium L, Leng S, Yang J, Yu M. Stock Securities industry Autoregressive Dynamics: A Multinational Comparative Examine with Quantile Regression. Mathematical Problems in Engineering. 2022;2016. [Google Scholar]

10. Chowdhury SSH, Rahman MA, Sadique MS. Stock return autocorrelation, day of the week and volatility. Limited review of Accounting and Finance. 2022. 10.1108/RAF-12-2014-0146 [CrossRef] [Google Learner]

11. Xue WJ, Zhang LW. Stock generate autocorrelations and predictability in the Chinese stock securities industry—Evidence from threshold quantile autoregressive models. Economic Modelling. 2022;60:391–401. 10.1016/j.econmod.2016.09.024 [CrossRef] [Google Scholar]

12. Fama EF, French KR. Dividend yields and expected stock returns. Diary of Financial Economics. 1988;22(1):3–25. 10.1016/0304-405X(88)90020-7 [CrossRef] [Google Scholar]

13. Jaffe J, Keim DB, Westerfield R. Profit yields, grocery store values, and stock returns. The Journal of Finance. 1989;44(1):135–148. 10.1111/j.1540-6261.1989.tb02408.x [CrossRef] [Google Scholar]

14. Fama EF, French Atomic number 36. The crossbreeding-section of unsurprising stock returns. The Journal of Finance. 1992;47(2):427–465. 10.1111/j.1540-6261.1992.tb04398.x [CrossRef] [Google Scholar]

15. Silvennoinen A, Teräsvirta T. Multivariate autoregressive conditional heteroskedasticity with smooth transitions in conditional correlations. SSE/EFI Working Theme Series in Economics and Finance; 2005.

16. Dennis P, Mayhew S, Stivers C. Strain returns, inexplicit volatility innovations, and the asymmetric volatility phenomenon. Diary of Financial and Numeric Analysis. 2006;41(2):381–406. 10.1017/S0022109000002118 [CrossRef] [Google Scholar]

17. Eom C, Park JW. Effects of lowborn factors on stock correlation networks and portfolio diversification. International Review of Fiscal Analysis. 2022;49:1–11. 10.1016/j.irfa.2016.11.007 [CrossRef] [Google Scholar]

18. Dimic N, Kiviaho J, Piljak V, Äijö J. Impingement of financial market doubt and macroeconomic factors on stock–bond coefficient of correlation in emerging markets. Search in International Business and Finance. 2022;36:41–51. 10.1016/j.ribaf.2015.09.001 [CrossRef] [Google Scholar]

19. Han L, Lv Q, Yin L. The effect of oil returns connected the stock markets network. Physica A: Statistical Mechanics and its Applications. 2022;533:122044 10.1016/j.physa.2019.122044 [CrossRef] [Google Student]

20. Augustyniak M. Utmost likeliness idea of the Markov-switching GARCH model. Computational Statistics danampere; Data Analysis. 2022;76:61–75. 10.1016/j.csda.2013.01.026 [CrossRef] [Google Scholar]

21. Stein C. Estimation of a covariance matrix, Rietz Lecture. In: 39th Period of time Meeting IMS, Capital of Georgia, GA, 1975; 1975.

22. Ledoit O, Wolf M. A well-conditioned reckoner for large-magnitude covariance matrices. Diary of Multivariate Depth psychology. 2004;88(2):365–411. 10.1016/S0047-259X(03)00096-4 [CrossRef] [Google Scholar]

23. Ohno S, Ando T. Stock return predictability: A factor in-augmented predictive regression system with shrinkage method. Econometric Reviews. 2022;37(1):29–60. 10.1080/07474938.2014.977086 [CrossRef] [Google Scholar]

24. Yang L, Couillet R, McKay MR. Minimum variance portfolio optimization with robust shrinkage covariance estimation. In: 2022 48th Asilomar Conference connected Signals, Systems and Computers. IEEE; 2022. 1326–1330.

25. Ledoit O, Wolf M. Improved estimation of the covariance matrix of stock returns with an application to portfolio selection. Journal of Experiential Finance. 2003;10(5):603–621. 10.1016/S0927-5398(03)00007-0 [CrossRef] [Google Scholarly person]

26. Wilder JW. Recent concepts in commercial trading systems. Trend Research; 1978. [Google Scholar]

27. Rougier J. Ensemble averaging and mean squared error. Daybook of Climate. 2022;29(24):8865–8870. 10.1175/JCLI-D-16-0012.1 [CrossRef] [Google Scholar]

28. Scharf LL, Demeure C. Statistical signal processing: detection, estimation, and time series analysis. vol. 63 Addison-Wesley Reading, MA; 1991. [Google Scholar]

29. Hotelling H. Depth psychology of a complex of applied math variables into principal components. Journal of Educational Psychological science. 1933;24(6):417 10.1037/h0071325 [CrossRef] [Google Scholar]

30. Jolliffe I. Of import component analysis In: International Encyclopedia of Applied mathematics Scientific discipline. Springer; 2011. 1094–1096. [Google Scholar]

31. Ince H, Trafalis T. Kernel principal component analysis and support vector machines for stock price prediction. IIE Transactions. 2007;39(6):629–637. 10.1080/07408170600897486 [CrossRef] [Google Scholar]

32. Fung W, Hsieh DA. Empirical characteristics of dynamic trading strategies: The event of hedge funds. The Review of Financial Studies. 1997;10(2):275–302. 10.1093/rfs/10.2.275 [CrossRef] [Google Scholar]

33. Alexander C. Market risk analysis, value at risk models. vol. 4 John Wiley danamp; Sons; 2009. [Google Scholar]

34. Shukla R, Trzcinka C. Sequential tests of the arbitrage pricing theory: a comparison of principal components and maximum likelihood factors. The Journal of Finance. 1990;45(5):1541–1564. 10.1111/j.1540-6261.1990.tb03727.x [CrossRef] [Google Scholar]

35. Yu H, Chen R, Zhang G. A SVM stock selection poser within PCA. Procedia electronic computer skill. 2022;31:406–412. 10.1016/j.procs.2014.05.284 [CrossRef] [Google Scholarly person]

36. Zhong X, Enke D. Forecasting daily stock market comeback using dimensionality decrease. Expert Systems with Applications. 2022;67:126–139. 10.1016/j.eswa.2016.09.027 [CrossRef] [Google Scholar]

37. Wang J, Wang J. Forecasting livestock market indexes using precept component analysis and stochastic time effective neural networks. Neurocomputing. 2022;156:68–78. 10.1016/j.neucom.2014.12.084 [CrossRef] [Google Scholar]

38. Pasini G. Corpus ingredient analysis for stock portfolio management. International Journal of Light and Applied math. 2022;115(1):153–167. 10.12732/ijpam.v115i1.12 [CrossRef] [Google Assimilator]

39. Narayan PK, Narayan S, Thuraisamy Kansas. Can institutions and macroeconomic factors predict stock returns in emerging markets? Emerging Markets Review. 2022;19:77–95. 10.1016/j.ememar.2014.04.005 [CrossRef] [Google Scholar]

40. Velicer WF, Capital of Mississippi DN. Component analysis versus plebeian factor analysis: Some issues in selecting an apt procedure. Variable behavioral search. 1990;25(1):1–28. 10.1207/s15327906mbr2501_1 [PubMed] [CrossRef] [Google Scholar]

41. Bartholomew DJ, Steele F, Moustaki I. Analysis of variable social group science data. Chapman and Hall/CRC; 2008. [Google Scholar]

42. Meglen RR. Examining large databases: a chemometric glide slope victimisation principal component analysis. Marine Chemistry. 1992;39(1-3):217–237. 10.1016/0304-4203(92)90103-H [CrossRef] [Google Scholarly person]

43. Hermus K, Wambacq P, et al. A review of signalise subspace speech enhancement and its application to noise robust oral communicatio realization. EURASIP Diary on Advances in Signal Processing. 2006;2007(1):045821 10.1155/2007/45821 [CrossRef] [Google Bookman]

44. Tufts DW, Kumaresan R, Kirsteins I. Information adaptive signal estimation past remarkable prise decomposition of a data matrix. Proceedings of the IEEE. 1982;70(6):684–685. 10.1109/PROC.1982.12367 [CrossRef] [Google Scholar]

45. Ballings M, Van den Poel D, Hespeels N, Gryp R. Evaluating aggregate classifiers for stock price direction prediction. Expert Systems with Applications. 2022;42(20):7046–7056. 10.1016/j.eswa.2015.05.013 [CrossRef] [Google Scholar]

46. Yao J, Suntan CL. A grammatical case study on using neural networks to execute technological forecasting of forex. Neurocomputing. 2000;34(1-4):79–98. 10.1016/S0925-2312(00)00300-3 [CrossRef] [Google Scholar]

47. Torun MU, Akansu AN, Avellaneda M. Portfolio risk of exposure in five-fold frequencies. IEEE Signal Processing Magazine. 2011;28(5):61–71. 10.1109/MSP.2011.941552 [CrossRef] [Google Scholar]

48. Byrnes 101, Lindquist A, Chow dynast Y. On the nonlinear dynamics of fast filtering algorithms. SIAM Diary on Control and Optimisation. 1994;32(3):744–789. 10.1137/S0363012991202422 [CrossRef] [Google Scholar]

49. Brock W, Lakonishok J, LeBaron B. Simple technical trading rules and the stochastic properties of stock returns. The Journal of finance. 1992;47(5):1731–1764. 10.1111/j.1540-6261.1992.tb04681.x [CrossRef] [Google Scholar]

50. Taylor Military police, Allen H. The role of technical analysis in the international exchange market. Journal of world Money and Finance. 1992;11(3):304–314. 10.1016/0261-5606(92)90048-3 [CrossRef] [Google Scholarly person]

51. Hong K, Satchell S. Time series momentum trading strategy and autocorrelation gain. Amount Finance. 2022;15(9):1471–1487. 10.1080/14697688.2014.1000951 [CrossRef] [Google Learner]

52. Pafka S, Potters M, Kondor I. Exponential weighting and random-matrix-theory-based filtering of financial covariance matrices for portfolio optimization. arXiv preprint cond-mat/0402573. 2004.