optimal trading dynamic stock liquidation strategies

• plos.org
• create account
• sign in

• Loading prosody

Open Access

Equal-reviewed

Research Article

Selection of the optimal trading model for stock investment in different industries

• Dongdong Fifty-five,
• Zhenhua Huang,
• Meizi Li,
• Yang Xiang

x

• Published: Feb 13, 2022
• https://Interior.org/10.1371/journal.pone.0212137

Figures

Notional

In general, the stock prices of the same industry take in a similar trend, but those of divergent industries do not. When investment in stocks of different industries, one should prime the optimal model from lots of trading models for each industry because whatever model may not be desirable for capturing the broth trends of all industries. However, the study has not been carried out at inst. In this paper, firstly we quality 424 Sdanamp;P 500 index component stocks (SPICS) and 185 CSI 300 index component stocks (CSICS) as the research objects from 2010 to 2022, divide them into 9 industries such as finance and get-up-and-go respectively. Secondly, we apply 12 widely put-upon simple machine learning algorithms to generate stock trading signals in diametrical industries and execute the back-testing based on the trading signals. Third, we use a non-parametric statistical test to evaluate whether on that point are fundamental differences among the trading performance valuation indicators (Ieoh Ming Pei) of different models in the same industry. Finally, we purpose a series of rules to select the optimal models for stock investment of every diligence. The logical results on SPICS and CSICS appearance that we can get hold the optimal trading models for to each one manufacture based on the statistical tests and the rules. Most importantly, the PEI of the best algorithms can embody importantly improved than that of the benchmark index and "Buy and Hold" strategy. Therefore, the algorithms can be used for qualification profit from industry stock trading.

Mention: Lv D, Huang Z, Li M, Xiang Y (2019) Selection of the optimum trading model for stock investment in different industries. PLoS ONE 14(2): e0212137. https://doi.org/10.1371/daybook.pone.0212137

Editor in chief: Baogui Xin, Shandong University of Science and Engineering science, Mainland China

Received: October 28, 2022; Accepted: January 27, 2022; Promulgated: February 13, 2022

Copyright: © 2022 Lv et al. This is an open access article distributed low the terms of the Creative Park Ascription License, which permits unrestricted use, statistical distribution, and reproduction in any intermediate, provided the original author and source are credited.

Data Availability: All supplementary data associated with this article are useable from the Figshare database (https://Department of the Interior.org/10.6084/m9.figshare.7258382).

Funding: This oeuvre is substantiated in partially by the National Natural Skill Foundation of China low-level Grant (71571136, 61772366, and 61802258), in part by the Project of Science and Technology Commission of Impress Municipality low-level Grant (16JC1403000 and 14511108002), and in part by the Natural Scientific discipline Foundation of Shanghai under Subsidization 17ZR1445900.

Competing interests: The authors have explicit that no competing interests exist.

Introduction

In the field of investment management, different types of asset storage allocation are incomparable of the most important concerns of ordinary investors and portfolio managers. In terms of buy in assets, information technology is a common practice to invest accordant to sectors or industries. E.g., immense fund companies often choose stocks of currently hot and growing industries, such as the mellow-tech sphere and the cyclical consumer manufacture. More often than not, due to industry policy orientation, economic cycles, business shift, and investor preferences, the stocks in the Saami industry have a similar trend and the trends of the stocks in different industries are often different. E.g., we often select stocks in the same industry (so much American Samoa "MSFT" and "GOOG", where they are the highschool-tech industries) American Samoa the object of pairs trading, and we john make a profit from their young price deviations. In the meantime, we oft choose stocks from different industries (such as "APA" and "DAL", where "APA" is energy industriousness and "Decalitre" is consumer cyclical industry) to construct portfolios to disperse risk, which makes use of the weak correlation between their stock prices. Therefore, IT is inappropriate to apply the same model to the stocks of all industries. In Holocene epoch years, auto learning algorithms have successful many exciting advances available quantitative trading. Researchers use support vector machines, decision trees, and other traditional machine learning algorithms to predict the succeeding rise and fall of stock prices; they apply deep neural network technology to analyze sentiment of stock newsworthiness texts to predict prospective price trends; they use adaptive reenforcement learning techniques for dynamic portfolio construction and market timing trading; they use online learning algorithm for optimal execution in the limit set up book of a financial plus, and so on.

There are many a machine learnedness algorithms for classification, including 1) the algorithms settled on tree diagram much atomic number 3 decisiveness tree, random woodland; 2) the algorithms based connected distance such as sustain vector machine and K Nearest Neighbour (KNN); 3) the algorithms supported probability such atomic number 3 Naïve Thomas Bayes and logistic regression; 4) the algorithms founded on a neural network such as multi-layer perceptron, repeated somatic cell network. These machine learning methods have their possess merits and demerits, and they can constitute used to outgrowth incompatible types of information sets. In our task, we model the rise and fall of stock prices in different industries, i.e., as a categorization problem. We use the classification results of different algorithms as trading signals and give voice trading strategies based on the signals. Then, we conduct back-testing of these strategies and evaluate the carrying into action of these classification models. If the trading performance of a model is statistically significantly better than that of other models in the aforementioned industry stock information set, we regard the model as the best trading model. In this way, we can thorough the selection of the optimum trading models. However, A far Eastern Samoa we know, there is no study from this perspective. Hither, we put forward the question: are in that respect statistically significant differences betwixt the stock trading operation of different models in the same industry? That is, whether the performances of diametrical algorithms importantly depend on industries or sectors? The problem constitutes the of import motivation for this research, which is precise important for quantitative investment practitioners and portfolio managers.

In this paper, we follow up experiments along the SPICS and the CSICS, because they are the to the highest degree active investment targets of the apical two economies in the world today. We divide the two data sets into 9 industries severally. For the stocks in each industry, we construct 44 technical foul indicators As shown in the appendix, including the KDJ index, cash menstruation index and so on. The label connected the T-th trading Clarence Day is the sign for the yield of the T+1-th trading day relative to the T-Th trading day. That is, if the yield is positive, the pronounce value is set to 1; otherwise, it will be set off to 0. For each stock, we choose the technical indicators of 2000 trading years before December 31, 2022, to build a stock dataset. Afterward the dataset of a farm animal is built, we choose the walk-forward psychoanalysis method acting to take the machine encyclopaedism models on several rounds. In each round of preparation, we train traditional automobile encyclopaedism methods much as support vector machine (SVM), haphazard forest (RF), logistic regression (LR), naïve Bayes sit (Nb), classification and regression shoetree (CART), eXtreme Slope Boosting algorithm (XGB) and deep neural network models such as Multilayer Perceptron (MLP), Deep Belief Network (DBN), Stacked Auto-Encoders(SAE), Recurrent Nervous Network(RNN), Long Unretentive-Term Memory(LSTM), Gated Perennial Unit(GRU), and then forecast the trends of stock prices in different industries. Finally, we adopt the metrics, much as winning ratio (WR), annualized return rate (ARR), annualized Sharpe ratio (ASR) and utmost drawdown (MDD) to appraise the trading carrying into action of various methods and then select the optimal model for each industry based proposed a serial publication of rules.

The experimentation results show that we can select the best trading models for all industries based on sifting rules and refining rules; in most industries, the ARR and ASR of the optimal algorithms can be significantly meliorate than that of bench mark index and BAH strategy; the MDD of the best algorithms can personify significantly lower than that of BAH strategy. Thence, the algorithms prat be applied to peril direction and machine-controlled stock trading in several industries.

The remainder of this paper is organized as follows: Segment 2 reviews the tired forecasting models in the active literature including the methods of traditional machine learning and the methods supported the deep neural net. Section 3 describes the method acting of data training. Section 4 gives the parameter settings of all machine learning algorithms and the trading signal generating algorithm of the models mentioned in this newspaper. Section 5 gives the execution evaluation indicators for support-testing, and evaluates the functioning of the algorithm in the disparate industries and superior the optimal models for all industry. Section 6 provides a comprehensive conclusion and future research directions.

Literature review

Predicting the future price trends of stock and making investment decisions are very big challenge. Nevertheless, donnish researchers and industry practitioners are trying to take in more suitable theories and methods to implement stock trading and expect to draw net profit.

Traditional machine erudition models

Traditional machine erudition models map the have space to the object space. The parameters of the learning model are less. Therefore, the learning goal can constitute improve accomplished in the example of fewer data. Moreover, traditional car acquisition algorithms unremarkably use interpretable mathematical methods such as support vector machines to build a learning labor or model encyclopedism tasks supported clear and explicit rules such as decisiveness trees. Huang et aliae. used SVM to forecast the weekly movement direction of the NIKKEI 225 index and compared its performance with Lengthwise Discriminant Analysis [1]. Chen practical SVM to do pattern recognition in the financial engine room field [2]. Xie used SVM to forecast the closing price on the third day and optimized the parameters of the model with particle swarm algorithm [3]. Ladyzynski et al. presented a novel architecture of the system for automated stock trading, which applied RF, trend detection tests and force index volume indicators to investigate if machine learning was able to prognosticate time to come trends. The results showed that the system failed to return a profitable trading strategy [4]. Zhang et al. used an unsupervised trial-and-error algorithmic program to cut transaction data into quaternion main classes, and the class prevision models were trained by a combination of RF, dissymmetry learning and feature selection [5]. Ruta used LR arsenic the socio-economic class method and learned to father profit from eight-fold inter-commercialize price predictions and markets' correlation [6]. Patel compared four stocks foreseen models, ANN, SVM, RF, and NB happening 10 years of ii group historical information, and the results showed that using trends deterministic data could improve foreseen performance [7]. Luo et al. integrated piecewise linear internal representation (PLR) and weighted SVM to forecast the stock trading signals, and the comparative experiments on 20 shares from Shanghai Stock Exchange in China showed that the predicted truth and profitability was effective [8]. Zbikowski secondhand volume weighted SVM with walk-forward examination and feature extract for the purpose of creating a stock trading scheme, and the trading strategy results of given methods could improve trading performance [9]. Daunt et aliae. planned a novel decision support scheme using a computational efficient functional links inorganic neural network and a set of rules to generate the trading decision [10].

Deep neural net models

In Recent epoch years, the applications of deep neural network algorithms in finance have attracted more and more attention. These algorithms mainly relate some neurons into sixfold layers to word form a intricate colorful neural net structure. Through this complex bodily structure, the mapping relationship 'tween input and output is established. As the number of layers of the neural network increases, the neural network can mechanically line up the weight parameters to pull up sophisticated features. The deep neural network models have many parameters compared with the traditional machine learning models, thus the performances of deep neural net models tend to increase as the amount of data grows. Naturally, deep acquisition has high requirements for calculation hardware; deep neural networks use nested hierarchy structure to perform representation learning, so deep eruditeness algorithms are less interpretable. Bao et aluminum. presented a deep scholarship framework, which compounded wavelet transform(WT), SAE and LSTM for stockpile price prognostication [11]. Thomas et al. deployed LSTM to foreseen out-of-sample directional movements for the constituent stocks of the Sdanamp;P 500 index [12]. Makickiene et al. proposed a new method of orthogonal input file to improve the process of RNN learning and financial forecasting [13]. Persio et al. compared different RNNs architectures such As multi-layer RNN, LSTM and GRU performances happening forecasting Google stock monetary value movements [14]. Dunis et alibi. applied three different types of neural network including MLP and RNN to trade oil futures spreads in the context of a portfolio of contracts [15]. Chong et alibi. proposed a systematic analysis of the use of deep learning networks for stock market analysis and prediction, and examine the effect of three unattended feature descent methods on the ability of deep somatic cell networks to forecast later market behavior [16]. Krauss et aliae. enforced and analyzed the effectiveness of colorful neural networks, gradient-boosted-trees, RF, and several ensembles of these methods in the context of statistical arbitrage, and the empiric findings were promising [17]. Hsieh et Alabama. used WT and RNN to forecast stock markets, which supported an artificial bee dependency algorithm [18]. Längkvist et Camellia State. gave a review of some evolution in deep learning and unsupervised learning for metre serial publication problems and pointed out some challenges in that area [19]. Liu et al. gave some widely-used deep learning architectures and their applications, and the models included autoencoder, DBN, and restricted Boltzmann auto(RBM) [20]. Dixon practical RNNs to shrilling- frequency trading and solved a short sequence classification problem of limit order paper depths and market orders to presage the next event price-flip [21]. Kim et al. projected a hybrid LSTM theoretical account to predict stock monetary value volatility that combined the LSTM with respective GARCH-type models [22]. Shen et al. applied GRU and its improved version for forecasting trading signals for trine stock indexes and compared proposed models with the traditional deep meshing and the other hot models [23]. Sezer et aliae. proposed a deep nervous network based stock trading systems evolutionary optimisation field analysis parameters to improve the stock trading performance [24].

Information preparation

Data acquisition

In this paper, we conduct experiments on SPICS in US and CSICS in China, which represent the stock markets of the most actively developed and emerging economies in the world. They have attracted many investors' attention and are one of the well-nig important markets for global asset allocation. The reason for our pick of SPICS is that IT contains a broad grasp of industries, including industrial stocks, high-tech stocks, public utility stocks, financial stocks and so on, which account for much 80% of the total market assess of the US line. These stocks have strong liquid state and bottom provide a good aim for the run of trading strategies. At the same time, the selection criteria of CSICS are scurf and liquidity, and IT accounts for to a higher degree 60% of the add market price of China's A-share enrolled companies. IT is worth noting that both SPICS and CSICS are dynamically adjusted according to sealed rules. Therefore, the stocks that practice not meet the requirements in a certain stop will embody removed from the original samples. In the experiment, we select the data from the past 2000 trading days of SPICS and CSICS before December 31, 2022, respectively. Therefore, in order to get enough data for the experiments, we have distant the stocks that have been suspended, delisting and less than 2000 trading days. Last, we prime 424 SPICS and 185 CSICS, which account for just about 85% and 60% of the add up number of stocks respectively.

We grab the price data (the highest price, the last price, the opening price, the closing price) and the bulk data of the SPICS from hypertext transfer protocol://finance.chawbacon.com and the data of the CSICS from http://quotes.money.163.com. The acquired data is not polished by ex-dividend/rights, and then we want to process these information reported to the dividend and rights cut proclaimed by registered companies. Because distributed shares, increase shares by transferring, and dividends can cause excessive jump and twisting of stock terms, which will affect the functioning of trading algorithms and back-testing.

Lineament generation

In this paper, we select 44 comparatively recovered-accepted technical indicators with a HF of use as the features, which include trend indicators, the volatility indicators, cash flow indicators, investor scientific discipline indicators and sol on, as shown in support information (S1 Table). These features describe the kinetic change of a line price and mass in the trading day. It is worth noting that the number of study indicators of stocks is heroic, and the comparable indicator can generate many different indicators because of the different parameters. In addition to around common indicators such as commodity channel exponent (CCI) and relative strong suit index (RSI), on that point are some separate indicators much as average true range (ATR), triple exponentially smoothed moving average (TRIX), because these indicators are of big significance for characterizing the movement pattern of stocks.

Data standardisation

Information normalization is an important whole tone in data preprocessing. Normalized data are generally used as inputs to motorcar learning and data mining models. The meaning of Normalisatio is to compress all data to [0,1]. In this means, a larger value of features throne be avoided having a strong influence on the outturn of the model, thus every bit to improve the hardiness of the model. In this clause, we adopt max-min normalization. That is, to each feature x∈R ⁿ , we ingest x* = (x−min(x))/(max(x)−Amoy(x)).

Trading algorithm and its pattern

Learning algorithm

Given a training dataset, D = {(x ₁,y ₁),(x ₂,y ₂),⋯,( x _P ,y _P )}, where x _i = {x _i1,x _i2,⋯,x _{information processing} } is an instance of input; Pis the figure of try out features; y _i = {0,1} is a class label; i = 1,2,3,⋯,N, where N is the taste size. D is a matrix of N*(P+1), where the P+1-th column of D is course label. The task of learning is to construct a learning model based on a presumption training dataset then that the model can classify class labels correctly. In this paper, we wish use the six time-honoured machine models, namely Atomic number 103, SVM, Go-cart, Releasing hormone, BN, XGB and six inscrutable neural networks, namely MLP [25], DBN [26], SAE [27], RNN [28], LSTM [28], and GRU [28] as classifiers to predict the rise and fall of the stock prices. The main model parameters and training parameters of these learning algorithms are shown in the above table.

In Table 1 and Remit 2, features and class labels are set according to the stimulation data format of various motorcar learning algorithms in R language. Matrix (m, n) represents a matrix with m rows and n columns; Array (p, m, n) represents a tensor (namely array in R language), where each layer of the tensor is Matrix (m, n) and the height of the tensor is p. c (h1, h2, h3, …) represents a transmitter, where the length of the vector is the number of hidden layers and the i-th factor of c is the number of neurons of the i-th layers. In the experiment, m = 250 represents that the data of the past 250 days (about 250 trading years in a year) are used every bit preparation samples in each round of paseo-forward psychoanalysis, because we think the model trained with one year's data is enough to predict the day in front; n = 44 represents that the data of each day has 44 features. In Table 2, the activation function of altogether deep neural network models is a line function. Other parameters much as learning plac, tidy sum size, and epoch are all the default values in the algorithm of R programs.

It is worth noting that experimental data is high-dimensional time series data. Previous studies have shown that clock time series information have autocorrelation and time dependencies, so it is different from the assumption of independent and identically distributed information in machine learning model. Therefore, we arrange non divide the data place into training dataset, establishment dataset and quiz dataset in the experiment, because the validation dataset can separate the training dataset and the test dataset, which will cause the dependency between time series to melt. Meanwhile, time serial publication data are not suitable for crossbreeding-validation to optimize parameters because it is logically wrong to use the information after a certain prison term to predict the data before that time. Therefore, we do not use validation dataset to choose hyper-parameters. The hyper-parameters mentioned in the paper much as the number of layers of the deep neural net and the number of neurons in each bed are empirically tuned based on previous experiments. For those insensitive parameters so much as the number of trees in the random forest algorithm and the prior probabilities of class membership in naïve Bayesian algorithm, we use the default parameters preset by R packages.

Walk-forward depth psychology

Walking-Forward Analysis [29] is a systematized and formalized fashion of playing what has been referred to as a rolling optimization or a periodic re-optimization (see Ficus carica 1). Extraordinary of the primary benefits of the walk-forward analysis is to determine the robustness of the trading scheme. Walk-forward analysis is to determine the degree of confidence with which the trader May expect that the strategy will do in real-metre trading.

Another momentous advantage of walk-forward analysis is to produce a better trading performance as markets, trends, and volatility change. Since this periodic re-optimization is done with a scheme-appropriate amount of current monetary value data, this also provides an efficient means to continuously accommodate a trading poser to ongoing changes in grocery conditions.

The algorithmic program for generating trading signals

Therein break u, we use machine learning algorithms as the classifiers to predict the ups and downs of the stocks in each industry of SPICS and CSICS and use the prediction results A the signals of daily trading. We use the walk-first analysis method to train each machine learning algorithmic program gradually. In each step, we use the information from the past 250 days (combined year) as the grooming set and the data for the next 5 years (cardinal week) as the test set. Each stock contains data for 2,000 trading days, indeed it takes (2000–250) / 5 = 350 training sessions to produce a total of 1,750 predictions which are the signals of daily trading. The algorithm for generating trading signals is shown in Algorithmic program 1.

Algorithm 1. Generating Trading Signals in R Language

Stimulant: Commonplace Code List for each industry (SCLEI)

Output: Trading Signals

dannbsp;dannbsp;dannbsp;dannbsp;1. N = length of Unoriginal Code List #424 SPICS, and 185 CSICS. N = 424, 185.

dannbsp;dannbsp;dannbsp;dannbsp;2. L = Phone number of Samples #L = 2000

dannbsp;dannbsp;dannbsp;dannbsp;3. P = Length of Features #P = 44

dannbsp;dannbsp;dannbsp;dannbsp;4. k = distance of Training Dataset #k = 250

dannbsp;dannbsp;dannbsp;dannbsp;5. n = Distance of Testing Dataset/Length of Paseo-Fresh Windowpane #n = 5

dannbsp;dannbsp;dannbsp;dannbsp;6. for i in 1:N

dannbsp;dannbsp;dannbsp;dannbsp;7.dannbsp;dannbsp;dannbsp;dannbsp;Stock Data = SCLEI[i]

dannbsp;dannbsp;dannbsp;dannbsp;8.dannbsp;dannbsp;dannbsp;dannbsp;M = (L-k)/n

dannbsp;dannbsp;dannbsp;dannbsp;9.dannbsp;dannbsp;dannbsp;dannbsp;Trading Signal0 = NULL

dannbsp;dannbsp;dannbsp;dannbsp;10.for j in 1:M

dannbsp;dannbsp;dannbsp;dannbsp;11.dannbsp;dannbsp;dannbsp;dannbsp;dannbsp;dannbsp;dannbsp;dannbsp;New_Data = Descent Data[(k+n*(j-1)):(k+n+n*(j-1)), ]

dannbsp;dannbsp;dannbsp;dannbsp;12.dannbsp;dannbsp;dannbsp;dannbsp;dannbsp;dannbsp;dannbsp;dannbsp;New_Train = New_Data[1:k,]

dannbsp;dannbsp;dannbsp;dannbsp;13.dannbsp;dannbsp;dannbsp;dannbsp;dannbsp;dannbsp;dannbsp;dannbsp;New_Test = New_Data[(k+1): (k+n),1:P]

dannbsp;dannbsp;dannbsp;dannbsp;14.dannbsp;dannbsp;dannbsp;dannbsp;dannbsp;dannbsp;dannbsp;dannbsp;Train_Model = Encyclopaedism Algorithmic program(New_Train)

dannbsp;dannbsp;dannbsp;dannbsp;15.dannbsp;dannbsp;dannbsp;dannbsp;dannbsp;dannbsp;dannbsp;dannbsp;Proba = Train_Model(New_Test)

dannbsp;dannbsp;dannbsp;dannbsp;16.if Probadangt; = 0.5 then

dannbsp;dannbsp;dannbsp;dannbsp;17.dannbsp;dannbsp;dannbsp;dannbsp;dannbsp;dannbsp;dannbsp;dannbsp;dannbsp;dannbsp;dannbsp;dannbsp;Trading Signal0 = 1

dannbsp;dannbsp;dannbsp;dannbsp;18.else

dannbsp;dannbsp;dannbsp;dannbsp;19.dannbsp;dannbsp;dannbsp;dannbsp;dannbsp;dannbsp;dannbsp;dannbsp;dannbsp;dannbsp;dannbsp;dannbsp;Trading Signal0 = 0

dannbsp;dannbsp;dannbsp;dannbsp;20.End if

dannbsp;dannbsp;dannbsp;dannbsp;21.dannbsp;dannbsp;dannbsp;dannbsp;dannbsp;dannbsp;dannbsp;dannbsp;Trading Signal = c(Trading Signal, Trading Signal0)

dannbsp;dannbsp;dannbsp;dannbsp;23.End for

dannbsp;dannbsp;dannbsp;dannbsp;24.dannbsp;dannbsp;dannbsp;dannbsp;return (Trading Signals)

dannbsp;dannbsp;dannbsp;dannbsp;25. End for

Performance evaluation and optimal trading model selection

Performance evaluation indicators

Investment performance is an fundamental tool to evaluate the effectualness of a quantitative trading algorithm. In that paper, we use 12 machine learning algorithms and walk-forward analysis to predict the future trends of stock prices. Then, we use the classification presage results atomic number 3 the trading bespeak to demeanour the backrest-testing and apply the WR, ARR, ASR, and MDD as the indicators of the trading performance evaluation. These indicators reflect the investment ability of investors surgery trading algorithms.

WR is the ratio of the number of days with empiricism earnings to the add numerate of the trading Day. IT is noteworthy that our trading strategies do not allow short sale. So, we cannot trade when our trading algorithms predict that the stock prices will fall. WR is a measure of the truth of trading signals, and a better algorithm for generating trading signals will lead to a higher WR. American Samoa the nigh basic rating indicator, WR can be utilised to evaluate whether the current transaction performance is consistent with the previous one. The go down in the WR may betoken that the trading strategy has reached impregnation.

ARR is a theoretical rate of proceeds, not the substantial yield of investing strategy. It is derived from the average rate of return in the past investment full stop by the annualized calculation process and is non example of future carrying into action. Suppose that the holding period of an investment puppet is H, the return rank of the investment puppet is RR _H , and there are m single periods in uncomparable year. ARR is conferred by the following formula.

This method takes into consideration the continuous three-lobed worry of eightfold periods. In some cases, ARR = m*R _H /H can constitute utilized to calculate ARR. In general, we first calculate the yield of a single catamenia and then cipher the ARR.

ASR is a performance evaluation index organized by Sharpe in 1966 [30]. IT is a risk-keyed return. Suppose that the holding period of an investment instrument is H, and there are m several periods in a twelvemonth. In the H period, the ARR of the investment tool is ARR _H , the standard deviation of return plac is σ _H , and R _f is the benchmark such as risk-free pass. In that paper, we lay R _f = 0. ASR is given as follows.

Drawdown is a measure of historical loss. It is the largest loss compared to the previous highest value (urine level) of the net value curve. Investment managers normally flummox performance fees after their investing returns exceed the water level. MDD shows the largest decline in the price or value of the investment period H, which is an important risk assessment indicator. In the period of investment τ, we first calculate the D _τ at any meter τ≤H. Then, we give the axe drive the MDD _H when we go traverse the whole time interval. where p _t denotes the value of the net esteem curve with time t; D _τ represents the drawdown at the time τ, i.e., the difference betwixt the maximum value in [0,τ] and the value of at the time τ. MDD _H denotes the maximum drawdown in [0,H].

It is noteworthy that we serve non consider transaction costs when calculating these performance evaluation indicators. Stocks may atomic number 4 listed only once in a few days when we implement stock daily trading strategy and short marketing is not allowed. So, transaction costs are fewer and even negligible.

Valuation and depth psychology of trading carrying out for the two datasets

In order to study the significant difference among the application of dissimilar machine learning algorithms in other industries, we divide the industry into 9 categories supported on finance.sina.com.cn, which including Basic Materials (BM), Consumer Cyclical (CC), Communicating (COM), Energy (EN), Finance (FIN), industry (IND), Non-Consumer Cyclical (NCC), Public Inferior (PU), and Technology (TECH) equally shown in supporting information (S2 Postpone). The number of SPICS and CSICS in various industries is shown in Board 3.

In Order to compare whether there are statistically significant differences between the stock trading performance of opposite algorithms in the same industry, we redact forward the following test hypotheses:

For any industry i∈{BM,CC,COM,Nut,IND,NCC,Pu,TECH}, for any performance rating indicator j∈{WR,ARR,ASR,MDD}.The null hypothesis a is Hija, alternative hypotheses b is Hijb.

Hija: in the industry i, the evaluation indicator j of altogether trading strategies are the same;

Hijb: in the manufacture i, the evaluation indicator j of whol trading strategies are not the equal.

Surrendered the significance level is 0.05. We apply cardinal non-parametric applied mathematics test method. Firstly, we use the Kruskal-Wallis rank sum test [31]to carry out the analysis of variance. Second, if the disjunctive hypothesis is established, we need to apply the Nemenyi test [32] to do the multiple comparisons between trading strategies. In this process, we use the index (Sdanampere;P 500 index finger and CSI 300 index) and BAH scheme as the benchmark.

Relation analysis of sensible trading algorithms for each manufacture in SPICS.

In that part, we will analyze whether there are statistically significant differences in the WR, ARR, ASR, and MDD among various algorithms for each industry in SPICS, which can ply guidance for using contrastive trading algorithms in distinct industries.

From Table 4, we can pick up that in whol industries, MLP achieves the highest WR among all algorithms. Through with the multiplex equivalence analysis, we can find that the WR of MLP is not importantly different from that of DBN, SAE, RNN, GRU, LSTM, and SVM in Ordure, merely the WRs of MLP, SAE, and DBN are significantly greater than that of other algorithms. In CC, Phoebe, IND, NCC, Technical school, in that location is no significant difference between MLP, DBN, and SAE for WR, but the WRs of them is significantly greater than other algorithms. In COM, the WR of MLP is not significantly different from that of DBN and SAE, just the WR of MLP is significantly greater than that of unusual algorithms except for SAE and DBN; the WRs of DBN and SAE are not significantly different from that of LSTM, but the WRs of them is significantly greater than other algorithms. In EN, there is no significant difference 'tween the WR of all trading algorithms. In PU, the WR of MLP is not significantly different from that of DBN that SAE, but the WRs of MLP, SAE, and DBN are importantly greater than other algorithms.

From Table 5, we can find out that the ARR of CART is the highest in the BM, COM, EN, and IND; the ARR of DBN is the highest in the Ml and PU; the ARR of MLP is the highest in FIN and NCC; the ARR of SAE is the highest in TECH. Through the multiple equivalence analysis, we derriere encounte that the ARR of Sdanamp;P 500 index finger is not significantly different from that of BAH scheme, but the ARRs of all machine encyclopedism algorithms are significantly greater than that of Sdanamp;P 500 index and BAH in the BM, CC, COM, EN, IND, PU, and Technical school; other, there is none significant dispute between the ARRs of whatsoever deuce algorithms. In the 5, the ARR of Sdanamp;P 500 index is non significantly different from that of BAH strategy, but the ARRs of all machine learning algorithms are significantly greater than that of SdanAMP;P 500index and BAH scheme; the ARR of MLP is significantly greater than that of LR; the ARRs of MLP, DBN, and SAE are importantly greater than that of LSTM; otherwise, there is no world-shaking difference between the ARRs of whatever two algorithms. In NCC, the ARR of the Sdanadenylic acid;P 500 index is importantly frown than that of BAH scheme; the ARRs of wholly machine learning algorithms are significantly greater than that of Sdanadenosine monophosphate;P 500 forefinger and BAH scheme; otherwise, there is nobelium significant difference between the ARRs of any 2 algorithms.

From Table 6, we can see that the ASR of XGB is the highest in the BM and COM; the ARR of SAE is the highest in the EN; the ARR of RF is the highest in the CC, FIN, IND, NCC, PU, and Technical school. Through the multiple comparison analysis, we can find the ASR of Sdanampere;P 500 index is not importantly different from that of BAH strategy, but the ASRs of all trading algorithms are significantly greater than that of Sdanamp;P500 index and BAH strategy in altogether industries except EN. Otherwise, in that location is no world-shattering difference between the ASRs of any two algorithms in the Stool, COM, PU, and TECH. In the CC, the ASR of Lr is significantly greater than that of CART; other than, there is no evidential difference between the ASRs of any two algorithms. In the EN, the ASR of SdanA;P 500 index is non significantly different from that of RNN, LSTM, GRU, CART, NB, RF, LR, SVM, and XGB, but the ASRs of BAH strategy, MLP, DBN, and SAE is significantly greater than that of Sdanamp;P 500 index; the ASRs of wholly trading algorithms are significantly greater than that of BAH strategy; otherwise, on that point is atomic number 102 significant difference between the ASRs of any two algorithms. In the FIN, the ASRs of all trading algorithms are significantly greater than that of Handcart; other, there is no significant difference between the ASRs of any two algorithms. In the IND, the ASR of RF is significantly greater than that of DBN; the ASR of CART is significantly lower than that of NB, RF, SVM, XGB; otherwise, at that place is no portentous departure between the ASRs of any two algorithms. In the NCC, the ASR of RF is significantly greater than that of MLP, DBN, SAE; the ASR of CART is importantly let down than that of RF, SVM, XGB; differently, there is no significant difference between the ASRs of any two algorithms.

From Table 7, we toilet realise that the MDD of Sdanamp;P 500 index is the lowest and the MDD of BAH is highest in all trading strategies including all machine learning algorithms and BAH strategy in each industry. Through the multiple comparison analysis, we can find that the MDDs of all machine learnedness algorithms and BAH strategy are significantly greater than that of the Sdanadenylic acid;P 500 index in every last industries except PU. Otherwise, there is no operative difference between the MDDs of any two algorithms in the BM, COM, and TECH. In the CC, the MDD of BAH is significantly greater than that of RF and XGB; otherwise, there is No significant difference between the MDDs of any two algorithms. In the EN, the MDD of BAH is significantly greater than that of MLP, DBN, and SAE; other than, there is no significant difference between the MDDs of any cardinal algorithms. In the FIN, the MDD of BAH is significantly greater than that of RNN, LSTM, GRU, SVM, XGB, and Atomic number 10; differently, there is no significant difference between the MDDs of any two algorithms. In the IND, the MDD of BAH is importantly greater than that of GRU, CART, Atomic number 103, NB, SVM, XGB, and RF; otherwise, at that place is nobelium significant conflict between the MDDs of any two algorithms. In the NCC, the MDD of BAH is significantly greater than that of RNN, GRU, CART, NB, RF, and XGB; otherwise, there is nary monumental difference between the MDDs of any two algorithms. In the PU, there is no significant difference between the MDDs of all trading strategies including BAH strategy and Sdanamp;P500 index.

Relation analytic thinking of intelligent trading algorithms for all industry in CSICS.

In the CSICS, we still use the analysis method mentioned above. We obtain the best trading algorithm which can be suitable for the stock trading of the given industry by comparing the carrying into action of different algorithms. This can provide some guidance for the formulation of an investment strategy.

We can see from Table 8 that the WR of MLP is the highest in all industry. In the BM, Millilitre, COM, FIN, IND, NCC, PU, and TECH, the WR of MLP is not a significant deviation from that of DBN, SAE, and SVM through ten-fold comparison psychoanalysis, but the WRs of MLP, DBN, and SAE are importantly higher than that of other algorithms. In the EN, the WR of MLP is non significantly different from that of DBN, SAE, RNN, SVM, and NB, but the WRs of MLP, DBN, and SAE are significantly higher than that of other algorithms. Thence, MLP, DBN, and SAE perform well in all industries.

We fire see from Put of 9 that the ARRs of the CSI 300 index and BAH strategy are to a lesser degree that of all other trading algorithms in each industry. The ARR of Nb is the highest all told industries except for the Technical school, and the ARR of MLP is the highest in the Technical school. Through with the analysis of variance and multiple comparative analysis, the ARRs of all trading algorithms are significantly higher than that of CSI 300 index and BAH strategy. All told the mentioned automobile learning algorithms in this paper, although the ARRs of just about algorithms look very major than that of other algorithms, in that location is no significant departure between the ARRs of them in all industries except the IND. It is worthy noting that the ARR of RF is significantly lower than that of other algorithms in the IND, but there is nary significant difference between other algorithms. Equally far as ARR is concerned, the traditional machine erudition algorithms are not worse than that of all the algorithms supported on the deep neural network in just about industries.

We can see from Remit 10, the ASRs of CSI 300 index and BAH scheme are lower than that of all machine learning algorithms. In BM, COM, FIN, and NCC, the ASR of LR is the highest; in CC, IND, and TECH, the ASR of LSTM is the highest; in Nut and Plutonium, the ASR of GRU is the highest. Done the analysis of variance and multiple relative analysis, the ASRs of CSI 300 index and BAH strategy are significantly depress than that of all former machine learning algorithms. In the CC, EN, Pu, TECH, there is no more significant difference betwixt the ASRs of all algorithms. In the BM, the ASR of N.B. is importantly turn down than that of LR and GRU, otherwise, there is no world-shattering difference between whatsoever two algorithms. In the COM, the ASR of NB is significantly glower than that of LR, otherwise, in that respect is no significant difference between any two algorithms. In the FIN, the ASR of Atomic number 41 is significantly lower than that of all other algorithms; the ASR of CART is significantly lower than that of all other algorithms omit MLP, DBN, and SAE; otherwise, on that point is no significant difference between any two algorithms. In the IND, the ASR of NB is importantly glower than that of all other algorithms; the ASR of CART is significantly lower than that of LSTM; otherwise, there is nary significant difference between any two algorithms. In the NCC, the ASR of CART is importantly lower than that of LR; otherwise, in that respect is no significant difference between any two algorithms. As Army for the Liberation of Rwanda as ASR is solicitous, the NB and CART are non the ideal choices. It is worth noting that the traditional machine encyclopaedism algorithms are non worse than the popular algorithms based on the deep neural network in some industries.

We can run into from the Table 11, the MDD of CSI 300 index is lower than that of all machine acquisition algorithms and BAH strategy in completely industries except the Mil and NCC; in the CC, the MDD of Reticular formation is the last; in the NCC, the MDD of LR is the lowest. Through analysis of variance and quaternate comparative analysis, the MDD of the CSI 300 index is not significantly frown than that of LR and GRU in the Fecal matter, but significantly turn down than that of other algorithms and BAH strategy; the MDD of the BAH strategy is significantly higher than that of Lawrencium and GRU, just there is no significant difference between BAH and other algorithms; the MDD of NB is importantly higher than that of LSTM, GRU, CART, LR, and XGB; other, there is no meaning departure between any two algorithms. In the Ml, the MDD of the CSI 300 index is significantly lower than that of NB and BAH strategy, merely there is zero significant deviation between CSI 300 index and some other algorithms; the MDD of the BAH strategy is significantly higher than that of RNN, LSTM, GRU, and RF, but there is atomic number 102 significant difference 'tween BAH strategy and other algorithms; the MDD of the NB is significantly higher than that of RNN, LSTM, GRU, CART, SVM, RF, and XGB; otherwise thither is nary significant difference 'tween some two algorithms. In the COM, the MDD of the CSI 300 index is importantly lower than that of Atomic number 41 and BAH strategy, but there is no more significant difference between CSI 300 index and other algorithms; there is no significant dispute between BAH and other algorithms; the MDD of the N.B. is importantly higher than that of LR and XGB; otherwise, in that respect is no significant difference between whatever deuce algorithms. In the EN, the MDD of the CSI 300 index is significantly let down than that of NB and BAH, but there is no significant difference between CSI 300 indicant and other algorithms; at that place is atomic number 102 significant difference between BAH and other algorithms; the MDD of NB is significantly high than that of LSTM, GRU, RF, and Lawrencium; otherwise, on that point is nary significant difference between any two algorithms. In the Little Phoeb, the MDD of the CSI 300 index is significantly take down than that of MLP, DBN, SAE, NB, and BAH, but there is nobelium significant difference between CSI 300 index and other algorithms; the MDD of the BAH strategy is significantly higher than that of RNN, LSTM, GRU, Go-cart, LR, SVM, and RF, but thither is no significant difference 'tween BAH and other algorithms; the MDD of the NB is significantly higher than that of all other algorithms; otherwise, there is no significant difference betwixt any two algorithms. In the IND, the MDD of the CSI 300 index is significantly lower than that of complete other algorithms and BAH scheme; the MDD of the BAH scheme is significantly higher than that of RNN, LSTM, GRU, CART, LR, and XGB, but there is no significant difference between BAH scheme and other algorithms; the MDD of the NB is significantly high than that of all unusual algorithms; otherwise, there is no key difference between any two algorithms. In the NCC, in that location is no significant difference 'tween the MDDs of CSI 300 index, BAH strategy, and other algorithms. In the PU, there is no significant difference 'tween the MDDs of CSI 300 index, BAH and strange algorithms except Niobium, and the MDD of the NB is importantly higher than that of CSI 300 index, BAH, and other algorithms; otherwise, there is no significant difference between any cardinal algorithms. In the Technical school, the MDD of the CSI 300 index is significantly depress than that of MLP, DBN, SAE, NB and BAH strategy, but there is no world-shaking deviation between CSI 300 forefinger and other algorithms; the MDD of the DBN is significantly higher than that of LSTM; otherwise there is no significant difference between any two algorithms.

Selection of the best trading model for different industries

Next, we give the optimal trading algorithms (TOTAs) for stock trading of apiece industriousness supported the analysis results of the above. We give a series of rules as follows, where "adangt;b" represents that the carrying into action of algorithm a is importantly greater than that of algorithm b; "a = b" represents that the execution of algorithmic program a is no significantly diverse from that of algorithmic program b. (a = b)∧(bdangt;d) represents that "a = b"and "bdangt;d"are simultaneously ingrained.

For any industry i∈{Movement,CC,COM,EN,FIN,IND,NCC,PU,TECH}, for some public presentation evaluation indicator j∈{WR,ARR,ASR,−MDD}. We totally know that the greater the value of WR,ARR,ASR and −MDD, the better the trading carrying out of the strategy or trading models. The relationship of the performance of all strategies including machine learning algorithms, BAH strategy, and bench mark index finger can be denotive by the human relationship among the 3 strategies, which are expressed Eastern Samoa a,b, and c respectively.

Dominate 1. Sifting Regulation Depending on Single Index number

• If (adangt;b)∧(bdangt;c)∧(adangt;c), then the strategy a is the optimum altogether strategies;
• If (adangt;b)∧(bdangt;c)∧(a = c), then the strategy a is the optimal in all strategies;
• If (adangt;b)∧(b = c)∧(a = c), then the strategy a is the optimal in completely strategies;
• If (adangt;b)∧(b = c)∧(adangt;c), then the strategy a is the optimal in each strategies;
• If (a = b)∧(bdangt;c)∧(adangt;c), then the strategy a and b are the optimal in all strategies;
• If (a = b)∧(b = c)∧(a = c), and so the scheme a,b, and c are the optimal in all strategies.

Firstly, we want to choose the optimal machine scholarship algorithms for each industry in which the trading performance of the algorithm can atomic number 4 significantly better than that of the bench mark index, that is, algorithm trading strategy can beat the market; secondly, the trading performance of the optimum machine learning algorithm pot cost significantly better than the BAH strategy in each manufacture, which is conducive to take an active amount investment strategy for neckcloth trading while reducing risk. Therefore, if the trading performance of auto learning algorithms is not best than that of the index, we hope that it is importantly better than BAH scheme. Otherwise, the machine learning algorithmic program will non make gumption for stock trading. We select the best trading algorithms(TOTAs) which are significantly better than the rest of the algorithms, as shown in Set back 12.

From Table 12, we find that the optimal trading model based on the WR is always found in any industry, and MLP is the optimal algorithms altogether industries through the analysis of the industries in the SPICS; MLP is the optimal trading model supported on ARR in the Phoebe, and whatever algorithms can be used in different industries; the optimum trading model based on ASR seat be institute in the Milliliter, Nut, Quintet, IND, and NCC, and any algorithms can be used in other industries; the optimal trading model based on MDD can be ground in the Mil, EN, Little Phoeb, IND, and NCC, and whatever algorithms can be used in other industries. Through the analysis of the industries in the CSICS, the optimum worthy ass constitute institute in all industries supported on the WR, and SAE, MLP, and DBN are the best trading models in well-nig industries; TOTAs settled on ARR nates be found in the Phoebe and IND, and any algorithms can be put-upon in other industries; TOTAs based on ASR ass be found in the BM, COM, Quintet, IND, and NCC, and any algorithms can be used in separate industries; TOTAs supported MDD buns make up plant in all industries except NCC. From Remit 12, we can find that there is more than incomparable optimal trading algorithm in some industries, which is normal. As a matter of fact, there is no significant difference in performance among the multiple optimal trading algorithms selected. For example, for the industry BM in SPICS, we obtain the optimal trading algorithms which including MLP, DBN, and SAE supported WR. These tercet algorithms have No statistically significant difference for WR.

However, we can find from Postpone 12 that there are too more "ATAUs", which means that the optimal trading models planned for each industry are still non refined enough, so we propose a newborn set of rules based on Table 12 to opinionated the selection range of the optimal models. For each industry, ASR represents risk-familiarized returns, it is the most weighty indicator for evaluating a trading algorithm; secondly, ARR represents the return of a stock during a holding period, so ARR is also an important indicator for evaluating the trading algorithm without considering risk; thirdly, MDD describes the potential risks of trading algorithms which are practical to stock trading; finally, WR represents the carrying into action of a trading algorithmic program in predicting stock Leontyne Price trends, which is not a direct source of stock investment returns. Thence, we arrogate that ASR≻ARR≻MDD≻WR accordant to the importance of the four rating indicators, where "m≻n" represents the index number m is more important than the indicant n. The following refining rules are proposed.

Rule 2. Refining Rules Depending on Comprehensive Indicators

If ASR∩ARR∩MDD∩WR≠∅, then TOTAs = ASR∩ARR∩MDD∩WR;

else if ASR∩ARR∩MDD∩WR≠∅ and ASR∩ARR∩MDD≠∅, then TOTAs = ASR∩ARR∩MDD;

else if ASR∩ARR∩MDD≠∅ and ASR∩ARR≠∅, then TOTAs = ASR∩ARR;

other ASR∩ARR≠∅, so TOTAs = ASR.

For model, we can use the above rules to select TOTAs for NCC in SPICS: WR = {MLP,DBN,SAE}, ARR = ATAU = {MLP,DBN,SAE,RNN,GRU,LSTM,NB,SVM,XGB,LR,RF,Pushcart}, ASR = {RF}, MDD = {RNN,GRU,CART,NB,RF,XGB}, we have ASR∩ARR∩MDD∩WR≠∅ and ASR∩ARR∩MDD = Releasing hormone≠∅, thus the RF is the TOTA for NCC in SPICS. We obtain TOTAs for for each one industry in the SPICS and CSICS, as shown in Table 13.

As can be seen from Table 13, the number of optimal trading models selected according to Prescript 2 is elflike because Ruler 2 takes into account the importance between the ASR and the remaining indicators. The transaction models selected are more in working order. At the equal clock time, deep neural network algorithms have a good operation in most industries, simply LR and Element 104 are same prominent in some industries.

These empirical results prove that on SPICS and CSICS, we rump always select TOTAs supported the single index number and comprehensive indicators in all industries. We dismiss apply TOTAs to implement trading action in for each one manufacture of China A-share securities industry and American market.

Conclusion

In that paper, we adopt the 424 SPICS in the US market and the 185 CSICS in Republic of China market from 9 industries arsenic the research object. Then, for each stock in every industry, we pick out the information of the 2000 trading days before December 31, 2022, and build 44 technological indicators As the input features of the machine learning algorithm to predict the trends of the stock Mary Leontyne Pric. Then, we formulate trading strategies supported the trading signals, canva and measure the performance of these algorithms in different industries. Eventually, we use a set of rules to select TOTAs for stock trading in each industry. The experiment shows that on SPICS and CSICCS, we can blue-ribbon at least one of the C. H. Best trading models for each industry supported on the lonesome indicator and comprehensive indicators. The optimal trading models perform well for WR in all industries; the ARR and ASR of the model can be significantly better than that of the benchmark index and BAH strategy in most industries; the MDD of the worthy can be significantly less than that of BAH strategy in most industries. Thus, the algorithms tail end be applied to the stock investing in most industries, and it is a precise significant effect on investment yield and risk management.

In view of the rapid development of artificial intelligence engineering and the easy access to financial big data, the future research work can be carried out from the following aspects: (1) using the deep neural web to carry out propelling portfolio direction among different assets; (2) using the inexplicable nervous network to simulate high-frequency trading and rise strategies. The solution for these problems will help to develop an later and profitable machine-driven trading system supported financial big data, which including high-voltage portfolio construction, optimal execution, and risk direction according to the changes in market conditions.

Supporting data

References

1. 1. Huang W, Nakamori Y, Wang SY. Foretelling stock market apparent movement direction with support vector machine. Computers danadenosine monophosphate; Trading operations Research. 2005; 32: 2513–2522.
   • Survey Article
   • Google Scholar
2. 2. Chen JX. SVM application of financial metre series forecasting using empirical technical indicators. In World-wide conference on data, networking and automation; 2010.
3. 3. Xie CQ. The optimisation of share price prevision model based on support vector machine. In International conference along control, mechanisation and systems technology; 2011.
4. 4. Ładyżyński P, Żbikowski K, Grzegorzewski P. Stock Trading with Random Forests, Slue Espial Tests and Force Index Volume Indicators. International Group discussion connected AI and Soft Computation; 2022.
5. 5. Zhang J, Cui SC, Xu Y, Lithium QM, Li T. A novel data-driven breed terms trend prediction system. Adept Systems with Applications. 2022; 97:60–69.
   • View Article
   • Google Scholar
6. 6. Ruta D. Automated Trading with Machine Learning on big data. IEEE International Congress on Big Information; 2022.
7. 7. Patel J. Predicting stock and stock price level movement victimization Drift Deterministic Data Preparation and machine learning techniques. Expert Systems with Applications. 2022; 42:259–268.
   • Take i Clause
   • Google Scholar
8. 8. Luo LK, Chen X. Desegregation piecewise linear representation and weighted support vector automobile for stock trading signal prediction. Practical Soft Computing. 2022; 13(2):806–816.
   • View Article
   • Google Scholar
9. 9. Zbikowski K. Using Volume Heavy Support Vector Machines with walk forward testing and feature selection for the purpose of creating stock trading scheme. Expert Systems with Applications programme. 2022; 42:1797–1805.
   • Opinion Article
   • Google Scholar
10. 10. Dash R, Hyphen PK. A hybrid stock trading framework integration technological analysis with machine learning techniques. The Journal of Finance and Data Science. 2022; 2:42–57.
    • Scene Clause
    • Google Scholar
11. 11. Bao W, Yue J, Rao YL. A deep learning framework for financial time serial publication exploitation built autoencoders and prolonged unawares terminal figure store. Plos Same. 2022; 12(7):1–24.
    • View Clause
    • Google Scholar
12. 12. Thomas F, Chrisstopher K. Deep learning with long short-term memory networks for financial commercialize predictions. Fau Discussion Papers in Economics. 2022; 270(2):1–32.
    • View Article
    • Google Scholar
13. 13. Makickiene N, Rutkauskas AV, Maknickas A. Investigation of financial market prediction by recurrent neural network. Innovative Info Technologies for Science, Business and Education. 2011; 2(11):1–24.
    • View Article
    • Google Scholar
14. 14. Persio LD. Recurrent neural networks approach to the business enterprise foretell of Google assets. International Journal of Mathematics and Computers in Simulation. 2022; 11:1–7.
    • View Clause
    • Google Scholarly person
15. 15. Dunis CL, Pentateuch J, Evanset B. Trading futures spread portfolios: applications of high set up and recurrent networks. The Continent Journal of Finance. 2008; 14(6):503–521.
    • Reckon Clause
    • Google Scholarly person
16. 16. Chong E, Han C, Park FC. Rich scholarship networks for securities market analysis and prediction: Methodology, information representations, and incase studies. Expert Systems with Applications. 2022; 83:187–205.
    • View Article
    • Google Scholar
17. 17. Krauss C, Do A, Hucket N. Bass neural networks, gradient-boosted trees, haphazard forests: Statistical arbitrage on the Sdanampere;P500. European Journal of Operative Research. 2022; 259: 689–702.
    • View Article
    • Google Scholar
18. 18. Hsieh TJ, Hsiao H, Yeh W. Forecasting stock markets exploitation wavelet transforms and recurrent neural networks: An Integrated system based along counterfeit bee settlement algorithm. Applied Soft Computing. 2011;11(2):2510–2525.
    • Reckon Article
    • Google Scholar
19. 19. Längkvist M, Karlsson L, Loutfiet A. A retrospect of unsupervised lineament learning and deep learning for time-series modeling. Shape Recognition Letters. 2022;42:11–24.
    • Perspective Article
    • Google Scholar
20. 20. Liu Weber, Wang ZD, Liu XH, Zeng NY, Liu YR, Alsaadi DE. A survey of deep neural net architectures and their applications. Neurocomputing. 2022; 234:11–26.
    • Take i Clause
    • Google Learner
21. 21. Dixon M. Sequence assortment of the limitation order Bible victimisation recurrent neuronic networks. Journal of Computational Skill. 2022; 24:277–286.
    • View Clause
    • Google Scholar
22. 22. Kim HY, Won CH. Forecasting the unpredictability of stock price index: A hybrid model integrating LSTM with multiple GARCH-type models. Expert Systems with Application. 2022; 103: 25–37.
    • View Article
    • Google Scholar
23. 23. Shen GZ, Burn QP, Zhang HY, Zeng P, Xu JJ. Deep Learning with Gated Recurrent Unit Networks for Financial Sequence Predictions. Procedia Computing. 2022;131:895–903.
    • Catch Article
    • Google Scholar
24. 24. Sezer OB, Ozbayoglu M, Dogdu E. A Bottomless Neuronal-Network Based Stock Trading System Based on Organic process Optimized Technical Analysis Parameters. Procedia Reckoner Science. 2022; 114:473–480.
    • View Article
    • Google Scholar
25. 25. Murphy KP. Machine Learning: A Probabilistic Perspective. London: The MIT Urge; 2012.
26. 26. Hinton GE, Osindero S, Teh Y. A high-velocity acquisition algorithm for deep belief nets. Neural Computation. 2006; 18:1527–1554. pmid:16764513
    • View Clause
    • PubMed/NCBI
    • Google Scholar
27. 27. Vincent P, Larochelle H, Lajoie I, Bengio Y, Manzagol PA. Stacked Denoising Autoencoders: Erudition Useful Representations in a Low-pitched Electronic network with a Local Denoising Criterion. Journal of Machine Learning Research. 2010; 11:3371–3408.
    • View Article
    • Google Learner
28. 28. Goodfellow I, Bengio Y, Courvilleet Y. Deep Learning. Cambridge: The MIT Press; 2022.
29. 29. Pardo R. The Evaluation and Optimization of Trading Strategies. 2nd ED. Hoboken: Trick Wiley danamp; Sons; 2008.
30. 30. Aldridge I. High-Frequency Trading: A Practical Guide to Algorithmic Strategies and Trading Systems. 2nd ed. Hoboken: John Wiley danampere; Sons; 2022.
31. 31. Dutchman M and Wolfe DA. Nonparametric Statistical Methods. Hoboken: John Wiley danAMP; Sons; 1973.
32. 32. Nemenyi PB. Distribution-free multiple comparisons. Doctoral dissertations, State University of New York.1963.

optimal trading dynamic stock liquidation strategies

Source: https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0212137

Posted by: simmonssursen.blogspot.com

Share this post