Algorithmic Trading

Presented to the S. Daniel Abraham Honors Program

Stern College for Women

Yeshiva University

May 19, 2022

Talia Leitner
Mentor: Professor Lawrence Teitelman, Computer Science
Abstract:

This paper gives an overview of algorithmic trading. It is intended to inform interested traders of

the value of incorporating algorithms into their trading strategies. They must also be warned of

the dangers of implementing faulty algorithms. This paper highlights the usefulness of

categorizing algorithms. An overview of the most popular algorithms is also given. Methods for

comparing the performance of different algorithms in trading specific stocks is discussed. The

metrics used to evaluate their performance is also mentioned. Challenges within the field of

algorithmic trading are important for traders to consider when choosing which algorithms to

implement in specific situations. Finally, the integration of machine learning into the field of

algorithmic trading is discussed. This revelation has significantly broadened the scope of

algorithmic trading. Machine learning now serves as an essential tool for traders to compete in

the fast-paced markets of modern day.

2
Introduction:

Technology is integrated into all areas of life. The mid 1970s marked the beginning of the

computerization of trading with the introduction of the designated order turnaround (DOT)

system. This revolutionized the world of finance from the floor-based trading that existed in the

New York Stock Exchange (NYSE) to the electronic trading that is now widely available. High

frequency trading dominates the market and is the practice of trading large volumes of stocks in

infinitesimal periods of time. Electronic trading enables computer systems to control decision

making. Many of these programs are automated and do not require human input.1 Trading

algorithms have been developed to trade stocks at the most profitable prices and at the fastest

rates. Machine learning algorithms have further progressed the financial industry by

incorporating the insights of more data than ever before. These highly complex models result in

more accurate market predictions. With the potential to earn sizable profits comes the risk of

great loss. It is extremely important that traders are well versed in the different algorithms,

performance metrics, and challenges that exist in the field. Understanding all these components

will permit traders to have confidence in their trading algorithms.

An algorithm consists of a series of instructions intended for problem solving.

Algorithms are typically associated with computer programs because they are responsible for

their inner workings. In the realm of trading, algorithms are used for data processing, market

forecasting, decision making, and order execution. Non-executing algorithms are intended to

present parsed data to the trader to inform his or her decision making. The focus of this paper

will be on executing algorithms. These programs are connected to the stock exchange and

execute trades without any human input. Traders oversee these algorithms to ensure that they

1
McGowan (2010)

3
produce the intended results.2 As industry advances and markets change, the code behind these

algorithms is updated accordingly. Human intervention is still critical to ensure that the

algorithms behave in the desired manner.

A couple of challenges were presented with the writing of this paper. Firstly, many works

on algorithmic trading are kept private. This is because financial firms do not wish to share their

trading strategies with others. In addition to limiting the resources available for research, this

makes it difficult to evaluate the different trading algorithms from a computational perspective.

In the world of computer science efficiency is key. Computers are expected to get tasks done in

the fastest, and therefore simplest, way possible. It is difficult to evaluate the complexity of an

algorithm without its code being readily available. In consistency with most of the published

literature on this topic, this paper will mainly outline methods for evaluating the algorithms from

a financial perspective. To briefly explain the ideal features of an algorithm from a computational

perspective, the focus is on time and space efficiency. This interest in time efficiency is shared

with traders because they compete to make split-second decisions as new data is processed.

Programmers have different coding styles which may affect the efficiency of their code.

Additionally, the computing power of the machine used to analyze data and execute trades can

make a substantial difference. The data sources chosen by the trader can also account for the

differences in performance between algorithms. This is without even looking at the code behind

a program. When evaluating the code, it is important to count the number of operations. The idea

is to implement the algorithm with as few operations as possible.3 It is important for

programmers to remember this when designing trading algorithms because this is essential for

minimizing time complexity.

2
Arnoldi (2016)
3
Rinker (2002)

4
Categorization of Trading Algorithms:

There exist multiple ways to categorize trading algorithms. The categories highlight the

differences between algorithms and make it easier to select the most appropriate algorithms for

specific cases. One determining factor is whether it is a single-order or list-based algorithm. The

former refers to an algorithm designed to take considerations of a single order into account

before executing a trade. The latter refers to an algorithm designed to take the considerations of

multiple orders into account before executing any of the trades. Single-order algorithms can be

fed lists of orders to execute; however, only list-based algorithms monitor the considerations of

all the orders in relation to one another. Another distinguishing factor is whether it is a

schedule-based, dynamic, or opportunistic algorithm. All algorithms split up orders and execute

them in parts over time. Schedule-based algorithms are unique in that they follow a predefined

schedule to trade the parts of an order. Dynamic algorithms are programmed to track market

conditions to find the most suitable rate at which to trade the parts of an order. Similarly,

opportunistic algorithms track market conditions to find opportune times to trade. The difference

between dynamic and opportunistic algorithms is that the former is guaranteed to execute all

parts of an order, even if at a slow rate. The latter may only initiate a trade if an opportunity in

the market arises.4

An additional way to categorize a trading algorithm is by whether it is an aggressive,

working order, or passive algorithm. An aggressive algorithm aims to execute all parts of an

order as fast as possible. This type of algorithm will trade all parts of an order at once if the price

is desirable. Many schedule-based algorithms are designed as aggressive algorithms because they

trade large volumes in short durations of time. A working order algorithm trades at a rate that
4
Bacidore (2020), pp.1-3

5
best minimizes the cost and risk. Many dynamic algorithms are designed as working order

algorithms and vary their rates of trading to keep losses to a minimum. A passive algorithm only

trades if its activity will not disturb the market. Many opportunistic algorithms are designed as

passive algorithms that only trade at times when they will not affect market liquidity.5

Schedule-Based Trading Algorithms:

Schedule-based single-order algorithms constitute one category of trading algorithms and

include time- and volume-weighted average price algorithms and the scheduled-based

implementation shortfall algorithm. The time-weighted average price (TWAP) algorithm is

simply a program that executes the parts of an order in proportion to time. The algorithm divides

the order into parts and evenly spaces out the execution of the resulting trades. It is possible to

set a minimum quantity for each trade to prevent the machine from trading millions of individual

stocks at increments of less than a second. An issue with this method is that it executes all orders

at market price without allowing the trader to set limits on trade prices. To incorporate limit

orders into the TWAP algorithm, the program may be allowed to go slightly off schedule to trade

more when the market price is favorable. A limit order is a trade that is executed on the condition

of being filled at a set price or better.6 Because it is a schedule-based algorithm, there are upper

and lower bounds that control the volume of trades that can be executed at each interval. The

lower bound ensures that the algorithm does not fall too off schedule while waiting for the

market price to match the price prescribed by a limit order. The upper bound ensures that the

algorithm does not act too aggressively and execute too many trades when the price is favorable.7

5
Kissell (2014), p.16
6
Cesari, et al. (2012)
7
Bacidore (2020), pp.31-35

6
 Another schedule-based single-order algorithm is the volume-weighted average price

(VWAP) algorithm. This algorithm tracks the volume of a stock that has been historically traded

at each time of day. It then divides up the order and trades different proportions of the total

volume at each interval of the order life. The order life is split up into intervals of time called

bins. A portion of the order is assigned to each bin, reflecting the quantity that has been

historically traded at that time. At each of these bins, the program follows the TWAP algorithm

in that it executes that piece of the order in proportion to time. Similar to the TWAP algorithm,

the VWAP algorithm accommodates limit orders by allowing itself to veer slightly off schedule.

It sets upper and lower bounds that limit the volume traded throughout the order life. It should be

noted that the TWAP algorithm is usually implemented as a VWAP algorithm set to a constant

rate of trade throughout the day.8

The schedule-based implementation shortfall (IS) algorithm is another scheduled-based

single-order algorithm. This algorithm uses the IS performance measure which compares the

arrival price, the midpoint between the buy and sell offers, with the realized price of the trade.

The realized price of the trade takes all costs, including any changes in market price, into

account.9 The goal of the algorithm is to buy the stock at a price below the arrival price or sell

the stock at a price above the arrival price. The algorithm assesses the current market prices and

chooses the behavior that minimizes the costs and risks of trading. It may choose to implement

the schedules of the TWAP or VWAP algorithms if deemed most appropriate. The program also

uses these algorithms as a lower bound to ensure that the order is completed on time. If the

schedule generated by the algorithm will result in too few trades, the program is compelled to

trade the volume prescribed by the lower bound.10

8
Ibid, pp.37-40
9
Hayes (2021)
10
Bacidore (2020), pp.43-47, 65

7
Dynamic Trading Algorithms:

Dynamic single-order algorithms constitute another category of trading algorithms and

include the standard, forecasted, and optimal percent of volume (POV) algorithms. The POV

algorithm tracks the trades of a specific stock in the market and ensures that its own participation

rate is kept constant. Similar to the VWAP algorithm, the POV algorithm monitors the volume of

trades in the market at each point during the order life. The difference is that the POV algorithm

monitors real-time data and executes orders to maintain responsibility for a percentage of trades

in the current market. There is no guarantee regarding the total volume that will be executed by

the POV algorithm, as there is with the VWAP algorithm. 11 The algorithm usually trades at

market price because it reacts to other orders in the market. Upper and lower bounds may be set

to accommodate limit orders and ensure that the algorithm does not deviate too far from its

prescribed participation rate.12

One variation of the aforementioned algorithm is the forecasted POV algorithm. The

algorithm divides the order life into intervals. At the end of each interval, it uses a model to

predict the volume of trades it expects to see in the next interval. It uses real-time data to forecast

the expected volume and calculates the number of trades it needs to initiate to maintain its

participation rate. It does this until the entire order has been filled or until the order life ends. The

forecasted POV algorithm differs from the standard POV algorithm in that it predicts future

volumes. In contrast, the standard POV algorithm tracks previous trades to keep up its

participation rate. The forecasted POV algorithm can be implemented similarly to the VWAP

algorithm if the historical data is replaced with real-time data and the end time is made dynamic.

11
Cesari, et al. (2012)
12
Bacidore (2020), pp.56-59

8
The key difference between the two is that the VWAP algorithm is concerned with the

percentage of an order that it fills at each point during the order life. The forecasted POV

algorithm cares about the percentage of trades in the market that it contributes at each interval.13

The optimal POV algorithm is a variation of both the POV algorithm and the IS

algorithm. Its goal is to determine the optimal participation rate based on real-time market data.

It can be fed definitions for aggressive, moderate, and passive participation rates to aid the

algorithm in choosing the optimal rate. This algorithm is dynamic in that its participation rate is

intended to fluctuate with time. The standard POV algorithm will not fill an entire order if the

market volume and participation rate call for too few trades. The optimal POV algorithm is also

an IS algorithm and can therefore be instructed to complete an entire order. It may use the TWAP

or VWAP algorithm as a lower bound to keep it on track. The algorithm will be compelled to

trade even when the optimal participation rate is very low to ensure that the trade is completed

by the end of the order life.14

Opportunistic Trading Algorithms:

Single-order opportunistic algorithms constitute the final category of single-order trading

algorithms. This category includes the “hide and take” algorithm and the adaptive arrival price

algorithm. The “hide and take” algorithm remains hidden until an opportunity in the market

presents itself. Unlike the algorithms of the other categories, it is not concerned with keeping to a

schedule or to a participation rate. It waits for an opportunity to arise rather than settling for

market prices. One factor that this algorithm may monitor is the price of the stock. It waits for

the stock to move into its desired price range before executing any trades. Another factor that the

13
Ibid, pp.60-63
14
Ibid, pp.64-65

9
algorithm may consider is market liquidity. The algorithm considers it an opportune time to trade

when market liquidity substantially increases. The algorithm may also create its own

opportunities. If it is used to monitor an illiquid market, it may execute some desirable orders to

draw traders to the other side of the market. The algorithm will then fill further orders before the

market turns illiquid again. For large quantities, the algorithm may seek out opportunities to

trade in the dark market rather than in the lit market. This way it can avoid impacting the market

price or liquidity. Similar to the bounds applied to the other algorithms, the “hide and take”

algorithm can uphold a minimum participation rate. This ensures that it will execute a reasonable

number of trades even if there are few opportunities. The algorithm may also allow the user to

indicate the level of urgency of the order. If the order is of high urgency, then as mentioned

before, the algorithm can create its own opportunities instead of waiting for them to arise.15

Another single-order opportunistic algorithm is the adaptive arrival price algorithm. This

algorithm is another variation of the IS algorithm. The algorithm monitors the market price and

trades more when the price moves in the desired direction and less when it moves the other way.

Unlike the “hide and take” algorithm, this algorithm seeks to constantly trade. The rate at which

it trades varies based on the current market conditions. The algorithm only exploits opportunities

related to the price of the stock, while the “hide and take” algorithm considers many factors. The

adaptive arrival price algorithm seeks to take advantage of current opportunities rather than

making predictions about future market conditions. Although this is the safer option, some

opportunities will inevitably be missed since it does not look ahead.16

Multi-Order Trading Algorithms:

15
Ibid, pp.68-70
16
Ibid, pp.71-72

10
 Moving away from single-order algorithms, the next category of trading algorithms is

multi-order, or list-based. The two types of list-based algorithms are pairs trading and portfolio

algorithms. Pairs trading algorithms monitor the price of two stocks in relation to one another.

The goal of the algorithm is to buy one of the stocks and sell the other when their price ratio

drastically changes.17 There are upper and lower trigger values set for the price ratio. If the ratio

exceeds the upper trigger value, the algorithm will buy units of one of the stocks and sell units of

the other. If the ratio falls short of the lower trigger value, the algorithm will buy and sell the

stocks in reverse. The algorithm stops trading when the ratio returns to an exit trigger value, or

the algorithm reaches the maximum position size. The exit trigger is a value that indicates that

the price ratio is returning to normal, and that the algorithm should refrain from any further

trading. The maximum position size is set by the user to prevent the algorithm from buying too

much of a stock at once. A leader stock is chosen based on which of the assets is more difficult to

trade. Units of the leader and follower stocks are traded in repeated intervals and at amounts

proportionate to their price ratio. These units are kept small to minimize the risk that the price

ratio changes before the follower stock is traded.18

Portfolio algorithms are the other type of multi-order algorithms. They monitor many

stocks at a time, making it very complex to apply the same logic as in the pairs trading algorithm.

Therefore, most portfolio algorithms implement an upgraded version of the schedule-based IS

algorithm. The schedule-based IS algorithm evaluates the costs and risks of a stock when

creating its trading schedule. In contrast, portfolio algorithms take the costs and risks of all the

stocks into consideration when creating individual trading schedules. The trading cost used in

generating each of the schedules is simply the sum of all the trading costs. The risk is more

17
Huang, et al. (2015)
18
Bacidore (2020), pp.85-89

11
complicated to calculate because it accounts for the effect that the trading of one stock in the

portfolio has another. The total risk is multiplied by a covariance term to account for the risk

each order creates in the markets of the others. Once the schedules have been produced, each of

the orders can be executed by a single-order algorithm, such as the VWAP algorithm. The

historical data in the VWAP algorithm is replaced with the schedule created by the portfolio

algorithm. The orders can either be sent in their entirety or in increments to the single-order

algorithms. The latter approach allows for the portfolio algorithm to act similarly to an

opportunistic algorithm. It can search for profit-making opportunities to trade the remaining parts

of the orders before the next increment begins. The algorithm must be careful not to disturb the

level of risk in any of the markets while looking for such opportunities, or else it would need to

recalculate all the trading schedules.19

Metrics for Measuring Performance:

It is not enough to understand the intricacies of the aforementioned algorithms. Traders

must also learn to compare their performances to select the best algorithms for specific cases.

There are different methods for comparing the performance of trading algorithms including

paired observations and independent samples. The paired observations approach takes an order

and divides it into two equal parts. The halves are fed into different algorithms and are given the

same start and end times for trading. This approach should only be used for algorithms that have

a static trading strategy. This includes the standard TWAP, VWAP, and POV algorithms. The

volume of trades may vary over time; however, the trading strategies of these algorithms do not

change. These algorithms will act nearly identically even when implemented by different traders.

The TWAP algorithm will always trade at a constant rate for a given volume size and order life.
19
Ibid, pp.90-98, 104-106

12
The VWAP algorithm follows the schedule prescribed by historical data and does not assess

current market conditions. The POV algorithm trades at the volume required to maintain its

predefined quota in the market. The paired observations method serves as a strong means of

comparison because the algorithms trade in the same market at the same time. The reason its

implementation is limited to algorithms with static trading strategies is because the results may

be skewed if the algorithms compete with one another. If the algorithms reacted to the trades

executed by their counterparts, then the results would not be a fair assessment of their behaviors.

Since the trading strategies of these algorithms do not vary with market conditions, it is safe to

have the pair of algorithms trade at the same time.20

Another approach for comparing trading algorithms is through independent samples. This

method is ideal for algorithms that use dynamic trading strategies. This includes the IS

algorithm, forecasted and optimal POV algorithms, “hide and take” algorithm, and adaptive

arrival price algorithm. The strategies of these algorithms vary with the change of market

conditions. If two of these algorithms were programmed to trade at the same time, they may end

up competing with one another. If one of the algorithms disturbs the market liquidity, it may

trigger the other to trade more or less aggressively than it would normally. These results would

not appropriately reflect the behaviors of the algorithms. The algorithms used in independent

samples should be given identical orders and time periods. The samples are independent because

the algorithms trade at different times. They should also be given large enough orders to ensure

that the algorithms trade at their intended capacities.21

Data from paired observations and independent samples is used to calculate metrics for

comparing the algorithms. Financial analysts most commonly use the IS and VWAP benchmarks

20
Kissell (2014), pp.115-116
21
Ibid

13
to evaluate the performance of trading algorithms. The IS benchmark compares the arrival price

and average execution price of an order. The arrival price is the midpoint between the buy and

sell offers at the time the decision to trade is made. The IS benchmark calculates the difference

between the arrival price and realized price. This metric considers the loss endured due to the

inability of the algorithm to trade large volumes all at once. The algorithm will inevitably buy

above or sell below the arrival price due to transactional delays. For each part of an order, the IS

benchmark measures the difference between the price at the time of the decision and the average

execution price. The IS benchmark and volume size of all the parts of an order must be

considered in evaluating the overall performance of a trading algorithm.22

Another metric used for measuring the performance of trading algorithms is the VWAP

benchmark. This metric takes the typical price of a stock over a set period and divides it by the

volume traded. The typical value is the average of the highest, lowest, and final price at which

the stock was traded. This is repeated for all periods during the order life and the values are

added all together. The VWAP benchmark is calculated by dividing this value by the total

volume traded during the order life.23 It is typically preferred to use the VWAP benchmark when

comparing trading algorithms through paired observations. This method is used for algorithms

that have static trading strategies and therefore trade more aggressively. The VWAP benchmark

is preferred for static algorithms because it measures the weighted average price. This is relevant

for aggressive algorithms since they constantly execute trades with the intention of maintaining a

desirable average execution price. The IS benchmark is preferred for comparing trading

algorithms through independent samples. This method is used for algorithms that have dynamic

trading strategies and therefore trade more passively.24 The IS benchmark is preferred for

22
Markov (2019)
23
Fernando (2022)
24
Kissell (2014), pp.116

14
dynamic algorithms because it measures the execution and opportunity costs.25 This is relevant

for passive algorithms since they aim to trade when costs and risks are minimized.26

Challenges of Algorithmic Trading:

Algorithmic trading presents many challenges that traders must consider in choosing

which algorithms to implement in specific cases. One challenge is selecting the best approach for

trading in illiquid markets. Illiquid markets have few participants which makes it difficult to

trade. As mentioned earlier, opportunistic algorithms can be used. Algorithms such as the “hide

and take” algorithm wait for opportunities to present themselves. They may also resort to

creating their own opportunities by temporarily drawing more participants to the market. The

issue with waiting for opportunities to arise is that there are more risks associated with large

orders. If the algorithm waits for an opportunity it will try to exploit it by issuing trades at large

volumes. A better solution may be to use algorithms with static trading strategies such as the

TWAP, VWAP, and POV algorithms. These algorithms produce better results in illiquid markets

than algorithms with dynamic trading strategies because their focus is on the average execution

price. The intention of dynamic algorithms is to minimize the costs and risks of every trade. If an

illiquid market were to attract more participants, a dynamic algorithm may trade aggressively to

take advantage of the larger spread of buy and sell bids. This approach could drain the market of

its resources, adding risk to the resulting orders. Instead, it may be better to use static algorithms

that trade at a steady rate. These algorithms can still trade more aggressively when opportunities

arise; however, their upper bounds prevent them from disturbing the liquidity of the market.27

25
Hayes (2021)
26
Markov (2019)
27
Shen (2021)

15
 Another challenge of algorithmic trading is handling special days. Special days include

half trading days, index rebalancing days, and federal announcement days. The occurrences on

these days may impact the overall stock market or individual stocks. Historical data alone cannot

accurately forecast market movements on these days due to the uncertainty of external factors.

Algorithms that incorporate advanced statistical models or machine learning should be used to

predict market changes on these days. These models make better predictions than standard

trading algorithms because they take factors other than historical data into account. Some of

these models can even be implemented to accommodate special periods of time. These periods

can last up to a week or a month. This is relevant when a stock is first introduced to the public

and its trading patterns are uncertain. Models for special periods are not regularly used because

they are difficult to compute; however, it is in the best interest of traders to produce these models

and take advantage of the opportunities presented on special days.28

An additional challenge of algorithmic trading is the potential for human error in

programming. Algorithms are improved over time to correct mistakes, update strategies, and

implement new features. Turnover is common in industry and new people are often assigned to

work on existing projects. Programmers do not all share the same coding styles. This makes it

difficult for new employees to work on existing programs and often results in coding errors. This

is especially an issue in algorithmic trading because even a small mistake can lead to a

significant loss of money. Solutions to this problem include the use of object-oriented

programming and unit and regression testing.29

The first solution to the issue of human error is to use programming languages that

support object-oriented programming such as Java and C++. Object-oriented programming is a

28
Ibid
29
Ibid

16
programming style that focuses on objects and their transactions. Objects represent real-world

entities and are assigned unique identities, attributes, and behaviors. Classes of code are designed

to initialize objects and regulate their transactions. Programmers can limit the accessibility of

features within an object’s class to ensure that they are only used in the intended manners. The

reason that this form of programming is protective against human error is because it is

formulated in layers of classes that are limited in their interactions with one another. The idea is

to restrict changes made in one area of the code from affecting other areas.30 This is not a full

proof solution because objects are designed to interact with one another. These interactions may

not work as originally intended if later changes are made by new programmers. It is therefore

important to understand the intricacies of a program before making any changes to it.31

Regardless of whether a program has the protective features of object-oriented

programming, it is important to thoroughly test all aspects of the code. Unit tests are used to test

individual functions of a program. Regression testing is used to test the overall program after

changes are made to it. The issue that arises with these tests is that they often do not check for all

possible edge cases. To thoroughly test a trading algorithm, simulations of extreme market

conditions must be automated to assess the reactions of the program. The reason that this is

difficult to achieve is because most algorithms are comprised of series of conditional statements.

It is not simple to ensure that all combinations of an algorithm’s features are tested. A unit test

could be used to check the functionality of an update to the program. Further regression testing

would be necessary to examine how that change affects the remainder of the program. It is

important to note that some updates to a program will produce more possible outcomes.

30
Prabhu (2022)
31
Shen (2021)

17
Therefore, although testing is sometimes viewed as a tedious process, it is important to update

the tests as the program is improved to ensure the effectiveness of the trading algorithm.32

Integration of Machine Learning:

With the rise of big data came the integration of machine learning into the trading

industry. Machine learning algorithms are not explicitly programmed; rather, they are designed to

train themselves. This differentiates them from the trading algorithms previously discussed

because those require the programmers to give explicit directions in the code. The three

categories of machine learning include supervised, unsupervised, and reinforcement learning.

Supervised learning is used to label sets of data. This type of machine learning model uses

labeled data to train itself to predict labels for unseen data. It produces a function that takes

unseen data as input and returns its predicted label. Unsupervised learning is similar, except that

it is not trained with labeled data. Rather, it must learn to categorize data into groups on its own.

It produces a mapping system that classifies the data that it takes as input and returns its labels.

Reinforcement learning trains an agent to choose the actions that will maximize its reward. The

actions performed by the agent move it between the predefined states of its environment. The

agent learns which actions to choose based on the occasional rewards it receives. This form of

machine learning is used for making predictions and is the category most used in trading.33 When

applying reinforcement learning to trading, the environment refers to the collection of data that

the agent, or machine, uses to make decisions. The states refer to the data that describes the

current market conditions and the agent’s position. The agent uses a learned policy to inform it of

32
Ibid
33
Koshiyama, et al. (2020)

18
the best action to take based on the current state. In cases of portfolio management, the agent

may have to make multiple decisions to claim rewards and bring it to the next state.34

Machine learning is applied in the trading industry to monitor market data and execute

trades. Natural language processing is the area of machine learning used for analyzing language

and speech. It can be used to examine data from social media and assess the public attitude

towards a stock. It can also be used to analyze statements released by a publicly traded company

and evaluate its economic well-being. Machine learning algorithms typically make more accurate

predictions than standard trading algorithms due to the inclusion of this data. These market

monitoring algorithms are paired with other machine learning algorithms that execute trades.

They act as smart order routers that monitor the price of a stock across different venues and

execute the desired trade on the platform with the best price.35 Machine learning algorithms are

also used to devise optimal trading schedules. Many of these algorithms are adaptations of the

VWAP algorithm. The difference is that they use reinforcement learning to inform their trading

strategies. This is advantageous because reinforcement learning algorithms are trained based on

past experiences. The VWAP algorithm only uses historical data in its decision making and

cannot learn from past successes or failures. Although using reinforcement learning is a strong

strategy for order execution, it is very difficult to implement in markets that are constantly

changing.36

Another area of machine learning is deep learning. It is categorized by its hidden

processing layers that transform inputted data. Each layer is designed to bring the data closer to

the desired results. This structure is known as a neural network and is intended to replicate the

functionality of the human brain. One algorithm that uses supervised deep learning is the

34
Ritter (2017)
35
Koshiyama, et al. (2020)
36
Hansen (2020)

19
long-short term memory (LSTM) algorithm. It is mostly used for predicting stock prices. It is

structured as a recurrent neural network that stores past data to inform future decision making.

This type of neural network is useful for recognizing patterns in times series data because of its

ability to reference past data. The LSTM algorithm categorizes data into long- and short-term

memory cells. It must limit the amount of data that it stores to optimize its efficiency. The

algorithm would work very slowly if it had to examine all the past data in making predictions.

This would not be useful for the split-second decision making required of the fast-paced

markets.37

Moving away from the different applications of machine learning in trading, it is essential

to understand how to choose which algorithms to implement. Traders must consider the tradeoff

presented between accuracy and interpretability. Many machine learning algorithms are called

black boxes due to the inability of humans to understand their inner workings. The reason for

this tradeoff is because algorithms tend to make more accurate predictions as they examine more

parameters. Algorithms also become more complex as they incorporate a larger number of

parameters. The more complex an algorithm, the less interpretable it becomes to humans.

Reducible errors occur when improper algorithms for specific cases are implemented. To

minimize these types of errors, it is essential to understand the data and choose the algorithm that

best approximates the relationship between inputs and outputs. Two types of reducible errors

exist and are known as bias errors and variance errors. Bias errors occur when the programmer

makes incorrect assumptions about the relationship between variables. The programmer chooses

to implement an algorithm that is incapable of approximating the function between inputs and

outputs. The results produced by such an algorithm are not accurate. Variance errors arise when

the algorithm is overfitted to the training data. Overfitting occurs when the function produced by
37
Koshiyama, et al. (2020)

20
the algorithm is too heavily influenced by trends specific to the training data. The algorithm will

then make predictions that are not accurate for new data. Deep learning algorithms tend to have

higher variance errors and lower bias errors. The reverse is true for simpler machine learning

algorithms. These types of errors are inevitable; however, they can be reduced if the programmer

is well versed in the different types of machine learning algorithms. He or she must study the

data set before selecting which algorithm to implement.38

Conclusion:

The intention of this paper is to emphasize the importance of education in the field of

trading. This paper was purposely written in a manner that is understandable to someone with

limited background in computer science or finance. With today’s technology, it is easy for

anyone to access stock prices and initiate trades. Trading can be expanded beyond this with the

incorporation of algorithms. Computers process data at speeds that far exceed the capabilities of

the human brain. It is therefore essential for traders to understand the potential of the powerful

tool that is algorithmic trading. As research in machine learning advances, these algorithms are

being implemented by more financial firms. Machine learning algorithms are capable of

processing more data than standard algorithms and implementing trading strategies that are

beyond human understanding. Standard algorithms are limited to the strategies programmed by

humans, but machine learning algorithms surpass human intellect. Although machine learning is

enticing, the field presents many of its own challenges. It is therefore important to acknowledge

these setbacks before choosing to implement a machine learning strategy. No matter which

approach is taken for trading, the trader must educate him or herself about the methods used to

38
Lee and Shin (2020)

21
compare and evaluate algorithms. Only through this process can he or she choose the best

algorithm to trade in specific scenarios.

22
Appendix:

Schedule-Based Dynamic Opportunistic

Single-Order TWAP, POV, “Hide and Take”,

VWAP, Forecasted POV, Adaptive Arrival

Schedule-Based IS Optimal POV (IS) Price (IS)

Figure 1: Classification of Single-Order Trading Algorithms

2 Orders 3+ Orders

Multi-Order Pairs Trading Portfolio Algorithm

Figure 2: Classification of Multi-Order Trading Algorithms

Static Trading Strategies Dynamic Trading Strategies

TWAP, Schedule-Based IS,

VWAP, Forecasted POV,

POV Optimal POV,

“Hide and Take”,

Adaptive Arrival Price

Figure 3: Classification of Single-Order Trading Algorithm Strategies

23
 References

1. Arnoldi, Jakob. “Computer Algorithms, Market Manipulation and the Institutionalization

of High Frequency Trading.” Theory, Culture & Society, vol. 33, no. 1, Oct. 2016, pp.

29–52., doi:10.1177/0263276414566642.

2. Bacidore, Jeffrey M. Algorithmic Trading: A Practitioner's Guide. TBG Press, 2020.

3. Cesari, Riccardo, et al. “Effective Trade Execution.” Portfolio Theory and Management,

22 June 2012, pp. 411–437., doi:10.1093/acprof:oso/9780199829699.003.0018.

4. Fernando, Jason. “Volume-Weighted Average Price (VWAP).” Investopedia,

Investopedia, 6 May 2022, www.investopedia.com/terms/v/vwap.asp.

5. Hansen, Kristian Bondo. “The Virtue of Simplicity: On Machine Learning Models in

Algorithmic Trading.” Big Data & Society, vol. 7, no. 1, 2020,

doi:10.1177/2053951720926558.

6. Hayes, Adam. “Implementation Shortfall.” Investopedia, Investopedia, 19 May 2021,

www.investopedia.com/terms/i/implementation-shortfall.asp.

7. Huang, Chien-Feng, et al. “An Intelligent Model for Pairs Trading Using Genetic

Algorithms.” Computational Intelligence and Neuroscience, vol. 2015, 14 Mar. 2015,

doi:10.1155/2015/939606.

8. Kissell, Robert. The Science of Algorithmic Trading and Portfolio Management.

Elsevier/Academic Press, 2014.

9. Koshiyama, Adriano, et al. “Algorithms in Future Capital Markets.” 11 Mar. 2020.

10. Lee, In, and Yong Jae Shin. “Machine Learning for Enterprises: Applications, Algorithm

Selection, and Challenges.” Business Horizons, vol. 63, no. 2, 2020, pp. 157–170.,

doi:10.1016/j.bushor.2019.10.005.

24
11. Markov, Vladimir. “Bayesian Trading Cost Analysis and Ranking of Broker Algorithms.”

22 Apr. 2019.

12. McGowan, Michael J. “The Rise of Computerized High Frequency Trading: Use and

Controversy.” Duke Law & Technology Review, vol. 16, 2010.

13. Prabhu, Rishabh. “Object Oriented Programming (OOPS) Concept in Java.”

GeeksforGeeks, 13 May 2022,

www.geeksforgeeks.org/object-oriented-programming-oops-concept-in-java/.

14. Rinker, Robert. “Algorithm Efficiency.” 2002.

15. Ritter, Gordon. “Machine Learning for Trading.” SSRN Electronic Journal, 21 Aug.

2017, doi:10.2139/ssrn.3015609.

16. Shen, Jackie. “Nine Challenges in Modern Algorithmic Trading and Controls.”

Algorithmic Trading and Controls, vol. 1, no. 1, 21 Jan. 2021, pp. 1–9.

25