TECH SPARK Summer 2018
How can I help you today? (p8)
THE DATA ISSUE
Is your business intelligent? 04
Artificial intelligence and machine learning
08
Virtual assistants in financial services
30
A hacker’s guide to ‌ artificial neural networks
CONTENTS
EDITORIAL INTRO
3 4 C O N T E N T S
8 18 30 40 2 | TECH SPARK | Summer 2018
EDITORIAL INTRO Wayne Ross, editor and global CTO of Excelian, wishes you a warm welcome to The Data Issue.
ARTIFICIAL INTELLIGENCE AND MACHINE LEARNING A short and simple-to-navigate introduction to the artificial intelligence, machine learning and data science landscape
VIRTUAL ASSISTANTS Most people are familiar with digital assistants, Siri and Alexa. But can chat bots give your business the edge?
UNRAVELLING DATA VISUALISATION
We examine and uncover the main differences between data visualisation and information visualisation
A HACKER’S GUIDE TO … ARTIFICIAL NEURAL NETWORKS
Tech Spark gets under the bonnet of artificial neural networks in finance for this issue’s Hacker’s guide to …
I would like to welcome you to this edition of Tech Spark – The Data Issue. Digital transformation will increase the volume of data your business generates. This increase in data will originate from your business interactions and more importantly will be generated as a function of operating your business in the digital space. An intelligent business knows how to harness this data to transform, present and combine it with other data sources in meaningful ways, which can be acted upon to provide insights into how to optimise your business and to provide better customer interactions. The advance of artificial intelligence (AI) and machine learning (ML) technologies presents further opportunities to intelligently automate business interactions, driven by this data. This edition presents our view on where AI/ML might be applied to data within financial services, with a focus on virtual assistants. We discuss how to enhance your experience with virtual assistants and to provide an example of how virtual assistants might operate internally to present you with intelligent outcomes. I hope you enjoy this edition of Tech Spark as much as we’ve enjoyed exploring these technologies. Andre Nedelcoux MD, Head of Digital Consulting, Excelian Luxoft Financial Services andre.nedelcoux@excelian.com
THE EXCELIAN SERVICE CATALOGUE See how Excelian can tackle complex financial services at scale and accelerate digital transformation programmes …
TECH SPARK | Summer 2018 | 3
ARTIFICIAL INTELLIGENCE AND MACHINE LEARNING
apacity, c e v i t i n g ludes co c n i g n i learn LP) and e N ( n i h g c n i a s s M delling. o ge proce a m u s g s n e a l natural vel busin e l r e ssistant h a g l i a h u r t r o i f s build a v o platform t r e h t e this tog Bring all
4 | TECH SPARK | Summer 2018
n o e s r u o c Dis method the TECH SPARK | Summer 2018 | 5
ARTIFICIAL INTELLIGENCE AND MACHINE LEARNING
D
ine ce, mach ata scien re a and AI learning d t are use a th terms gs in th t n re ean diffe m t hort u s b ill give a ngeably intercha ere we w H . le t p u o o nt pe e with to differe landscap let’s into the n ns. First, o ti io c it u n d fi e d intro n o an p m d in hu o hung u intereste y ll getting to a e re th t o r and fo we’re n at AI is, be clear, ed. realise th e W specialis . ly n h o emulati ill be, hig tter w e b re , e tu m ble fu , or beco e b foreseea s y a umans – rs will alw k than h Compute given tas a g . d in e be fram at perform rn ill always thms lea w ri k o s lg ta but the I is that a A o h d it w to rence imately The diffe data, ult cally with the e lv o e specifi v b e and er than th ta. ra a r, d e e tt be tion of th the task d in isola e m m ra pre-prog
lligence
l inte a i c fi i t r A ies t i n u t r o opp tries s u d n i l l in a ead r p s e d i are w tly n e u q e s and con ervices ls financia architect o must als at the or for AI – gent i l l e t n i t s very lea n tio automa 6 | TECH SPARK | Summer 2018
VIRTUAL ASSISTANT le r availab ng powe ti u p ore m m o makes w the c data age Right no e th f o t adven ata … and the first on d eflecting R . le ib s pos ring data e is secu c n ie c s ave atures h rt of data at data fe A big pa h w ns. g io in is d erstan nd dec and und insight a e v a ri d e k to a ntial n to m the pote t a huma ns ’t expec n. Huma n ld u o w rm fo atio in You t u o s h s acros ision wit nnection solid dec aking co m y to AI. b s e e li add valu ame app s e th iness d – an side bus data sets eaning in m and r d fo o g okin understo When lo ds to be . e e ta n a t d x e nte rnativ data, co g for alte and ns lookin a e business m r t e a id th w e (i.e. th m es fro m people This com nities, fro u s) m w m e o n c ent ons and governm municati m ). o T c , Io ia ed cluding social m world (in ined … the real m o can be m fr it d e n c a la p in data is Once the hine s to mac proache p a y n a m uction, There are ugh ded stics and (ML): thro g in r, rn o a rr stati e le d n a st-fit l ons, tria duce be dels pro connecti o M orks r. u tw o e ural n neighb nearest ssion, ne re s g m re th h ri g o throu adly, alg formula cted hes. Bro e c p a x ro e p d p n ra put a and othe h given in iven ained wit tr or just g e ) b g n in a rn c a le d ries e upervis d catego output (s tterns an a p en e th in eterm sons are data to d ing). Les rn a le d e is (unsuperv . new data to d e li p ap is domain, , like any L M itive f n o g o n licatio ve: 1) c The app xt we ha te er n th o o c d is n In th ages a im , h layered. c e e p n and to map s gorisatio capacity text, cate to to t ts x u te p to map sensor in ue and NLU g P lo L ia N d ) , 2 values r signals fo t, x d te n n s) co ration a (busines ge gene d langua n a usiness s b m l e te sys igher lev h r fo s ng all this rm lling. Bri e 3) platfo d o m d lysis an tant. data ana tual assis uild a vir b to r e togeth
AI/ML BIG DATA DATA ACQUISITION (ALTERNATIVE DATA)
RPA
DATA SCIENCE CLASSIC MACHINE LEARNING
NLP SENTIMENT
LUIS/BOT (MICROSOFT)
REWARD BASED
LEX (AWS)
DIALOG SYSTEMS
WATSON (IBM)
DEEP LEARNING
QUERY
25+ OTHER PLATFORMS
DIALOGFLOW (GOOGLE)
SUMMARY WATSON TONE ANALYSER
VISUALISATION
COGNITIVE VPA
VOICE OTHER
know Want to cial out artifi more ab arning achine le e and m c n e ig ll t inte ld benefi they cou and how il ma iness? E your bus om xcelian.c ynard@e marc.ma
WIT.AI (FACEBOOK)
NLG
ROBO ADVISORY
TECH SPARK | Summer 2018 | 7
8 | TECH SPARK | Summer 2018
ou have probably been frustrated, if not infuriated, by an automated customer service system. Many of us still have preconceived opinions of their uses. Most of us can recall orbiting an automated menu system or trying to trick an automated system to give up sending standard text replies so that you gain the attention of someone that can actually help. Unsurprisingly the chat bot technology has evolved and is now ready for mainstream business adoption. While retail banking is already investing heavily in dialogue system technology, particularly on mobile devices (in part driven by start-ups and PSD2 legislation), adoption in investment banking has been slow. The opportunities to improve
I can do for you …
me show you what
A well-trained dialogue system can establish intent and execute a service. It can also read sentiment, access client and market data, make connections and anticipate a client’s needs far better than a person – meet your new hard-working, round-the-clock VIRTUAL ASSISTANT …
VIRTUAL ASSISTANTS IN FINANCE
Turn over and let
VIRTUAL ASSISTANTS IN FINANCE
the client experience are widespread and those who choose to defer investment in virtual assistants (VAs) will inevitably become less relevant. A VA can monitor the tone of a client, will have instant access to all the client’s activities and can draw on wider and alternative data to infer client and market signals: anticipating their needs, returning information, making recommendations and executing. Even if an institution is built on relationships and personal trust, and insists that this is a core value, a VA still has a place helping its human co-worker. In the front office, for now, chat systems are still just an embedded RFQ (request-for-quotation) and execution channel for many salestraders. Turn the page to see how this is a natural domain for the virtual assistant/ client working relationship.
TECH SPARK | Summer 2018 | 9
VIRTUAL ASSISTANTS IN FINANCE User/virtual assistant example dialogue via smartphone ‌
&-/0 ( ..&./ */ %0 3
01
(( +*
%0 3
-")&0) #+- )+*/% ./-&'" / #+-2 -! ,-& " !"(/ &. /
%0 3
02
*! / )+*/%. %0 3
-")&0) #+- )+*/% ./-&'" / #+-2 -! ,-& " !"(/ &. /
"-" -" .+)" ,-")&0). #+- )+-" ./-&'".
%0 3
03
STEP 01 • Client makes an initial trade enquiry • System identifies likely product as a call option in gold, but doesn’t quite have enough data; strike and expiry are missing. • System checks appropriateness and client status, all good, assumes this is a buy and continues • System looks at past trade patterns for the client, and decides the most probable is a 12 month expiry and strike close to the forward price, assumes these parameters, calls pricing system to get price and continues ‌
, /+ -" +-! 1+& " )".. $"
"- 0
04
,* **# * , +,*#$ /" , /(-% &0 ,(, % )(+#,#(' (' "- 0
'$$ (". "& $ ( $$ *
-"*! '-& & '+* *
COMPONENTS CALL OUT TO: • Client profile • Picks up price from message bus • Presentation system STEP 03 • Client asks for a knock out at 1280, system checks that this is appropriate, and sees client has a mandate and has traded knock outs before. We’re still not sure on strike, so table is retained
$*
%%
"- 0
05 STEP 02 • Client was looking for a 12 month expiry, system now decides to expand response, and in accordance with customer profile determines an appropriate table format, and so offers a table of prices in a narrow strike and delta space
#*,- % ++#+, ',
COMPONENTS CALL OUT TO: • Product reference • Client profile • Picks up price from message bus • Compliance system
COMPONENTS CALL OUT TO: • Now it is a bespoke product the call out is to price/risk
10 | TECH SPARK | Summer 2018
*("#
*! &# '*+ ' +0/ / %0 3
' /" , " )) '+ # )*# + !( (/' 0 "- 0
$*
%%
"- 0
"- 0
COMPONENTS CALL OUT TO: • Natural language query engine • Double entry position system • Scenario system – adjusting client position • Presentation system
STEP 05 • Client asks two more questions COMPONENTS CALL OUT TO:
' . !
( ") , $ (
STEP 04 • Client asks more questions
• Natural language query engine • Scenario system – adjusting market price • Presentation system enriches reply
) ,( * (* .(# & ++ !
Up until this point I have saved my user five minutes. But, more importantly, I have been exceptionally responsive to our client ‌
TECH SPARK | Summer 2018 | 11
12 | TECH SPARK | Summer 2018 TECH SPARK | Summer 2018 | 13
dialogue concludes …
how the rest of our
Flip the page to see
In businesses that are essentially virtual – isn’t the advent of the virtual co-worker inevitable?
VIRTUAL ASSISTANTS IN FINANCE
(023 + 11(12 -2 '3 7
08
*134", 22*23".3
' -*1 /+$ 1$ /./ 2'$ 4$& + ##$0 .- ,7 /.02 + # 1' -# 1$-# ,$ 2$62 (% 4.+ .- 0(1$1 !.4$ '3 7
.-$ )4 "8
06
/4,% 8/4 ,*+& 3/ ,//+ "3 )&%(*.(
' 2 !.32 2'$ ./2(.- 2. 0$#3"$ $6/.130$ 2. (-"0$ 1$ 4.+ 2(+(27
09
COMPONENTS CALL OUT TO: • Portal/dashboard
)4 "8 '3 7
STEP 06 • System puts trade idea on client stack • Margin workflow engaged
. " -*.43& 6)"3 )"00&.2 3/ -"1(*. /. 3)& )4 "8
)&1& *2 ./ *.*3*", -"1(*. %4& '1/- 8/4 /. 3)& 041$)"2& /' " $",, /03*/. & $".;3 .&3 8/41 &70/241& 3/ 42 "("*.23 ".8 /' 8/41 &7*23*.( 0/2*3*/.2 & 6/4,% )/,% *.*3*", -"1(*. ".% 5"1*"3*/. -"1(*. %4& 3/ 8/4 *. " 2&(1&("3&% "$$/4.3 *. ,*.& 6*3) /41
' 2 #. 7.3 13&&$12 '3 7
.8 " ++ ./2(.- 2 1/.2 /0$,(3, -# .8 %.05 0# -.2$ 2'$ %.05 0# 5(++ !$ 13!)$"2 2. 25. 5 7 , 0&(- '(1 5(++ '$#&$ 7.30 #$+2 -# 4$& /.1(2(.- 2 "300$-2 /0("$1
COMPONENTS CALL OUT TO:
• Trade stack • Margin system
07
$1
+ ;,, 31"%& 231*+& %/6. ".% /43
'3 7 )4 "8
/ 8/4 "1& 041$)"2*.( " /.& 8&"1 #"11*&1 $",, /03*/. /. /9 "3 /9 231*+& %&,3" 6*3) " $/.3*.4/42 %/6. ".% +./$+ /43 "3 /9 "3 3)& 01*$& /' /9 (*5*.( " 01&-*4- /'
(1&&% )4 "8
!&2 3)".+2
STEP 07 • Client executes.
" ++ ./2(.- .- .8 2
.8 120(*$ #$+2 2 2'$ /0("$ .% .8 &(4(-& /0$,(3, .%
COMPONENTS CALL OUT TO:
"+&. :
&0$$# '3 7
$1 '3 7
*$- 9
)4 "8
,, %&3"*,2 )"5& #&&. &-"*,&% 3/ 8/4 "3 789 "#$ $/10 $/-
• Credit system • Pricing system • Margin system • Execution • Confirmation
! %.05 0# .- .8 2
.8
• Credit system check • Execution • Confirmation • Sentiment analysis – noticing the word
thanks, the sentiment engine detects the client is happy
)4 "8
STEP 09 • System recalls trade idea on client’s stack and margin enquiry.
COMPONENTS CALL OUT TO:
'3 7
)4 "8
STEP 08 • Client asks for dashboard enhancements
TRADE CONFIRMED Reference code and all relevant info is emailed automatically to the user
'3 7
++ #$2 (+1 ' 4$ !$$- $, (+$# 2. 7.3 2 678 !" ".0/ ".,
TRADE CONFIRMED Reference code and all relevant info is emailed automatically to the user
"0 3/ 1&$/1% " 5/*$& -&22"(& / 2. 0$".0# 4.("$ ,$11 &$
Overall I have saved my user 10 minutes but I have also directly fielded all client enquiries!
14 | TECH SPARK | Summer 2018
TECH SPARK | Summer 2018 | 15
VIRTUAL ASSISTANTS IN FINANCE IN SUMMARY AI technology in human-computer interfaces is often regarded as having three discrete layers:
16 | TECH SPARK | Summer 2018
marc.maynard@excelian.com
business please email
An “utterance” is anything the user says (or types), for example,
these could help you or your
assistants in finance and how
To find out more about virtual
with actions and prices
and prompting the sales-trader
– automatically parsing text
sales-trader correspondence
assistants to augment
are already using virtual
forward-thinking banks
It may be early days but
Converting voice to text, or extracting the utterance “Give me the price of Apple”
Mapping the utterance to intent
The “intent” is something that the user wants to do, as recognised by the system. For example, the system might have tens (or thousands) of intents, one of which might be “ShowListedStockPrice”
Response
This is what the system does in response to a user intent. For example, given a ShowListedStockPrice(Apple) intent, the system’s response might be to display the current price of Apple, or to plot a price chart
Most of the heavy lifting is in the first two layers – voice recognition and interpretation of the text – and there are plenty of third-party cloud services that can manage the necessary infrastructure and natural language processing very effectively - for example LUIS, Watson, Lex and DialogFlow. This leaves the system designer with the job of determining which user intents the system should recognise and respond to, and defining the appropriate responses to those intents. This is usually a non-trivial task for anything but a very simple chat-bot. (There also some subtleties that we have not covered here – in particular, context is almost always important when we are looking at a conversational system). The payoff for all this work is that a rich dialogue system can add tremendous value, both to clients as well as internally. A good dialogue system can present additional relevant information, promote additional relevant services and initiate execution quicker than a human operative. For example, when responding to a request-forquotation (RFQ), the system may realise that historically the client has always asked for three quotes and helpfully returns the other two without being asked; if the client trades on the RFQ, it might ask if the client wants an updated exposure or limit report, and provide it if necessary; alternatively the system might also present some further trade recommendations or analysis, or hand-off to a sales person or other human operative. In summary, we can see that a critical success factor for our virtual assistant is a well-analysed and well-designed utterance/intent/ response system. However, a further critical success factor is a welldesigned and integrated back-end layer for the virtual assistant – it is crucial that the virtual assistant receives all relevant information smoothly and seamlessly. Design and integration of the back-end layer is often a significant task. It is difficult to anticipate the path that technology will take and how different sectors and businesses will evolve. Nonetheless, it seems safe to assume that much of the world of finance and banking – in many instances, already a virtual business – will see the advent of the virtual co-worker, and sooner than we think.
TECH SPARK | Summer 2018 | 17
Unravellingth emeaningofda tavisualisation infoursteps DATA VISUALISATION/INFO VISUALISATION
DATA VISUALISATION/INFO VISUALISATION
What do we mean when we talk about Data Visualisation?
Such a common phrase has many different interpretations.
Our interpretation is formed around understanding the difference between data and information visualisation …
18 | TECH SPARK | Summer 2018
TECH SPARK | Summer 2018 | 19
DATA VISUALISATION/INFO VISUALISATION
DATA VISUALISATION/INFO VISUALISATION
FAST FACTS
D
What is data visualisation?
ATA VISUALISATION is a common term which is interchangeably used to mean very different things. Depending on who you speak to, it can have a very different purpose and execution. Common definitions include:
The visualisation layer put on top of datasets to make sense of the data The tool used to visually interrogate and analyse data and from which to gain insights The process of displaying data information in graphical charts, figures and bars Infographics and visual representation of complex data An automated visual design to provide predictions So it’s a layer, a tool, a process, an infographic and an automated visual design? It can be all of these and none … and that’s the problem. Data visualisation means different things to different people and industries. It’s confusing!
VISUALISATION IN BROAD TERMS Visualisation is powerful. Every visualisation we see is likely to give us some new insights, points of view and knowledge. Visualisation is ultimately based on data or information and can either be raw or curated. Some of those insights might be already known (but perhaps not proven) while other insights might be completely new or even surprising. Certain visualisations may give an incorrect or biased skew, or may misinform and actually make data more difficult to understand. If it’s the wrong visualisation for the data, or it’s badly executed, it means certain insights, and intelligence, can be lost or be inaccurate – which can do real damage to your credibility and reputation. Your data is only as good as your ability to interpret and communicate it, which is why having the right type of visualisation approach is critical. It’s not practical to expect that data visualisation tools and techniques will unleash a series of ready-made stories from datasets. They can only take you so far. There has to be some interpretation and analysis applied. Firstly, it makes sense to look for insights, which can be woven into stories and, in order to do this, you have to analyse the data, search for patterns, identify trends and discover connections. This can be done manually and by using visualisation tools. Secondly, how do you communicate those insights, define the stories you want to tell and the relationships you are looking to show? Knowing this information will help you define and design the visualisation to best deliver your message.
20 | TECH SPARK | Summer 2018
90% 90% of all information transmitted to our brains is visual
80% People remember
80% of what they SEE
20% BUT only remember
20% of what they READ
TECH SPARK | Summer 2018 | 21
DATA VISUALISATION/INFO VISUALISATION
DATA VISUALISATION/INFO VISUALISATION
Two of the most popular types of visualisation are:
Because the use of graphical data visualisations is growing so quickly, there are a lot of different views about the difference between a data visualisation and an information visualisation.
DATA VISUALISATION and INFORMATION VISUALISATION (often known as infographics and/or information design and can be static, interactive or animated).
Data visualisation versus information visualisation
The process that is used to create them could be applied to another data set They are likely to be objective presented without context and not edited
Information visualisation characteristics
and take a back seat to the data itself Data analysts will be more interested in the Gestalt principles (proximity, similarity, continuity, closure, and connectedness) rather than an editorial or thoughtleadership perspective
The visualisations focus on patterns and trends to allow analysts to determine insights
Often the raw datasets are so vast and unwieldy that its almost impossible to process without some sort of automated pattern analysis
The graphic design of the visualisation (the look and the feel) will be deliberately less obvious
The visualisation itself may take the form of more traditional library charts
The data will be qualitative, and therefore provided as informational
They are sensitively designed and presented in a way that is accessible for an audience
They are self-contained and discrete
They have an editorial focus or thought-leadership perspective
The information is presented in the context of the story it is trying to tell
The visualisations are often illustrations, iconography and other graphical representations to illuminate the information
The visualisations focus on telling a story around the insights The graphic design enhances the story and improves understanding of the information for a specific audience
The visualisation process Step 2 – Information visualisation
Step 1 – Data visualisation 1
2
3
4
5
6
7
8
Know your audience
Raw unrefined data
Organise raw data
Analyse patterns and trends
Refined data (information)
Themes (story to tell)
Visualise story
Share knowledge
Data that hasn’t been processed for use
Small Large simple complex data sets data sets
Using analysis tools or manual interpretation
Data that is specific and organised for a purpose
Using a relevant and meaningful story to aid understanding
Using illustrative techniques to bring the story to life
Presented in a format that is easy to share and gain momentum
+
Know the business problem
Sort by human
Sort by machine
TYPICALLY EXPRESSES ONE RAW DATA SET, USES PRE-CANNED VISUALISATIONS
22 | TECH SPARK | Summer 2018
In the real world these steps maybe more intertwined depending on the circumstances.
You may believe that the definition is clear but when you get into more complex
Data visualisation characteristics The data will be quantifiable, and therefore in the form of numbers
visualisations, you can start to wonder. We tend to see them as two steps in the same visualisation process and use the following model to explain the differences.
TYPICALLY EXPRESSES MULTIPLE CURATED DATA SETS IN A STORY, USES ORIGINAL ILLUSTRATIONS
TECH SPARK | Summer 2018 | 23
DATA VISUALISATION/INFO VISUALISATION
Know your audience, visualisation purpose and key message …
DATA VISUALISATION/INFO VISUALISATION
24 | TECH SPARK | Summer 2018
TECH SPARK | Summer 2018 | 25
DATA VISUALISATION/INFO VISUALISATION
AI and ML will do all the visualisation work, right?
DATA VISUALISATION/INFO VISUALISATION
Not quite. It depends on what the purpose is for the visualisation. How will AI and ML know what a good or bad visualisation looks like to the user? There is a human/ social aspect involved as much as there is a technological one. Machine learning (ML) is the science of making predictions based on patterns and relationships that have been automatically discovered and recognised in data. Based on teachings, ML can look for patterns from the raw data (input) and create a visualisation (output) that helps users to understand the next level of insight. The visualisation types need to be recognised, understood and applied appropriately through machine learning. Applying an incorrect visualisation type can give you inaccurate insights. So, in order to thrive, the machine learning model should be aligned to human storytelling and needs. It must be able to understand the difference between patterns WIndow to buy for communicating certain XAU($) insights. Every aspect of ML needs to be fuelled and managed by human judgement. 84.3 69.2 LAST PRICE
CURRENT
The virtual assistant example If a trader asks a VA for information on a trade, via a conversational interface, what is the best way for the VA to visualise the information back? In 90 per cent of cases, this would be a text response, but more detailed information could benefit from being graphically visualised, e.g. trade price prediction for XAU (see below, left). But how does an ML model know it’s the right visualisation for a trader to understand? An ML model needs to find out just how wrong it can be about the importance of the data in order to be as right as possible, as often as possible. This has to be carefully considered as the wrong trading decision based on an inaccurate prediction visualisation could lose millions. If you look at visualisation as a series of tasks/ smart tasks, without a doubt AI and ML can operate in that area. If you look at visualisation as strategic decisionmaking activity it’s a totally different spin. Unorthodox thinking, creative imagination, human experience, intensive interactions with clients and co-workers (soft-skills) are activities you can’t automate. Not yet anyway. 99.2 PREDICTED
BUY 26 | TECH SPARK | Summer 2018
TECH SPARK | Summer 2018 | 27
DATA VISUALISATION/INFO VISUALISATION
DATA VISUALISATION/INFO VISUALISATION
KEY FUNDAMENTALS OF A SUCCESSFUL DESIGN Consider your audience
The visual composition of your design is key …
Final points
Tell a story with your data
Make your
design accurate
80%
Organise your
design and tidy it up
Incorporate as
many guidelines as needed
People remember 80% of what they SEE Use colour
deliberately
Want to know more about how visualising information and/or data could benefit your business, please email andrew.hall@excelian.com 28 | TECH SPARK | Summer 2018
There are also some extra details that are worth paying attention to. The composition of your design is key — don’t just throw stuff in a blank page, make sure it is arranged as best as possible. Choosing a good font makes all the difference — avoid the use of fonts such as Times New Roman or Comic Sans as they look a bit careless, transmitting the wrong idea to your audience. Rather, choose fonts that are easy to read. Colour shouldn’t be used only to embellish your work – it should have a functional
purpose. For instance, it can be used to accentuate a certain detail. In the image above, the colour saturation used makes it easier for the user to differentiate between the higher and lower numbers. Finally, I have to mention storytelling. Storytelling is all about communicating efficiently with your audience, giving them context to what they’re seeing/reading and putting yourself in their shoes. You have to consider the audience’s profile (who you’re targeting) and what is the most relevant and interesting information for them.
20%
Choose fonts that are easy to read
Create the
most appealing BUT only remember composition 20% of what they READ
TECH SPARK | Summer 2018 | 29
ARTIFICIAL NEURAL NETWORKS
ARTIFICIAL NEURAL NETWORKS
A HACKERS GUIDE TO …
IN FINANCE
INTRODUCTION In recent years, inspired by their biological counterparts and the relatively recent increase in computing power, artificial neural networks have taken centre stage in an increasing number of industries. The increase in data collection has raised our desire to uncover relationships within our networks that could help in the exploitation of new opportunities. Neural networks give us the ability to construct extremely complex models from this data and provide a mechanism to predict future events. Artificial neural networks allow us to create an internal representation, or model, of a solution using a set of input scenarios and expected
30 | TECH SPARK | Summer 2018
outcomes. Instead of the hard work of creating a complex model using multiple input and output parameters, artificial neural networks construct a model by a process of learning which tunes the model to the expected outcomes we know about in much the same way a biological brain would. There are, of course, many differences between the biological counterpart and artificial networks – one of which is the dynamic nature of biological networks – connections are constantly being created and destroyed. Currently they only replicate the very basic structure of their biological counterparts but are coming up with extremely interesting results. Neural networks can solve many problems
that can also be solved by other existing algorithms (decision trees, k-nearest neighbour, etc.), but where they become particularly powerful in model creation is when there is a large number of inputs with a large number of predictions. So why use neural networks at all? In a world of ever expanding information where companies have large sets of structured and unstructured data, often held in vast enterprise stores, they often have an enormous task to uncover relationships hidden within. Finding relationships between data sets can be achieved through statistical analysis and often brute force searching but the resultant model is then only as good as the
TECH SPARK | Summer 2018 | 31
ARTIFICIAL NEURAL NETWORKS
ARTIFICIAL NEURAL NETWORKS
creator’s model. Neural networks will allow us to capture and model undefined relationships on our behalf. If we can train the network to a tolerable threshold then it might be a suitable model for predicting future results given new input data. There are plenty of frameworks available, most of which require a good understanding of neural networks at a high level with multiple levels of configuration to suit your data. However, in this article, we will chart our journey through the building of a rudimentary neural network to uncover its challenges and rewards. BUILDING THE FRAMEWORK To better understand how a neural network works, we decided to build one from scratch to understand how it fitted together and what impact different data had on outcomes and performance. By trying out a real set of sample data we wanted to understand the entire process: from manipulation of input data and the tuning of the weights, to the understanding of the results. Our goal was to construct a system that would discover the complexities hidden within our data without having
Fig 1
Input Layer
Hidden Layer
Output Layer
to understand them or discover them ourselves. We have data that, given an observed set of inputs, sees the model change to match the given set of outputs. We therefore wanted to create a system that will construct a model representation of our data, however complex, so that we can use it in the future with new data to predict an outcome that would have otherwise been very difficult to have achieved. Our system is a supervised system, which means we provide the inputs and the expected outputs. UNDERSTANDING THE FRAMEWORK Our first task was to discover what the standard neural network framework would need to contain. Each input corresponds with a unique parameter in our data and is represented by a single input neuron within the input layer. Each output in turn corresponds with a neuron in the output layer and represents an expected result given the input data. In between is a model representation of our specific dataset and the complex relationships within. See Fig 1 FRAMEWORK COMPONENTS At a basic level the system contains layers and neurons: Neurons Objects that associate input and output connections and house a mathematical function Layers A set of neurons Connections Inbound and outbound weightings
· ··
The mathematical (activation) functions that are invoked when the system is in training mode come in many different guises, each useful in their own way, depending on the data as we see later: Rectified linear unit (ReLU) Hyperbolic Tan Linear Binary step Softmax
·· ·· ·
There is always on a single input layer
32 | TECH SPARK | Summer 2018
The structure of an artificial neural network is actually based on our understanding of how a human brain functions
TECH SPARK | Summer 2018 | 33
ARTIFICIAL NEURAL NETWORKS
Getting started with neural nets – an opinionated roadmap Either get the basic training … Try some of the online courses – EdX Read the books – Neural Networks and Deep Learning, for example, or Ian Goodfellow’s Deep Learning … or jump straight in with Python … Get Anaconda, use its conda env (environment management) and Jupyter (widely-used notebook for data science). Alternatively, use Jupyter on Azure or AWS. Start with existing libraries … Keras is an easy-to-use Python library which uses other well-known libraries (like Google’s TensorFlow). There are many examples (keras-team/Jupyter Notebook) where datasets are included Scikit-learn also has many examples and broad functionality H2O.ai and MATLAB are two of the better-known libraries in a broad field … but is the project suitable for neural networks? Data availability you have thousands/ millions of observations … neural nets often come into their own at large scale Well-defined evaluation function you have a well-defined problem - for example, “predict mortgage pool prepayment rates with less than five per cent error” Combinatorial explosion You have many different inputs, outputs and either unknown – or complex – interactions between variables. Neural nets are relatively strong in these situations Critical success factors Data preparation/analysis often overlooked, but crucial Well-defined evaluation function a vital prerequisite: define the problem Manage bias and variance are you poorly-, correctly- or over-fitted? Your bias and variance will help tell you … Smooth integration productionising neural nets does need some work …
34 | TECH SPARK | Summer 2018
and output layer at least one “hidden layer” housed between the input and output layers. See Fig 2 Our build started by first constructing the framework to hold the multiple layers of neurons. The input layers designed to take on the rows of input data, the output layer for the expected data and a system to insert multiple hidden layers in between. The number of hidden layers and their number of neurons will be determined later. Each neuron within the hidden and output layers has a concept of a weighted connection which houses the weightings between neuron and the next. See Fig 2, left to right Training the network using our dataset requires tweaking each weighting within the network so that whenever we introduce a new set of input values, the resultant outputs will be as we expect. Training the model can be done from left to right or from the input layer through the hidden layers to the output layers by tweaking the weightings incrementally each iteration. This mechanism is called feed-forward. However, it is very slow and prone to either not converging or, at the very least, proving very expensive computationally. Examples include Extreme learning machines where parameters and weights of hidden neurons do not need to be tuned but instead randomly assigned. Back propagation (or back propagation of errors) was introduced which, using various activation functions, works from the output layer back to the input layer. Making use of back-propagation hugely increases the chances of convergence. Whilst there are many activation functions, we chose to build only a few, which we mention above. In our implementation we decided all neurons within a specific layer would have the same activation function and that each layer could have different activation
functions. Activation functions simply sum all the input weights, adds a bias and then determines how the neuron should fire. The weights on these neurons are what encode the knowledge (or model) of the system. PREPARING THE DATA Our dataset contained values where a combination of input values resulted in an expected category. In this example, we found that the best results were attained by ensuring that the output values or expected results were binary, so either 0 or 1. It was therefore very easy to identify if the values were near to 0 or to 1. We found that if a single output value was used that had a range from 0 to 1 then it was often difficult to interpret the result when a prediction was attained. For final validation of the model it is important to take out a random set of training data for use as validation. LEARNING STRATEGIES We decided on supervised learning because we had expected results with our input values. It also appeared to be the easiest way to see if our framework was doing its job correctly. Our framework was to be tested using two different sets of data. The first set was very simple with two inputs and a single output and would emulate the Exclusive OR (XOR) gate. The two inputs are either 0 or 1 however the result is a 1 if the input values are not the same as each other and 0 if they are. Before we could start our training our network we had to choose which activation function we were going to use as initially we wanted to restrict it to just one for all layers. Our initial choice was the Sigmoid function Fig 3 which has a gradient for most values other than at the tails where the learning rate decreases as the derivative tends to 0 asymptote, this issue is known as the vanishing gradient problem and can cause the learning rate of the network
Fig 2
Hidden Layer 1
Input Layer
Hidden Layer 2
Output Layer
Fig 3
Sigmoid
1.5 1 0.5
-5
-10
0
5
10
Fig 4
ReLU and Softplus
5 4 3 2 1 0 -6
-4
-2 ReLU
0
2
4
6
Softplus
TECH SPARK | Summer 2018 | 35
ARTIFICIAL NEURAL NETWORKS
ARTIFICIAL NEURAL NETWORKS
HOW ARTIFICIAL NEURAL NETWORKS CAN RECOGNISE PRECIOUS METAL TYPES AND ADVISE WHETHER TO BUY OR SELL ‌
Introducing ReLU, while still slower than we hoped, increased the convergence rate and allowed us to train our network faster and therefore allowed us to work with larger amounts of data.
ES TI XI
uy B % 90
10
0
ll
tpu
u 6O
ase
h tP
Se
on
%
0
eC
as Ph
iti gn
co
e 5R
OUR LEARNING EXPERIENCE By using the XOR function we were able to construct a dataset. This was also the case for the unit circle example and both XOR and unit circle were improved by the ReLU activation function.
LE
Ph
cog
e
R 04
on niti
B ase
P
P
ET A
IL
S
C
cog
e 3R
0
on niti
M
In
P
A
O
02
t pu
e has
C
0
e has
D
in
rain 1T
O
LO
P
U
R
al la PA diu LL m
Si l SL ver V
G o G ld LD
P
la t P inu P m LT
However, in our XOR example, we found that it didn’t perform as well as we had expected. If the values were close to 0 or 1 then they were less likely to be changed by the activation function and convergence was slow so we tried a different one. Our next choice was rectified linear unit (ReLU). See Fig 4, p35
ase
h gP
01
02
03
04
05
06
TRAINING PHASE
INPUT PHASE
RECOGNITION PHASE A
RECOGNITION PHASE B
RECOGNITION PHASE C
OUTPUT PHASE
exposed to multiple
of a precious metal
respond to different
to more complex
to highly complex
what the object most
A network is
labelled images of different types of
precious metal in
An unlabelled image is exposed to the
pre-trained network
order to classify
36 | TECH SPARK | Summer 2018
to greatly slow if not stall entirely. This is important in the back propagation to tune or tweak the weightings.
The neurons
simple parameters such as colour
Neurons respond structures in this case perhaps
individual detailing
Neurons respond
abstract concepts that we would
identify as different types of precious metals
The network predicts likely is, based on its
training, and advises whether to buy or sell according to
In the two examples above, our data was created using the two functions. We then moved onto a simple categorization example where our dataset consisted of approximately 10 inputs and a single categorisation or output. We again used the Sigmoid function but found that the training was outperformed by using ReLU. Our datasets were relatively small. If they increased into very large datasets, then training the network would take longer to perform and there may be a very real chance training might not complete in a manageable time. It is at this point that we choose the use of stochastic gradient descent which employs randomisation of mini-datasets to achieve training success. This is a popular method for large datasets and large model size and has become industry standard. Finally, we also had to work out what
Our goal was to construct a system that would discover the complexities hidden within our data
latest price info
TECH SPARK | Summer 2018 | 37
ARTIFICIAL NEURAL NETWORKS
our learning rate would be. This is the rate to which we climb or fall down the gradient. If the learning rate is small then the training will be more reliable, but it may take a long time to converge. However, if the learning rate is too high then convergence may not happen. We therefore started very low (0.0001) and slowly increased it to tune to our dataset. The eventual range was between 0.0001 and 0.01. The last thing to think about is the number of layers. This can depend on the complexity of the relationships within the data. In our case we decided on a single hidden layer but after starting with two neurons, however, we did experiment with additional neurons and layers. We found that the more neurons/layers added the more reliable the learning, although there is a cost in computational time. Deep learning employs the use of many hidden layers and for very complicated models this may be necessary. TRAINED NETWORK Once our network was trained using our dataset, our network was ready to be used for new predictions. The XOR function is not a good example because the entire combination of inputs was used to train the network, so there was no new data to predict. For the categorisation example, we held back some of the input data and their expected results to use as a way of measuring the trained network prediction capability with what we know are expected results. Our results were within tolerance when it came to predicting values that were inside of the training set inputs but as we strayed outside of these values we
ARTIFICIAL NEURAL NETWORKS
found the chances of the prediction being wrong increased. Our second dataset represented requests for quotations (RFQs) and their eventual outcome. The inputs include date, time, traded price, quantity, buy/sell, product, outcome and client. The goal was to predict the likelihood of a customer successfully executing a specific product. Initially, date and time were taken out of the data set but in possible subsequent examples they could be used to determine a day of the week to be used to narrow the prediction criteria. Once the system was trained, we were able to provide it with new trades and predict the likelihood (as a percentage) of a successful outcome. Resultant information of this kind is useful to add to the decision making of an organisation when trades are being done in real-time. DO WE NEED TO UNDERSTAND THE MODEL ANY MORE? Up until this point we have been very keen to understand the world around us by modelling it in as much detail as possible. For simple models that have few parameters with defined intervals, this is achievable. But as the number of parameters starts to increase the complexity can increase exponentially. In the past the model was most important with the data as a poorer cousin. However, in the modern world, data is so vast that the models can be extremely complex to create and understand. Artificial neural networks give us the chance to build models that are immensely complex without having to understand them.
They give us the chance to build models that are immensely complex without having to understand them
Want to know more about artificial neural networks and how they could support growth in your business?
WHERE DO ARTIFICIAL NEURAL NETWORKS FIT IN THE FINANCIAL LANDSCAPE? Most banks collect huge amounts of data and are now placing them in large data lakes. This data is sometimes well understood but often, with any large repository of unrelated data, there is the opportunity to gather it all into a single learning system and see if relationships can be uncovered. Artificial neural networks allow us to tackle this problem by processing large amounts of data and data types resulting in predictions that otherwise could not be achieved in the same time frame. The models can be used to predict financially driven models, for example the likelihood of success of a trade given the customer. Any predictive system is an important tool in an investment bank and used in concert with existing tools allows insights that may never have been conceived without the intervention of AI. SCEPTICISM AND CRITICISMS OF NEURAL NETWORKS There is scepticism when it comes to artificial neural networks and this can often come from trying to create predictions from datasets that sit outside of the training datasets. The prediction can prove to be very inaccurate as the model is very specific to the dataset. Understanding your requirements is therefore paramount when employing an artificial neural network as it is not a panacea. Understanding the model weakness is also important if model reliance is high.
Email paul.hewitt@excelian.com for more information ‌
38 | TECH SPARK | Summer 2018
TECH SPARK | Summer 2018 | 39
OUR SERVICE CATALOGUE
OUR SERVICE CATALOGUE
The Excelian Service Catalogue
ACCELERATED DIGITAL TRANSFORMATION MODEL THE JOURNEY
We are Excelian. We help our customers maximise efficiency, reduce risk and increase speed of delivery by helping them to adopt better practices and more advanced systems …
DIGITAL INNOVATION
DIGITAL TRANSFORMATION AT SCALE
DIGITAL TEAMS
DIGITAL EXPERIENCE
SERVICE OFFERING
Our unique value proposition is that we tackle Financial Services complexity at scale and accelerate Digital Transformation programmes by applying the following approaches: DESIGN-DRIVEN THINKING Agency grade digital experiences produced to solve business problems with a solid design process and immersive and engaging visualisation INNOVATIVE SOLUTIONS ACCELERATION Solution blueprints for UX and engineering accelerate the path to solving innovative and disruptive business problems DIGITAL DELIVERY End-to-end incremental deliveries, always read-to-ship and cloud ready with optimal onshore/near-shore price-points for the client TAILORED FOR FINANCIAL SERVICES Focussed on financial services business problems with secure, compliant, globally-enabled technologies and supported by data centric analytics and AI/ML solutions
CONTACT US If you want to know what Excelian can do for you or your business please do not hesitate to get in touch. Call 020 7336 9595 or email wayne.ross@excelian.com
40 | TECH SPARK | Summer 2018
INSIGHTS
Excelian solves complex problems, at scale.
MARKET INSIGHTS WORKSHOP
CONCEPTS
BLOCKCHAIN EXPLORATION
DATA SCIENCE AND ARTIFICIAL INTELLIGENCE EXPLORATION
VIRTUAL TRADING ASSISTANT
DIGITAL STRATEGY
DIGITAL SOLUTIONS
MACHINE LEARNING FOR CLIENT SIGNALS
DEEP LEARNING FOR CAPITAL OPTIMISATION
DIGITAL ENGINEERING CENTRE
DATA LAKE PLATFORMS
CLIENT CHANNEL FOR MARKETS
WM OMNI CHANNEL PLATFORM
RETAIL BANKING CUSTOMER JOURNEYS
INTELLIGENT WORKBENCHES
ENTERPRISE VISUALISATION
CLOUD COMPUTE FABRIC
DIGITAL CAPACITY
DIGITAL HYBRID TEAMS
DIGITAL ENABLERS
UI TECHNOLOGY
DIGITAL ARCHITECTURE UI DESIGN AND USABILITY
BIG DATA ENGINEERING DATA SCIENCE
DEVOPS DLT AND BLOCKCHAIN
CLOUD AND GRID COMPUTING
TECH SPARK | Summer 2018 | 41
42 | TECH SPARK | Summer 2018 TECH SPARK | Summer 2018 | 43
•
er in
e
th
xt
ne
e
su
is
th ea an es d a W d e f ll o is in m ea f TE su ter ore tu C e o 20 in res H f 1 t 8 h SP e AR K
tc
S s c v os ien erle or om t s cy ss ga e av , fl an nis int in ex d at ere gs ib e io ilit xp ns stin y eri ha g an en v d ce e d R
Th e m g: Th e s ai er n a e ke ve d rle va y be ss nta ne g fit ma es in s of In t p o la se us ad y ou ye rv sw e c diti o r b rs erl in itc as n e c es sig re he s w we us d s a ni se to h lo in er ok fic d es an effi ser e F a s t
• • Th
C ov
In
W e e th x e plo w r or e ld of
I believe people are smart and some people want to share more data than other people do. Ask them, ask them every time. Make them tell you to stop asking them if they get tired of your asking them. Let them know precisely what you’re going to do with their data. Steve Jobs (2010)