Review of Recent Research Directions and Practical Implementation of Low-Frequency Algorithmic Trading

ABSTRACT


INTRODUCTION
As in most life aspects, technology has tremendously advanced financial systems.It includes many financial systems such as credit business, real estate, insurance, and financial markets.This advancement motivates more quantitative financial mathematics, financial engineering, actuarial science, and risk management.This paper focuses on the evolution of trading in financial markets.Overseas business growth during the industrial revolution at the beginning of the 17 th inspired joint-stock companies and the Dutch East India Co. to issue the first paper shares.The paper shares make it very convenient to transfer the stocks' ownership, increasing the issue of paper shares rapidly.The place where the buyers and sellers gathered to trade the paper shares is called the stock exchange, and the first established exchange was the Amsterdam Stock Exchange (Braudel & Reynolds, 1983).The worldwide exchanges continued until the 90s of the previous century when it shifted to electronic trading (Johnson, 2014).The shift quickly increases the trading volume.However, with the increase of computational power and cloud service availability, the middle of the first decade of this century promoted computers to perform trading on behalf of individuals.It allows for automated trading to be achieved through a finite sequence of steps algorithms.As of 2020, 80% of the trading volume is effectuated through algorithms, and most hedge funds use algorithms to set up their trading strategies.H. Simon (Simon, 1955) prevents declare the bounded rationality the human from making rational decisions due to human emotions, the mind's cognitive limitations, and time availability.
Algorithmic trading (AT) allows for reducing the limitation on rationality.Algorithmic tradings have advantages over discretionary trading by removing emotions and coming up with consistent decisions.In addition, it can monitor the market all the time and implement back testing to ensure the strategy's effectiveness.However, the algorithmic trading results depend on the quality of the developed model and its ability to capture the right signals.In other words, the algorithmic is superior to discretionary trading only if the algorithm itself besteads the discretionary traders.However, the main advantage of AT, according to Johnson (Simon, 1955), are its ability to minimize the effect of emotions, back testing, maintain discipline and consistency, improve placing order speed, and diversification.How-ever, he stated that the main disadvantages of AT centers with the possibility of expense increase and machine failure.The area of algorithmic trading requires multidisciplinary knowledge and skills in finance, mathematics, engineering, and programming.Algorithmic trading is usually executed through automated trading, Robot trading, and black box (Kissell, 2013).The paper proceeds as follows: in Section 2, we defined and characterized algorithmic trading.Section 3 surveys the programming languages and platforms to research algorithmic trading.Section 4 demonstrates the domains of algorithmic trading.Section 5 shows the matrix of performance measures for backtesting.Section 6 explores the main methods used to develop an algorithm for trading.Section 7 survey-related search on various domains.Finally, Section?Provides the conclusion remarks.

LITERATURE REVIEW
The definition of algorithmic trading (AT) and highfrequency trading (HFT) varies among researchers.Jarnecic et al. (Jarnecic & Snape, 2010) define AT as computer algorithms executing predetermined trading decisions to minimize price impact.Domowitz (Domowitz & Yegerman, 2005) characterizes it as automated equity order execution via direct marketaccess channels.Hendershott et al. (Hendershott et al., 2011) describe AT as using algorithms for automatic trading decisions, order submissions, and management.HFT, a primary type of AT, relies on speed for profits.Jarnecic et al. (Jarnecic & Snape, 2010) define HFT as high-speed algorithms generating and executing trades for capital returns.Cvitani et al. (Cvitanic & Kirilenko, 2010) define it as rapid, automated programs creating, directing, and executing orders in electronic markets, engaging in substantial order submissions and cancellations.Gomber et al. (Gomber & Haferkorn, 2015) highlight typical AT and HFT characteristics involving pre-designated decisions, live market data observation, and automated order submission and management.However, HFT differs with numerous orders and cancellations, profiting as a middleman and holding assets briefly.The development of HFT is chiefly by financial institutions, emphasizing algorithmic trading's researcher development and implementation for retail investors.Choosing between buying or building trading software presents trade-offs.Johnson (Johnson, 2020) notes that buying existing software offers easy implementation and customization but can be costly and potentially contain loopholes.Building software, although time-consuming, offers control and customization.Numerous references like "Learn Algorithmic Trading" (Donadio & Ghosh, 2019), "Hands-On Machine Learning for Algorithmic Trading" (Jansen, 2018), "Trading Evolved" (Clenow, 2019), and "Algorithmic Trading" (Johnson, 2020) provide valuable hands-on experience in developing trading systems.Open-source trading platforms like Quantopian, Quant-Connect, and Quant-Insti provide cloud-based services for algorithm development, backtesting, and live trading (Cohan; QuantConnec Profile, 2011; Oberoi).Algorithmic strategies' domains are crucial, as strategies may perform differently based on financial instrument types or trading environments.Derivatives like forwards, futures, swaps, and options can impact trading strategy effectiveness (Hull, 2003).Back-testing using historical data is vital to evaluate algorithm performance.Various performance measures such as portfolio diversification, concentration, and risk-return ratios (Markowitz, 1952) help assess algorithm reliability and effectiveness.

Materials and Methods
Financial markets employ technical analysis, a tool reliant on historical prices to predict market patterns and facilitate trading decisions.Originating from Charles Dow's Dow Theory in 1900 (Achelis), it focuses on interpreting price movements through charts.The analysis mainly revolves around momentum and mean reversion strategies.Momentum strategies advocate following existing trends, assuming their continuation, whereas mean reversion anticipates securities returning to their average prices.Various technical indicators, such as moving averages (MA), exponential moving averages (EMA), and Bollinger Bands (BB), aid in analyzing market trends.For example, the Double Exponential Moving Average (DEMA) utilizes two EMAs for mean reversion signals, while Bollinger Bands offer confidence intervals depicting potential overbought or oversold situations for both strategies.However, technical analysis lacks adaptiveness and learning capabilities.Integrating modern algorithms like machine learning enhances pattern detection and prediction capabilities.Contrarily, fundamental analysis estimates assets' intrinsic value based on financial statements and economic factors, evading the Efficient Market Hypothesis (Fama, 1970).Techniques include financial ratios, discounted dividends, or free cash flow models (Graham & Dodd, 2008).Financial time series analysis evaluates and predicts security prices over time, including linear (AR, MA, ARMA, ARIMA) and nonlinear models (Threshold AR, Markov Switching AR).Volatility prediction models like ARCH and GARCH forecast variations in asset returns due to changing volatility.Moreover, computational mathematics, involving AI, machine learning, and data mining, supports these analyses (Overby, 2011;McMillan, 2002;Kastenholz, 2019).Decision trees (e.g., CART, C4.5) merge fundamental analysis with decision-making, offering actionable rules for stock actions (Rokach, 2014;Larose & Larose, 2014).These methods complement each other, enhancing market understanding and trading strategies (Box et al., 2011;Tsay, 2005;Tsay, 2013).Furthermore, the evolution of algorithmic trading is significantly influenced by advancements in neural networks, particularly Long Short-Term Memory (LSTM) networks introduced by Hochreiter and Schmidhuber (Hochreiter & Schmidhuber, 1997).LSTM, a form of recurrent neural network (RNN), features context neurons representing short-term memory dynamically updating during a time sequence.Differing from feedforward neural networks, RNNs transmit output from context nodes back to hidden layers, involving input and forget gates, and output gate mechanisms.Weight optimization in LSTM employs backward training methods like back-propagation, resilient-propagation, and genetic algorithms.Reinforcement learning, another AI branch, aims to maximize reward through iterative actions based on observed states, as described by Q learning.Studies explore AI-driven trading strategies, from pair-switching approaches (Maewal & Bock, 2011) and technical indicator-based decision support systems (Dash & Dash, 2016;Henrique et al., 2018) to employing machine learning like deep learning neural networks (Gao & Chai, 2018;Yu & Yan, 2020) and sentiment analysis (Bernile & Lyandres, 2011;Putra & Kosala, 2011).Various other strategies underline the diversity and complexity of algorithmic trading models, ranging from leveraging historical market data to predicting stock trends and exploiting market anomalies (Cohen et al., 2010;Gerlein et al., 2016;Kishore et al., 2008;Nair et al., 2010).

RESULTS AND DISCUSSION
On stock options weekly or based on options covered by the underlying assets of stocks with monthly data.Table 1 shows the most common domains of algorithmic strategies.They analyze the algorithms with k = 10 before and after transaction costs on the stocks of the S&P 500 for the period from Dec 1989 to Oct 2015 using various performance measures.Table 4 shows the comparative results after transaction costs obtained by (Krauss et al., 2017).
In a similar design, Fisher and Krauss (Fischer & Krauss, 2018) utilized the LSTM type of deep learning networks to compare the results of LSTM with a set of benchmarked memoryless models such as RF, deep neural network (DNN), and logistic regression (LOG).

Figure 5 :Figure 6 :
Figure 5: Return on equity from 2005 to 2020 for AAPL and MSFT

Figure
Figure7shows the profits of the options strategies over a variety of possible prices at maturity.TIn this paper, we do not aim to define, classify, differentiate or illustrate the methods in these areas, but alternatively, we focus on their application in trading and explore some of the standard methods used to develop algorithms for trading.However, machine learning and data mining algorithms aim to solve estimations,

Figure 7 :Figure 8 :
Figure 7: Possible Profits of various option strategies

Figure 9
Figure9shows an example of simplified decision tree rules.The rules are achieved through constructing a path for the tree, starting the root node to leaves through the branches of decision nodes.The classification and regression algorithm (CART) is a classical decision tree algorithm.CART partitions the tree in a binary manner by splitting the tree into two branches at each decision

Figure 9 :
Figure 9: Example of decision tree rules

Table 1 :
Common domains of algorithmic strategiesTrend indicators attempt to detect a trend in the prices of the assets.Calculating the moving average is a common procedure to identify the up or down trends by smoothing the prices.The momentum See section 6.2 indicators estimate the speed in the changes of prices in a given time-space.The volatility indicators focus on the trading activities, possible range, and security risk.Finally, the volume indicators measure the attraction of financial assets.Table2shows a comprehensive classification of the technical indicators.

Table 2 :
Comprehensive classification of technical indicators

Table 3 :
Financial ratios High value implies high expectations Market price per share/Book value per share

Table 5 :
(Fischer & Krauss, 2018)tained by Fisher and Krauss(Fischer & Krauss, 2018) three strategies as a single trading strategy.The portfolio is balanced monthly, and the results are compared to a benchmark of global bonds and stocks.The results exhibit a superior performance of the composite strategy, as shown in table 6.The model of Fernandez et al.(Fernandez-Perez et   (Zaremba et al., 2019)analyze 15 commodity factors from 1986 to 2017 to find if the momentum effect exists.The results confirm the assumption that buying a commodity with the highest past returns or selling a commodity with the lowest past returns supplement a significant profit.

Table 7 :
Research considerations Objective Prediction (