Session 3 Upskilling for Predictive Analytics Travis M Short, FSA - - PDF document

SOA Predictive Analytics Seminar – Malaysia

27 Aug. 2018 | Kuala Lumpur, Malaysia

Session 3 Upskilling for Predictive Analytics

Travis M Short, FSA

Pac ific L ife Re

Upskilling for Predictive Analytics

SOA Predictive Analytics Seminar – Kuala Lumpur Travis Short, FSA

—27 August 2018 —

Pac ific L ife Re

The Actuary — What’s the future Actuary?

‘T he Onc e & F uture Ac tua ry’ – Da vid Holla nd, 1997 SOA Pre side nt – “I

n his 1949 a ddre ss a s the first pre side nt o f the So c ie ty o f Ac tua rie s, E dmund M. Mc Co nne y a ske d: ‘What ar

e ac tuar ie s?’”

– … “T

he a c tua ry in re a lity is a so und, pra c tic a l ra the r tha n to o the o re tic a l ma the ma tic ia n a pplying simple princ iple s o f pro b a b ilitie s to huma n a ffa irs in the unkno wn future .”

– “T

his is no t a b a d de finitio n fo r 1949, o r e ve n fo r 1997.”

– “T

he “Onc e a nd F uture Ac tua ry” is the mo de l b uilde r a nd ma na g e r, the fina nc ia l a rc hite c t a nd e ng ine e r, who c a n la y the fo unda tio n fo r a se c ure fina nc ia l future . I t is o urs to inve nt.”

– T

he Ac tua ry ma g a zine “a na lytic s” wo rd c o unt: 0

– T

he Ac tua ry ma g a zine “CD-ROM” wo rd c o unt: 3

2 SOA Pre dic tive Ana lytic s Se mina r 2018 Aug - Upskilling fo r Pre dic tive Ana lytic s

Pac ific L ife Re

Why ‘upskill’?— What’s your ‘data science’ value proposition?

Amy He ine ike , Prime r AI: – “ “…da ta sc ie nc e is a b it o f a ho t

to pic , a nd so I think the re a re a lo t o f pe o ple who think tha t if the y c a n ha ve the ‘ da ta sc ie nc e ’ la b e l, the n ma g ic , ha ppine ss, a nd mo ne y will c o me to the m. So I re a lly sug g e st fig uring o ut wha t b its o f da ta sc ie nc e yo u a c tua lly c a re a b o ut.”

Upskilling a c tua rie s – Wha t suppo rts yo ur g o a ls, a s a n individua l o r a s

a n o rg a niza tio n? Wha t’ s yo ur e nd g a me ?

– Sha ping yo ur pa th fo r upskilling a nd to o ls to

c o nside r

– Ac tua rie s a nd no n-a c tua rie s wo rking to g e the r,

struc turing te a ms fo r suc c e ss

3

https:/ / www.kdnug g e ts.c o m/ 2018/ 05/ 8-use ful-a dvic e s-a spiring -da ta -sc ie ntists.html>

SOA Pre dic tive Ana lytic s Se mina r 2018 Aug - Upskilling fo r Pre dic tive Ana lytic s

Pac ific L ife Re

EY—

• n applications of advanced analytics for insurance

4

https:/ / www.e y.c o m/ Pub lic a tio n/ vwL UAsse ts/ Adva nc e d_a na lytic s_fo r_insura nc e / %24F I L E / Adv-a na lytic s_insura nc e _AUNZ00000335.pdf

SOA Pre dic tive Ana lytic s Se mina r 2018 Aug - Upskilling fo r Pre dic tive Ana lytic s

Pac ific L ife Re

The Actuary: “analytics” word count

5

2012

“Adva nc e d Busine ss Ana lytic s fo r Ac tua rie s is a se t o f to o ls a nd te c hniq ue s use d to de sc rib e , pre dic t, a nd re c o mme nd b usine ss c o urse s o f a c tio n b a se d o n c o nsume r a nd distrib uto r b e ha vio r. I t dra ws fro m ma ny disc ipline s. I t re lie s he a vily o n va st a mo unts o f da ta a nd c o mputing po we r, sta tistic s, mo de ling ,

• ptimiza tio n, da shb o a rd a nd a le rts, ma rke t

re se a rc h, a nd c luste ring . Adva nc e d b usine ss a na lytic s pro vide s e mplo ye rs with insig htful de c isio n ma king a nd a ffo rds the o ppo rtunity to a sse ss a ma rke tpla c e fro m a to ta lly ne w pe rspe c tive .” Ac tua rie s in Adva nc e Busine ss Ana lytic s White Pa pe r, 2012, L isa T

• urville (Cha ir)

SOA Pre dic tive Ana lytic s Se mina r 2018 Aug - Upskilling fo r Pre dic tive Ana lytic s

Pac ific L ife Re

The Actuary: “analytics” word count

6

2018: PA Pilot a nd E xa m PA

“Mo dule 1: Pre dic tive Ana lytic s T

• o ls

Mo dule 2: E ffe c tive Pro b le m De finitio n a nd Pro je c t Ma na g e me nt Mo dule 3: Da ta De sig n, T ra nsfo rma tio n a nd Visua liza tio n Mo dule 4: Da ta E xplo ra tio n Mo dule 5: F e a ture Ge ne ra tio n a nd Se le c tio n Mo dule 6: Mo de l De ve lo pme nt a nd Va lida tio n Within e a c h mo dule , the re we re kno wle dg e c he c ks, e xe rc ise s, e nd-o f-mo dule te sts a nd

• ppo rtunitie s to inte ra c t with o the r pa rtic ipa nts via

a priva te disc ussio n fo rum. At time s, the mo dule instruc tio ns wo uld a sk fo r spe c ific inte ra c tio ns. At a ny time , pa rtic ipa nts we re fre e to use the fo rum to ma ke c o mme nts o r a sk fo r he lp. T he pa rtic ipa nts wo rke d o n a va rie ty o f da ta se ts, using RStudio to pe rfo rm the a na lyse s.”

SOA Pre dic tive Ana lytic s Se mina r 2018 Aug - Upskilling fo r Pre dic tive Ana lytic s

Pac ific L ife Re

Exam PA: Predictive Analytics

7

2018 De c e mbe r E xa m PA

• 1. Pre dic tive Ana lytic s Pro b le ms a nd T
• o ls (R, RStudio )
• 2. T
• pic : Pro b le m De finitio n
• 3. T
• pic : Da ta Visua liza tio n
• 4. T
• pic : Da ta T

ype s a nd E xplo ra tio n

• 5. T
• pic : Da ta I

ssue s a nd Re so lutio ns

• 6. T
• pic : Ge ne ra lize d L

ine a r Mo de ls

• 7. T
• pic : De c isio n T

re e s

• 8. T
• pic : Cluste r a nd Princ ipa l Co mpo ne nt Ana lyse s
• 9. T
• pic : Co mmunic a tio n

T he PA E xa m is a dministe re d a s a five -ho ur pro je c t re q uiring a na lysis o f a da ta se t in the c o nte xt o f a b usine ss pro b le m a nd sub missio n o f a re po rt.

https:/ / www.so a .o rg / E duc a tio n/ E xa m-Re q / e du-e xa m-pa -de ta il.a spx

SOA Pre dic tive Ana lytic s Se mina r 2018 Aug - Upskilling fo r Pre dic tive Ana lytic s

Pac ific L ife Re

Exam PA Syllabus Textbooks

8

F

• ur te xtbooks on sylla bus

– I

SL te xtb o o k – de finite ly (fre e o nline )

– R fo r e ve ryo ne – a n o ptio n to le a rn R – Da ta visua liza tio n … c urre ntly o nline

• http:/ / so c viz.c o /

– Re g re ssio n mo de ling …

• Mig ht b e a re fre she r o ptio n fo r yo u

– No te tha t I

SL a nd the re g re ssio n b o o k a re a lso o n the pre -re q E xa m SRM sylla b us

SOA Pre dic tive Ana lytic s Se mina r 2018 Aug - Upskilling fo r Pre dic tive Ana lytic s

Pac ific L ife Re

Intro to Statistical Learning – videos too!

9

ISL Vide os online – Yo ur ne w c o mmute ma te ria l – https:/ / la g unita .sta nfo rd .e du/

c o urse s/ Huma nitie sSc ie nc e s/ St a tL e a rning / Winte r2016/ info

– On Yo uT

ub e to o ( se e link b e lo w)

https:/ / www.r-b lo g g e rs.c o m/ in-de pth-intro duc tio n-to -ma c hine -le a rning -in-15-ho urs-o f-e xpe rt-vide o s/

SOA Pre dic tive Ana lytic s Se mina r 2018 Aug - Upskilling fo r Pre dic tive Ana lytic s

Pac ific L ife Re

ISL: flexibility vs interpretability

10

Sta tistic a l le a rning – Yo u’ ll le a rn a va rie ty o f

po te ntia l mo de ls

– Ho w muc h do e s inte rpre ta b ility

ma tte r?

• GDPR, rig ht to a n

e xpla na tio n?

SOA Pre dic tive Ana lytic s Se mina r 2018 Aug - Upskilling fo r Pre dic tive Ana lytic s

Pac ific L ife Re

ISL: training and holdout testing

11

Mode lling princ iple s… – T

he g re e n tra ining mo de l a ppa re ntly ha s b e tte r fit – ho we ve r tha t do e s no t ho ld fo r a te sting sa mple

SOA Pre dic tive Ana lytic s Se mina r 2018 Aug - Upskilling fo r Pre dic tive Ana lytic s

Pac ific L ife Re

ISL: practical application of many models…

12

Re g re ssion Cla ssific a tion – L

• g istic re g re ssio n (le ft), line a r

disc rimina nt a na lysis, QDA, K NN

– E

xa mple he re o f ho ldo ut fit dive rg ing fro m the tra ining fit fo r the ‘ g re e n’ mo de l

T re e Ba se d Me thods Support Ve c tor Ma c hine s Cla ssific a tion – PCA, K

• me a ns c luste ring ,

hie ra rc hic a l c luste ring ,

SOA Pre dic tive Ana lytic s Se mina r 2018 Aug - Upskilling fo r Pre dic tive Ana lytic s

Pac ific L ife Re

Data Visualization: A Practical Introduction (ggplot for R)

13

Da ta Visua liza tion – so c viz.c o – g g plo t pa c ka g e a g o to fo r R – No te the ma p c o de b e lo w is

simple r tha n c o de fo r visua ls sho wn to the le ft

SOA Pre dic tive Ana lytic s Se mina r 2018 Aug - Upskilling fo r Pre dic tive Ana lytic s

Pac ific L ife Re

This map was the US Presidential election

14

– Re d (Re pub lic a n) i.e . T

rump

– Blue (De mo c ra t)

SOA Pre dic tive Ana lytic s Se mina r 2018 Aug - Upskilling fo r Pre dic tive Ana lytic s

Pac ific L ife Re

Beyond Intro to Statistical Learning

15

E le me nts of Sta tistic a l L e a rning – F

ro m I SL : “…I n this ne w b o o k, we c o ve r ma ny o f the sa me to pic s a s E SL , b ut we c o nc e ntra te mo re o n the a pplic a tio ns o f the me tho ds a nd le ss o n the ma the ma tic a l de ta ils.”

– E

SL pro vide s mo re tho ro ug h ma the ma tic a l de ta il

Va rious othe r options – Ma c hine L

e a rning : A Pro b a b ilistic Pe rspe c tive : K e vin P. Murphy

– Pa tte rn Re c o g nitio n a nd Ma c hine L

e a rning b y Christo phe r Bisho p

SOA Pre dic tive Ana lytic s Se mina r 2018 Aug - Upskilling fo r Pre dic tive Ana lytic s

Pac ific L ife Re

The Role of the Actuary in Data Science (IFoA ‘MAID’)

16

We fe d a ma c hine le a rning a lg orithm the te xt of our job a dve rts a nd a ske d: (pa ra phra se d) “wha t doe s a n a c tua ry ne e d to be c ome a da ta sc ie ntist?” – Pytho n – R / SAS – Ma c hine le a rning – Visua lisa tio n – ma ths – da ta b a se

SOA Pre dic tive Ana lytic s Se mina r 2018 Aug - Upskilling fo r Pre dic tive Ana lytic s

Pac ific L ife Re

Data Scientist Skills, 4 Types (Udacity blog)

17

“F inding a g re a t da ta sc ie ntist invo lve s finding so me o ne who ha s so me wha t c o ntra dic to ry skill se ts: inte llige nc e to handle

data pr

• c e ssing and c r

e ate use ful mode ls;

a nd a n intuitive unde r

standing of the busine ss pr

• ble m the y’ re trying to so lve , the

struc ture a nd nua nc e s o f the da ta , a nd ho w the mo de ls wo rk, sa ys L e e Ba rne s, he a d o f Pa ytro nix Da ta I nsig hts a t b usine ss so ftwa re pro vide r Pa ytro nix Syste ms.”

https:/ / www.c io .c o m/ a rtic le / 3263790/ da ta -sc ie nc e / the -e sse ntia l- skills-a nd-tra its-o f-a n-e xp e rt-da ta -sc ie ntist.html>

https:/ / b lo g .uda c ity.c o m/ 2018/ 01/ 4-type s-da ta -sc ie nc e -jo b s.html

SOA Pre dic tive Ana lytic s Se mina r 2018 Aug - Upskilling fo r Pre dic tive Ana lytic s

Pac ific L ife Re

Data Scientist Skills, 4 Types (Udacity blog)

18

“F inding a g re a t da ta sc ie ntist invo lve s finding so me o ne who ha s so me wha t c o ntra dic to ry skill se ts: inte llige nc e to handle

data pr

• c e ssing and c r

e ate use ful mode ls;

a nd a n intuitive unde r

standing of the busine ss pr

• ble m the y’ re trying to so lve , the

struc ture a nd nua nc e s o f the da ta , a nd ho w the mo de ls wo rk, sa ys L e e Ba rne s, he a d o f Pa ytro nix Da ta I nsig hts a t b usine ss so ftwa re pro vide r Pa ytro nix Syste ms.”

https:/ / www.c io .c o m/ a rtic le / 3263790/ da ta -sc ie nc e / the -e sse ntia l- skills-a nd-tra its-o f-a n-e xp e rt-da ta -sc ie ntist.html>

SOA Pre dic tive Ana lytic s Se mina r 2018 Aug - Upskilling fo r Pre dic tive Ana lytic s

Pac ific L ife Re

You said there’d be Pandas

19

R vs Python??

Pytho n Data Analysis L ib rary pandas is a n o pe n so urc e , BSD-lic e nse d lib ra ry pro viding hig h-pe rfo rma nc e , e a sy-to -use da ta struc ture s a nd da ta a na lysis to o ls fo r the Pytho n pro g ra mming la ng ua g e .

SOA Pre dic tive Ana lytic s Se mina r 2018 Aug - Upskilling fo r Pre dic tive Ana lytic s

Pac ific L ife Re

Learn SQL & managing data

20

De finite ly SQL

… mo st o f yo ur da ta is like ly in RDBMS whe re yo u’ ll use SQL fo r sto ring , ma nipula ting , a nd re trie ving suc h da ta A le a rning o ptio n: https:/ / www.w3sc ho o ls.c o m/ sq l/ Also c he c k o ut T idy da ta fo r R: http:/ / vita .ha d .c o .nz/ pa pe rs/ tidy- da ta .pdf (o r Pa nda s fo r Pytho n)

SOA Pre dic tive Ana lytic s Se mina r 2018 Aug - Upskilling fo r Pre dic tive Ana lytic s

Pac ific L ife Re

Gartner Magic Quadrant for Analytics and BI Platforms

21

‘Se lf se r vic e ’ BI Platfo r ms

Ac c o rding to Ga rtne r: T a b le a u, Mic ro so ft, a nd Qlik a re le a de rs in this spa c e A so me wha t ma ture spa c e o f so ftwa re pla tfo rms

SOA Pre dic tive Ana lytic s Se mina r 2018 Aug - Upskilling fo r Pre dic tive Ana lytic s

Pac ific L ife Re

Tableau: an example

22

Re vie wing c anc e r c laim e xpe r ie nc e

T

• p ha lf: A/ E

’ s re la tive to a n e xpe c te d b a sis, with the A de c o mpo se d in to the unde rlying c a nc e r site s Bo tto m ha lf: c la im c o unts a sso c ia te d with A/ E s, simila rly de c o mpo se d b y unde rlying c a nc e r site s

SOA Pre dic tive Ana lytic s Se mina r 2018 Aug - Upskilling fo r Pre dic tive Ana lytic s

Pac ific L ife Re

Tableau: another example

23 SOA Pre dic tive Ana lytic s Se mina r 2018 Aug - Upskilling fo r Pre dic tive Ana lytic s

Pac ific L ife Re

Tableau example: cancer year over year correlation at Lag 1

• X-a xis de no te s the

ma g nitude a nd dire c tio n o f ye a r-

• n-ye a r c ha ng e s
• Y-a xis sho ws the

c o rre la tio n a t la g 1

• Mo re c o nc e ntra tio n
• n the b o tto m ha lf

E mpiric a l ye a r- on- ye a r c ha ng e s a re g e ne ra lly ne g a tive ly c orre la te d

Re duc tion F

• llowe d by

Re duc tion Inc r e ase F

• llowe d by

Inc r e ase Inc r e ase F

• llowe d by

Re duc tion Re duc tion F

• llowe d by

Inc r e ase

24 SOA Pre dic tive Ana lytic s Se mina r 2018 Aug - Upskilling fo r Pre dic tive Ana lytic s

Pac ific L ife Re

Gartner Magic Quadrant for Data Science and Machine Learning Platforms

25

Mo ving a lo t ye ar

• ve r

ye ar

K MI NE , Ra pidMine r, H20.a i, Alte ryx, SAS a ll c urre ntly ‘ visio na ry+le a de rs’ Da ta Ro b o t no t sho wn he re Ana c o nda – sho wn pre vio usly, po pula r distrib utio n o f Pytho n, R, Rstudio , Jupyte r…

SOA Pre dic tive Ana lytic s Se mina r 2018 Aug - Upskilling fo r Pre dic tive Ana lytic s

Pac ific L ife Re

DataRobot — Software helping to unite end to end process

26 SOA Pre dic tive Ana lytic s Se mina r 2018 Aug - Upskilling fo r Pre dic tive Ana lytic s

Pac ific L ife Re

rapidminer— Software helping to unite end to end process

27 SOA Pre dic tive Ana lytic s Se mina r 2018 Aug - Upskilling fo r Pre dic tive Ana lytic s

Pac ific L ife Re

KNIME — End to end for data analytics

28

Pre dic tive mo de lling , just o ne pie c e o f puzzle

SOA Pre dic tive Ana lytic s Se mina r 2018 Aug - Upskilling fo r Pre dic tive Ana lytic s

Pac ific L ife Re

Actuaries vs Data Scientists

Ric ha rd Wa rne r L V= on da ta sc ie ntists:

“We ’ ve a lwa ys ha d da ta sc ie ntists; the y’ ve b e e n c a lle d a c tua rie s a nd the y’ ve ha d trie d-a nd-te ste d mo de ls fo r ma ny ye a rs. But with the rise o f b ig da ta , a rtific ia l inte llig e nc e a nd ma c hine -le a rning , pe o ple a re unde rsta nding tha t we c a n c o mb ine multiple da ta so urc e s o f o ur o wn, a s we ll a s e xte rna l da ta , a nd b e tte r ta rg e t spe c ific c a se s o r nic he s, a nd g e t b e tte r insig ht, whe the r it is fo r pric ing o r c o unte ring fra ud,” sa ys Mr Wa rne r.

“Purple Squirre ls” a nd Unic orns

29

https:/ / www.ra c o nte ur.ne t/ risk-ma na g e me nt/ rise -da ta -sc ie ntist-insura nc e

SOA Pre dic tive Ana lytic s Se mina r 2018 Aug - Upskilling fo r Pre dic tive Ana lytic s

Pac ific L ife Re

Insurance innovation challenges —

Global insurance executives were asked about the challenges they face in ability to innovate

30

PWC 2017

Talent (87%)

Data storage, privacy , reg’s (63%)

Digital i.d. reg’s, new bus reg’s IT Security (53%)

https:/ / www.ra c o nte ur.ne t/ risk-ma na g e me nt/ rise -da ta -sc ie ntist-insura nc e

SOA Pre dic tive Ana lytic s Se mina r 2018 Aug - Upskilling fo r Pre dic tive Ana lytic s

Pac ific L ife Re

PLRe Structure

31

Divisio n Da ta Ana lytic s Busine ss Units, UWMe BU Ana lytic s Gro ups (R&D, Pric ing …) BU Ana lytic s Pro je c ts Divisio n R&D, Pric ing … Div Ana lytic s Pro je c ts

SOA Pre dic tive Ana lytic s Se mina r 2018 Aug - Upskilling fo r Pre dic tive Ana lytic s

Pac ific L ife Re

Structuring teams — Data engineers, could complement upskilled actuaries

32

Da ta sc ie ntists Ac tua rie s Da ta e ng ine e rs

SOA Pre dic tive Ana lytic s Se mina r 2018 Aug - Upskilling fo r Pre dic tive Ana lytic s

Pac ific L ife Re

Structuring for predictive analytics — Keeping nimble with cross team collaboration

33

Legal & Risk Underwriting & Claims Marketing and Distribution Actuarial

• Ag re e me nts
• Re g ula tio ns
• GDPR, priva c y la ws
• Co nne c ting to ma rke ting a nd

distrib utio n

• Priva c y c ha lle ng e s
• Ag e nc y g ro ups, b ro ke rs, third

pa rtie s

• Ma rke ting sc ie nc e
• Whe n pre dic tive mo de ls b e c o me

a ssumptio ns

• Pro fe ssio na lism

SOA Pre dic tive Ana lytic s Se mina r 2018 Aug - Upskilling fo r Pre dic tive Ana lytic s

Pac ific L ife Re

Upskilling for actuaries: plunge into it

34 SOA Pre dic tive Ana lytic s Se mina r 2018 Aug - Upskilling fo r Pre dic tive Ana lytic s

"I think, ultima te ly, le a rning ho w to do da ta sc ie nc e is like le a rning to ski. Yo u ha ve to do it. Yo u c a n o nly liste n to so ma ny vide o s a nd wa tc h it ha ppe n. At the e nd o f the da y, yo u ha ve to g e t

• n yo ur da mn skis a nd g o do wn tha t hill."

Cla udia Pe rlic h, Chie f Sc ie ntist a t Dstille ry

https:/ / www.kdnug g e ts.c o m/ 2018/ 05/ 8-use ful-a dvic e s-a spiring -da ta -sc ie ntists.html

Pac ific L ife Re

Upskilling for predictive analytics — Many avenues…

35

SOA Offerings Online Learn from others University … examples

• E

xa m PA

• PA Ce rtific a te Pro g ra m
• Co urse ra (Andre w Ng ML

), Uda c ity, K ha n Ac a de my…

• K

a g g le , disc

• Jupyte r no te b o o ks o nline
• K

dnug g e ts, r-b lo g g e rs, …

• Ma ste rs in Da ta Sc ie nc e (Be rke le y)
• Online
• Oppo rtunity c o st?

SOA Pre dic tive Ana lytic s Se mina r 2018 Aug - Upskilling fo r Pre dic tive Ana lytic s

Reboot.

Do n’ t wo rry a b o ut the Ac tua ry. I ’ m sure he ’ ll b e a ll rig ht. He ’ s q uite c le ve r, yo u kno w… fo r a huma n b e ing .