On Present Use of Machine Learning based Automation in Finance

Vibha Tripathi

MIO, U.S.A.

Keywords: Data Analytics, Big Data, Artificial Intelligence (AI), Machine Learning, Deep Learning, Natural Language

Processing, Artificial Neural Networks (ANN), Gradient Boosted Ensembles (GBM), Market Prediction,

Portfolio Selection, Algorithmic Trading, Statistical Models, Biases, Data Confinement, Transfer Learning,

Model Locality, Interpretability, Explainable AI.

Abstract: In this paper, we survey the current known applications of Machine Learning based Data Analytics and

automation in finance industry. We look into the challenges involved in furthering this technology,

particularly in employing more Deep Learning approaches proven for successful automation in other

domains. We enumerate observations on some of the barriers faced by the industry in effectively adopting

and accelerating use of AI techniques, and finally propose more areas that we believe could further benefit

from application of Machine Learning.

1 INTRODUCTION

In the recent past, Machine Learning (ML) based

data analytics has made significant advances in

certain industry sectors e.g. location-based services

and US retail (MGI report 2016), however other

sectors like manufacturing, the EU public sector, US

health care and finance and banking industry are still

lagging behind in effectively using ML approaches

in solving existing problems or in exploring areas to

expand where applying AI may help. As AI

techniques are turning into a fundamental

component of business growth and remodelling

across a wide range of industries, the spectrum of

possible applications of AI in these sectors is

continuously widening as we gather and organize

more and more data for training AI models.

The amount of captured data is roughly doubled

every year. According to a 2016 IDC report, by 2025

we would have created 180 Zettabytes of data. Still,

at present the progress in capturing value from these

data has been uneven across the industries due to

many reasons; some of which have been identified

as lack of analytical talent, siloed data amongst

different companies or groups within a company and

scepticism within leaderships with regards to the

impact.

In particular, the financial domain is fast

catching up in its attempt to utilize AI methods in

data analytics-based automation (Kolanovic et al.,

2017). The Wall Street looks keen on investing

hugely in this area of technology; but there still

remains a dearth of technological discussions and

literature on identifying areas where e.g., trading

strategies, market prediction or correction, risk

calculations, pricing models and portfolio

management, can leverage from Machine Learning.

At best, the financial firms are forging into AI usage

in silos thereby creating an even greater need for

debate on the future implications of AI usage across

this domain.

In this paper, we study the current known

applications of AI in finance and banking industry

and survey the literature available on Deep Learning

approaches both proposed and demonstrated in

finance applications. We discuss the known as well

as emerging barriers to using AI in this industry and

conclude with proposals for further application areas

that we believe could benefit from this branch of

Machine Learning.

We start with a brief definition of Deep Learning

and its related terms and go onto exploring some of

the known present applications of Deep Learning in

financial sector, discuss some hurdles to ML

application in finance and then move on to proposals

for additional areas of usage.

412

Tripathi, V.

On Present Use of Machine Learning based Automation in Finance.

DOI: 10.5220/0008480504120418

In Proceedings of the 11th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2019), pages 412-418

ISBN: 978-989-758-382-7

2 DEEP LEARNING

Deep Learning is a branch of Machine Learning that

aims at discovering a mapping function between

given input data sets and expected output(s). A Deep

Learning module would analyse data in multiple

steps or layers of learning, which means it will begin

by learning a few simple concepts and proceed to

learning more complex ones by combining the

simpler ones in an iterative fashion in these layers.

A generally well-known goal of Artificial

Intelligence is to automate human tasks that would

be simple and easy for a machine to learn and

perform. Deep Learning is different from such

automation in that it works on more abstractly

defined problems where the learning is attempted by

the machine based on data sets on both input and

expected output. So far, a prominently used method

in Deep Learning is use of Neural Networks which

mimic how neurons perform learning in a human

brain.

Unlike other Machine Learning techniques, Deep

Learning does not involve automating human tasks

that are easy to define and perform. Deep Learning

aims to discover a mapping function between large

and variously sourced input and output datasets

thereby eventually developing an ‘intuition’ to arrive

at an accurate output prediction given an input

dataset.

2.1 Definitions

A generic Deep Learning module can be described

as follows. If there are n layers of Neurons in the

Deep Neural Network, let f1… fn be given

univariate activation functions for each of the n

layers.

These activation functions are nonlinear

transformations of weighted data. The quality of a

good predictor depends on the selection of univariate

activation functions f (i) at each layer of the neural

network. The algorithm extracts hidden factors or

features at each layer. Since the weights are

matrices, a deep learning predictor has greater

flexibility to uncover nonlinear features of the data.

Most commonly we divide the data sets into

three subsets, for training, validation, and testing.

The training set is used to adjust the weights of the

network. The validation set is used to minimize the

over-fitting and pertains to model selection. Testing

data set is used to confirm the actual predictive

power of a deep learner (Heaton et al., 2016).

2.2 Types of Deep Neural Networks

Deep neural networks (DNNs) are more

sophisticated artificial neural networks (ANNs) that

use several hidden layers. DNNs have proven very

successful in retail commerce sector with the use in

speech transcription and image recognition

(Krizhevsky et al., 2012) due to their superior

predictive properties and robustness to overfitting.

The most popular ways Deep Neural Networks

can be implemented are as follows:

 Multi-Layer Perceptrons or Feed Forward

Networks

 Restricted Boltzmann Networks

 Convolutional Neural Networks

 Long Short-Term Memory

2.2.1 Multi-Layer Perceptron (MLP)

As one of the first designs of multi-layer neural

networks, Perceptron is implemented in a Feed

Forward way such that the input passes through each

node of the neural network exactly once.

2.2.2 Restricted Boltzmann Networks (RBN)

RBNs are based on dimensionality reduction such

that the neurons in an RBN form two layers. The

first being the visible units (returns of assets) and the

second hidden units (hidden factors or features).

Neurons within the same type of layer hidden or

visible are not connected to each other creating a

restriction in the neural network.

2.2.3 Convolutional Neural Networks (CNN)

Convolutional Neural Networks (CNNs) are

popularly used for classifying and detecting objects

in images. A CNN extracts data features by using

multiple filters on overlapping segments of the

image, deriving first the simple parts and then

connected features of an image.

2.2.4 Long Short-Term Memory (LSTM)

Unlike Multi-Layer Perceptron which can only feed

forward, Long Short-term memory (LSTM) is a

neural network that includes feedback loops between

its elements.

LSTM networks work well with time series

analysis, as they can recognize patterns across

different time scales.

On Present Use of Machine Learning based Automation in Finance

413

3 DEEP LEARNING USE CASES

IN FINANCE

3.1 Portfolio Selection and Pricing

Although most Hedge Funds keep their portfolio

pricing methodologies strictly confidential, it’s been

demonstrated by a few that using Machine Learning

to price portfolios in overly aggressive markets,

firms can discover hidden features in their portfolio

data and adjust pricing to maximize profit.

(Heaton et al., 2016) Worked on finding a

selection of investments for which good out-of-

sample tracking properties of their objective could

be discovered. They used weekly returns data for the

component stocks of the biotechnology IBB index

for a period of about four years and trained a deep

learner without knowledge of the component

weights.

By implementing a four-step Deep Learning

algorithm which goes from auto-encoding,

calibrating, validating, to verifying they show that a

data-driven and model-independent Deep Learning

approach can be a new paradigm for prediction.

3.2 Financial Markets Predictions

Due to the computation complexities involved in

using Deep Neural Networks (DNNs) in algorithmic

trading, their proposed usage has posed a few

concerns. The financial research communities and

practitioners are still to find a meeting ground where

the traditional financial econometrics joins hands

with practical machine learning.

In their research (Dixon et al., 2016), use DNNs

to model complex non-linear relationships between

the independent variables and dependent variable

with reduced tendency to over-fit. To do so they

utilize low cost multi-core

accelerator platform to

train and tune the parameters of their model.

They

exploit state-of-the-art parallel computing

architecture to implement a feed-forward topology

for multivariate forecasting analysis. Using 5 minute

interval prices over five years, they simultaneously

train a single model from a large number of signals

across multiple instruments instead of one model for

each instrument. The aggregated model is thus able

to capture a more complete set of information to

describe time-varying movements for each

instrument’s price. They claim that their model is

able to predict the direction of instrument movement

to average 42% accuracy with a standard dev of 11%

across the instruments. The parallel trained model

also yields back testing accuracy which directly

translates into P&L for simple long-only trading

strategy.

3.3 Bond and Options Pricing

Bond markets suffer from a comparative lack of

trading information as opposed to equity trading.

Many bond prices are days old and do not represent

latest market developments due to the fact that the

information available on bond trades is scarcely

available often on a fee for data contract basis

(Benchmark Solutions, 2014).

(Ganguli et al., 2017) Show in their experiments

that Neural Networks give very accurate results

without overfitting in reasonable amounts of time

(order of hours). The success of Neural Networks on

this dataset implies that investigating the application

of multilayer networks and Deep Learning methods

to this problem may yield better bond price

predictions.

(Deoda et al., 2011) investigated option pricing

performance of non-parametric machine learning

techniques for Nifty index call options versus

parametric Black-Scholes model for 1-day & 1-week

price forecast. Their results suggest non-parametric

machine learning techniques outperform parametric

Black-Scholes model. They observed that the

nonparametric machine learning techniques adjust

more rapidly to changing market behavior and are

able to capture the pattern more effectively

compared to parametric models.

Using more sophisticated techniques such as

Deep Learning for calibrating the models and

filtering the data, pricing performance can be

improved

even further.

3.4 Stock Pricing based on Sentiment

Analysis

The stock market is influenced by various factors

including a plethora of human-related ones some of

which can be regime change, political scenarios,

climate, news media etc. Financial predictive

analytics can benefit from Deep Learning by using

most such factors to train the analytical models and

predict trends and signals.

In their research (Ding et al., 2015) demonstrate

that Deep Learning is useful for event-driven stock

price movement prediction. They propose a novel

Neural Tensor Network for learning event

embedding, and use a Deep Convolutional Neural

Network (CNN) to model the combined influence of

long-term events and short-term events on stock

KMIS 2019 - 11th International Conference on Knowledge Management and Information Systems

414

price movements. Their results show that event

embedding-based representations perform better

than discrete events-based methods, and deep CNN

capture longer-term influence of news event than

standard feed-forward neural network. They use a

simple greedy strategy to simulate the market to

make their model yield relatively more profit.

3.5 Real-time Fraud Detection and

Compliance

Machine Learning algorithms have been in use in

the area of detecting credit card fraud for some time

now. However the presently used ML techniques are

still rule-based and individual transaction oriented.

Fraud monitoring systems today are not able to

detect complex transactional patterns resulting into a

huge amount of false positive fraud alerts which

require human assessment and filtering.

There are nonetheless, huge credit card

transactions datasets available with the providers

that can be used to train systems that not only detect

but also predict fraudulent transactions.

Deep Learning has the potential to improve

detection of fraudulent and money laundering

activities. Deep Neural Networks have been shown

to identify complex patterns in the data and combine

transactions information at network speed, utilizing

data from many different sources to yield a more

complete picture of a client’s activity. These systems

have been shown to bring false positives down

significantly as well

3.6 Statistical vs. Deep Learning

Models

According to a report, more financial institutions are

now replacing their older statistical-modelling

algorithms with machine learning techniques. These

institutions have reportedly observed 10 percent

increases in sales of new products, 20 percent

savings in capital expenditures, 20 percent increases

in cash collections, and 20 percent declines in churn.

Deep Learning can help banks implement more

outreaching and effective recommendation engines

for clients in retailing and in small and medium-

sized companies. These systems can employ micro-

targeted models that more accurately forecast who

will cancel service or default on their loans, and how

best to intervene (Lalithraj et al., 2017).

4 BIASES AND BARRIERS

4.1 Biases

As we survey the use of Machine Learning based

analytics of big data in finance, we cannot ignore the

ubiquity of biases in learning from the data. Bias is

present in every type of learning. Bias in learning is

defined as any basis for choosing one generalization

over another, other than strict consistency with the

instances (Mitchell, 1980). Categorizing originally

as only two broad ways as representational and

procedural (Gordon et al., 1995) suggest that

selection and evaluation of biases is critical task for

any intelligent systems. These biases in the present

Machine Learning based analytic applications can be

further specified as algorithm bias and data bias

which further subdivides into Sample, Prejudicial,

and Measurement Biases on Data. Creating

standards and frameworks to allow a balance of

biases and variance in training data for financial

applications is a task impending on the regulatory

bodies and consortiums.

In addition to biases, there are multiple barriers

that the industry faces in application of AI to its

fullest potential. Some of which are discussed

briefly here.

4.2 Data Confinement

A firm’s data is its competitive advantage over

others. In some instances, the analytical value

derived from this data allows the firm to disrupt the

industry by strengthening its core business while in

others whole new business models can be adopted as

a result of learning from the data at its disposal.

However, data confinement is a prominent issue

facing many of the finance and banking industry.

Additionally, as opposed to consumer applications,

finance and banking industry have less data to

individualize their services beyond their present

consumers. Transfer Learning is proposed as a more

promising technique than Deep Learning in such

cases where labelled training data from a related

domain can be used in further learning from outside

data.

4.3 Disparate Data Formats

Another practical reality about data present in

finance and banking industry is the multitudes of

data formats like PDFs, MS Word documents, Excel

Sheets, to name a few. Deriving interchangeable

information from these formats becomes a challenge

On Present Use of Machine Learning based Automation in Finance

415

for AI methods like Natural Language Processing

(Carlton et al., 2017).

Cleaning and consolidating gathered data for

training, from different data sources within a firm is

also an organizational challenge faced in creating

useful AI based applications as data often sits in

silos amongst different groups across the firm.

4.4 Interpretability and Model Locality

The fierce competition to apply machine learning in

finance has been motivating data scientists to

experiment with complex predictive modelling using

for example artificial neural networks (ANN) and

gradient boosted ensembles (GBM). While this is

being done with a goal to achieve more predictive

accuracy, the models are increasingly becoming

black-boxes due to their unexplainable inner

workings.

Model interpretability is critical for

documentation and regulatory oversight, business

and human adoption. Finance and banking industry

face stricter regulatory and documentation

requirements, forcing financial data scientists to

continue to use traditional linear modelling to create

their predictive models.

Additionally, as finance and banking industry

adopt more machine learning based automation into

their decision-making systems, interpretability

becomes ever more important for ruling out any

deliberate or prejudicial decisions (Patrick et al.,

2017).

It is well known that learning algorithms can

produce multiple accurate models with very similar

internal architectures (Chiyuan et al., 2017),

provided the same input dataset and targets. This

phenomenon is known as model locality and is

another hurdle to model interpretation.

Various testing techniques like model

visualization, reason code generation, and sensitivity

analyses have been proposed to ensure

interpretability, however this area demands more

practicable research and tools development and

eventually industry wide standardization and

adoption.

The area of testing predictive models for

interpretability, fairness, accountability and

trustworthiness (FAT) is fast evolving, however it

has proven to be more intricate to apply in finance

industry.

5 PROPOSED AREAS OF USAGE

FOR DEEP LEARNING

5.1 Pricing and Risk Calculations

Financial risk management today relies on acquiring

and using several sources of data to run

sophisticated algorithms to compute results in

advance (Liebergen, 2017).

Deep Learning has proven ability to analyse very

large amounts of data from multiple sources and can

be used to produce in-depth predictive analysis with

high granularity (Härle, et al., 2016), thereby

holding the potential to greatly enhance analytics in

risk management as well as compliance.

There are also emerging systems where User

Defined Functions enable analytics at the data level

within the infrastructure. (Kinetica., 2017) is one

such example. UDFs enable a next generation risk

management platform that will also allow real-time

drill-down analytics and on-demand custom XVA

library execution.

5.2 Financial Time Series Forecasting

One of the important components of financial time

series forecasting is stock index prediction.

Derivative trading vehicles based on stock indices

provide the means to hedge against systemic risks

and diversify a portfolio. Finding better techniques

for stock index prediction is a continuous challenge

as it equips the market participants to make better

investment decisions.

(Krollner et al., 2010) observe that Artificial

Neural Networks (ANNs) have been identified as the

dominant machine learning technique in this area.

However, while the market researchers agree on the

importance of stock index forecasting, the task of

experimenting with deep neural networks is yet to be

done.

5.3 Technical Analysis

As mentioned earlier, Convolutional Neural

Networks have proven to be very efficient in

classifying images and object detection.

There are proposals to use CNN in the trading, in

particular in the area of technical analysis to detect

price chart patterns of technical analysts which are

difficult to define mathematically. As technical

analyses can have many variations based on time

scale.

(Krizhevsky et al., 2012) propose that various

technical patterns and even specific calls from

KMIS 2019 - 11th International Conference on Knowledge Management and Information Systems

416

prominent technical analysts can be used to train

CNNs, and then tested for their predictive power is

in the specific pattern, or specific analyst. Patterns

with significant forecasting power can be automated

and applied over a broad range of assets at a scale

that would be impossible to achieve by a human

technical analyst.

5.4 Audit and Reporting

Audit and Reporting processes involve massive

amounts of data and require auditors to solve, such

as text analysis, speech recognition, and parsing

images and videos.

Deep learning can help by automating the routine

tasks to improve audit efficiency and effectiveness

by facilitating repetitive audit procedures and

supporting audit judgments. Automating some

substantive procedures, such as confirmation and

examination will allow auditors to perform tasks that

are currently cost prohibitive or too complex, for

example exhaustively examining all corporate

contracts.

6 CONCLUSIONS

Machine Learning has proven to be an effective

framework in the areas of speech and image

recognition. It offers a system to use large data sets

to learn abstract mathematical definitions.

Financial industry continues to utilize statistical

models to make decisions like portfolio selection,

stock market prediction, risk calculations, pricing

models etc. Machine learning has the potential to

improve on predictive performance in financial

applications. However, there are many known

obstacles to adopting AI in finance in the current

state of affairs.

Inherent data and model biases underlying the

machine learning based automation have created an

ever-increasing need for regulatory oversight and

hence force financial data scientists to continue to

rely on their traditional linear predictive models.

While non-linear models created by trained

machine learning algorithms may produce more

accurate predictions resulting into better financial

margins, the approvability of such models still

remains in the hands of business partners, and

regulators.

Explainable AI and Machine Learning

interpretability are areas of research that are subject

to rapid changes and expansions at the moment.

Interpretable Machine Learning models are still very

difficult to achieve and hence finance and banking

industry have to start focusing more on explainable

models and their interpretability before any real

applications of ML.

In this paper we surveyed several areas in

financial data science where Machine Learning is

either presently being used or has been demonstrated

beneficial to use through research. We considered

some of the biases and barriers in the application of

Machine Learning in financial domain. In the end,

discussed several other areas in financial domain

where Deep Learning can be utilized effectively.

Areas that require further work both in terms of

research and in tools and processes development

include but are not limited to bias selection and

model interpretability. Among other hurdles to using

AI in finance are, rethinking solutions to data

confinement and utilizing disparate data hosted by

financial firms for training Machine Learning

models.

REFERENCES

McKinsey Global Institute, in collaboration with

McKinsey Analytics, Dec 2016. The Age of Analytics

– Competing in a Data Driven World. In McKinsey &

Company Report.

IDC Report 2016, https://www.forbes.com/sites/

gilpress/2016/08/05/iot-mid-year-update-from-idc-

and-other-research-firms/#6bd63a6955c5

Kolanovic, M., Krishnamachari, R. May 2017, Big Data

and AI Strategies Machine Learning and Alternative

Data Approach to Investing, JP Morgan Report.

Liebergen, Bart van 2017 Machine Learning: A

Revolution in Risk Management and Compliance? –

The CAPCO Institute Journal of Financial

Transformation, 2017

Bjoern Krollner, Bruce Vanstone, Gavin Finnie, 2010

"Financial Time Series Forecasting with Machine

Learning Techniques: A Survey", ESANN 2010

proceedings, European Symposium on Artificial

Neural Networks - Computational Intelligence and

Machine Learning. Bruges (Belgium).

Deoda, Amit June 2011 "Option Pricing using Machine

Learning techniques", Indian Institute of Technology,

Bombay

Dixon, Matthew Francis and Klabjan, Diego and Bang, Jin

Hoon, July 2016, “Classification-Based Financial

Markets Prediction Using Deep Neural Networks”,

Algorithmic Finance.

Heaton, J.B. and Polson, Nick and Witte, Jan Hendrik,

September 5, 2016 , “Deep Learning for Finance:

Deep Portfolios” Wiley Online Library

Ganguli, Swetava Dunnmon, Jared 2017 "Machine

Learning for Better Models for Predicting Bond

Prices" Statistical Finance (q-fin.ST); Computational

On Present Use of Machine Learning based Automation in Finance

417

Engineering, Finance, and Science (cs.CE)arXiv:

1705.01142 [q-fin.ST]

Benchmark Solutions, 2014, "Benchmark Bond Trade

Price Challenge", www.kaggle.com.

Härle, Philipp, Andras Havas, and Hamid Samandari, July

2016 "The future of bank risk management",

McKinsey & Company Report

Karthik Lalithraj, March 2017, "How GPUs and Deep

Learning Are Fuelling the Financial Industry",

NVIDIA Blog

Kinetica, GPU-Accelerated Database for Financial

Services, https://www.kinetica.com/solutions/finance/

Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin

Recht, Oriol Vinyals, 2017 "Understanding deep

learning requires rethinking generalization" ICLR

2017 arXiv:1611.03530 [cs.LG]

Carlton E. Sapp, 2017 "Preparing and Architecting for

Machine Learning" Gartner Report, January 2017 ID:

G00317328

Krizhevsky, A, Sutskever, I, Hinton, GE, 2012 "Imagenet

classification with deep convolutional neural

networks" - Advances in neural information

processing systems, 2012

Xiao Ding, Yue Zhang,Ting Liu, Junwen Duan, 2015

"Deep Learning for Event-Driven Stock Prediction",

International Joint Conference on Artificial

Intelligence (IJCAI 2015)

Mitchell, T, 1980, The need for biases in learning

generalizations. Technical Report CBM-TR-117,

Rutgers University, (1980)

Gordon, Diana F., Marie Desjardins 1995, “Evaluation

and Selection of Biases in Machine Learning”

Machine Learning, 20, 5-22 (1995)

Patrick Hall, Phan Wen, and Sri Satish Ambati. 2017

“Ideas on interpreting machine learning.” O’Reilly

Ideas. 2017

KMIS 2019 - 11th International Conference on Knowledge Management and Information Systems

418