Change-point detection in insurance portfolios Laub Work in - - PowerPoint PPT Presentation

▶
change point detection in insurance portfolios
SMART_READER_LITE
LIVE PREVIEW

Change-point detection in insurance portfolios Laub Work in - - PowerPoint PPT Presentation

Aug 22, 2023 228 likes •607 views

Change-point detection in insurance portfolios Dr Patrick J. Change-point detection in insurance portfolios Laub Work in progress Introduction Problem Data Dr Patrick J. Laub ML Options Post-doctoral researcher with Chaire DAMI BNP

slide-1
SLIDE 1

Change-point detection in insurance portfolios Dr Patrick J. Laub Introduction Problem Data ML Options

Change-point detection in insurance portfolios

Work in progress Dr Patrick J. Laub

Post-doctoral researcher with Chaire DAMI BNP Paribas Cardif Technical Seminar

March 27, 2019

1 / 17

slide-2
SLIDE 2

Change-point detection in insurance portfolios Dr Patrick J. Laub Introduction Problem Data ML Options

Who am I

Mathematician + Software Engineer ≈ Data Scientist PhD in Applied Probability between University of Queensland (Brisbane) and Aarhus University Arrived to Lyon in September

PhD Thesis

2 / 17

slide-3
SLIDE 3

Change-point detection in insurance portfolios Dr Patrick J. Laub Introduction Problem Data ML Options

Since arriving

Hosting Visiting Professors Séminaire Lyon-Lausanne Server/GPU access Apartment hunting D e p a r t m e n t S é m i n a i r e Conference submissions Bank account French classes Reporting Re/submitted 3 papers Préfecture Teaching R e s i d e n c e P e r m i t I n t e r v i e w s Database access Surgery CPAM

3 / 17

slide-4
SLIDE 4

Change-point detection in insurance portfolios Dr Patrick J. Laub Introduction Problem Data ML Options

Since arriving: Paper accepted

Shortened recording of a seminar on the topic: https://youtu.be/8Ih2NxrLrmg

3 / 17

slide-5
SLIDE 5

Change-point detection in insurance portfolios Dr Patrick J. Laub Introduction Problem Data ML Options

Since arriving: Created and delivered a course on rare-event estimation/simulation

Lecture notes & recordings at: https://pat-laub.github.io/rare-events/

3 / 17

slide-6
SLIDE 6

Change-point detection in insurance portfolios Dr Patrick J. Laub Introduction Problem Data ML Options

The BNP Netherlands contracts tables

C2014 = C2015 =

4 / 17

slide-7
SLIDE 7

Change-point detection in insurance portfolios Dr Patrick J. Laub Introduction Problem Data ML Options

The problem

Did our risk exposure change in some significant way? ⇒ We can update premiums/reserves. Look at Ct and Ct+1 given C0, C1, . . . , Ct+1 and throw up an alarm if distance(Ct, Ct+1) > threshold. Restriction: Don’t assume some parametric form of Ct.

5 / 17

slide-8
SLIDE 8

Change-point detection in insurance portfolios Dr Patrick J. Laub Introduction Problem Data ML Options

Prior work

Nothing has been done on exactly this topic... Similar to “fraud detection” in the sense that the events of interest are rare... but otherwise no similarity. Standard change-point detection is a huge field, mostly parametric (e.g. hidden Markov models, Markov regime switching models). Mostly uni- or low-dimensional. Require samples to be piecewise stationary/i.i.d. and to keep a fixed dimension. Similar to the discriminator in a GAN; that tries to learn a distribution and discern perturbations from the learned distribution.

6 / 17

slide-9
SLIDE 9

Change-point detection in insurance portfolios Dr Patrick J. Laub Introduction Problem Data ML Options

Data: Big Picture

Number of payouts made across the different coverage types

7 / 17

slide-10
SLIDE 10

Change-point detection in insurance portfolios Dr Patrick J. Laub Introduction Problem Data ML Options

Data: Big Picture

Total payouts made across the different coverage types

7 / 17

slide-11
SLIDE 11

Change-point detection in insurance portfolios Dr Patrick J. Laub Introduction Problem Data ML Options

Data: Big Picture

Claim outcomes

7 / 17

slide-12
SLIDE 12

Change-point detection in insurance portfolios Dr Patrick J. Laub Introduction Problem Data ML Options

Data: Big Picture

Claim number distribution

7 / 17

slide-13
SLIDE 13

Change-point detection in insurance portfolios Dr Patrick J. Laub Introduction Problem Data ML Options

Data: Big Picture

Claim number distribution (zoomed)

7 / 17

slide-14
SLIDE 14

Change-point detection in insurance portfolios Dr Patrick J. Laub Introduction Problem Data ML Options

All Columns

Contracts table Claims table

policy-nr, coverage-type, insured-id, product-code, broker-nr, agency-nr, network-nr, certificate-nr, professional-status, contract-status, coverage-status, last-coverage-status-change, premium-type, premium-schema, commission-schema, commission-type, birth-date, gender, effect-date, dm, currency, percentage-premium-increase, claim-waiting-period, type-of-loan, percentage-annuity-decreasing, insured-amount–capital, insured-amount-type, fiscality, premium-increase, incasso-machtiging, p, pm, percentage-chosen-commission, percentage-chosen-commission-y1, percentage-chosen-commission-y2, percentage-chosen-commission-y3, smoking, death-coverage-type, ao-coverage-type, ao-vaststelling, claim-payment-duration, child-addition, repatriation, dm1, premium-dm1, premium-dm2, payment-received, last-paid-billing-date, single-com, rev-com, surrenderdate, invoerdatum, discount-amount, discount-dm, discount-type, coverageendreason, tariff, pawntaker, occupation, occupation-code,

  • ccupation-class

loss-number, application-code, gender, birth-date, status-eng, policy-number, prod-code, product, insured-risk, insured-amount, start-date-contract, policy-duration, end-date-contract, end-date-benefit, date-of-loss-occurence, date-of-loss-declaration, date-loss-was-closed, type-of-loss, franchise-abs, decision, toekenningsreden, reason-commercial-coulance, coulance-monitoring, beroep, total-amount-paid-euro, first-payment, benefit-period-of-first-payment, latest-payment, benefit-period-of-last-payment, number-of-paid-periods, relatienummer, verz-kapitaal, invoerdat, oorzaak-type, reden-van-afwijzing, type-lening, ao-percentage, ao-vaststelling-, juridisch-proces-date, reopened-date, uitkeringspercentage, pemsa-insuredid, pemsa-coverage-type, coverage-duration, pemsa-product-code 8 / 17

slide-15
SLIDE 15

Change-point detection in insurance portfolios Dr Patrick J. Laub Introduction Problem Data ML Options

All Columns − Unused

Contracts table Claims table

policy-nr, coverage-type, insured-id, product-code, broker-nr, agency-nr, network-nr, / / / / / / / / / / / / / / / certificate-nr, professional-status, / / / / / / / / / / / / / / / / / contract-status, coverage-status, last-coverage-status-change, premium-type, / / / / / / / / / / / / / / / / / / / / premium-schema, commission-schema, / / / / / / / / / / / / / / / / / / / / commission-type, birth-date, gender, effect-date, dm, currency, percentage-premium-increase, claim-waiting-period, type-of-loan, / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / percentage-annuity-decreasing, insured-amount–capital, / / / / / / / / / / / / / / / / / / / / / / / / insured-amount-type, fiscality, premium-increase, incasso-machtiging, p, pm,/ / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / percentage-chosen-commission, / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / percentage-chosen-commission-y1, / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / percentage-chosen-commission-y2, / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / percentage-chosen-commission-y3, smoking, death-coverage-type, ao-coverage-type, ao-vaststelling, / / / / / / / / / / / / / / / / / / / / / / / / / / / / claim-payment-duration, child-addition, repatriation, dm1, premium-dm1, premium-dm2, payment-received, last-paid-billing-date, single-com, rev-com, surrenderdate, invoerdatum,/ / / / / / / / / / / / / / / / / / / / discount-amount, / / / / / / / / / / / / / / / discount-dm,/ / / / / / / / / / / / / / / / discount-type, / / / / / / / / / / / / / / / / / / / / / / coverageendreason, tariff,/ / / / / / / / / / / / / pawntaker, / / / / / / / / / / / / /

  • ccupation, /

/ / / / / / / / / / / / / / / / / /

  • ccupation-code,

/ / / / / / / / / / / / / / / / / / /

  • ccupation-class

loss-number, application-code, gender, birth-date, status-eng, policy-number, prod-code, product, insured-risk, insured-amount, start-date-contract, policy-duration, end-date-contract, end-date-benefit, date-of-loss-occurence, date-of-loss-declaration, date-loss-was-closed, type-of-loss, franchise-abs, / / / / / / / / / decision, / / / / / / / / / / / / / / / / / / / / / toekenningsreden, / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / reason-commercial-coulance, coulance-monitoring, beroep, total-amount-paid-euro, first-payment, benefit-period-of-first-payment, latest-payment, benefit-period-of-last-payment, number-of-paid-periods,/ / / / / / / / / / / / / / / / / / relatienummer, verz-kapitaal, invoerdat, oorzaak-type, / / / / / / / / / / / / / / / / / / / / / / / reden-van-afwijzing, type-lening, ao-percentage, ao-vaststelling-, / / / / / / / / / / / / / / / / / / / / / / / / juridisch-proces-date, reopened-date, uitkeringspercentage, pemsa-insuredid, pemsa-coverage-type, coverage-duration, pemsa-product-code 8 / 17

slide-16
SLIDE 16

Change-point detection in insurance portfolios Dr Patrick J. Laub Introduction Problem Data ML Options

All Columns − Unused/Unknown

Contracts table Claims table

policy-nr, coverage-type, insured-id, product-code, broker-nr, agency-nr, network-nr, / / / / / / / / / / / / / / / certificate-nr, professional-status, / / / / / / / / / / / / / / / / / contract-status, coverage-status, last-coverage-status-change, premium-type, / / / / / / / / / / / / / / / / / / / / premium-schema, / / / / / / / / / / / / / / / / / / / / / / / commission-schema, / / / / / / / / / / / / / / / / / / / / commission-type, birth-date, gender, effect-date, dm, currency, percentage-premium-increase, claim-waiting-period, type-of-loan, / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / percentage-annuity-decreasing, insured-amount–capital, / / / / / / / / / / / / / / / / / / / / / / / / insured-amount-type, fiscality, premium-increase, / / / / / / / / / / / / / / / / / / / / / / / incasso-machtiging, p, pm,/ / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / percentage-chosen-commission, / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / percentage-chosen-commission-y1, / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / percentage-chosen-commission-y2, / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / percentage-chosen-commission-y3, smoking, death-coverage-type, ao-coverage-type, ao-vaststelling, / / / / / / / / / / / / / / / / / / / / / / / / / / / / claim-payment-duration, / / / / / / / / / / / / / / / / child-addition, / / / / / / / / / / / / / / repatriation, dm1, premium-dm1, premium-dm2, payment-received, last-paid-billing-date, single-com, rev-com, surrenderdate, invoerdatum,/ / / / / / / / / / / / / / / / / / / / discount-amount, / / / / / / / / / / / / / / / discount-dm,/ / / / / / / / / / / / / / / / discount-type, / / / / / / / / / / / / / / / / / / / / / / coverageendreason, tariff,/ / / / / / / / / / / / / pawntaker, / / / / / / / / / / / / /

  • ccupation, /

/ / / / / / / / / / / / / / / / / /

  • ccupation-code,

/ / / / / / / / / / / / / / / / / / /

  • ccupation-class

loss-number, application-code, gender, birth-date, status-eng, policy-number, prod-code, product, insured-risk, insured-amount, start-date-contract, policy-duration, end-date-contract, end-date-benefit, date-of-loss-occurence, date-of-loss-declaration, date-loss-was-closed, type-of-loss, franchise-abs, / / / / / / / / / decision, / / / / / / / / / / / / / / / / / / / / / toekenningsreden, / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / reason-commercial-coulance, coulance-monitoring, beroep, total-amount-paid-euro, first-payment, benefit-period-of-first-payment, latest-payment, benefit-period-of-last-payment, number-of-paid-periods,/ / / / / / / / / / / / / / / / / / relatienummer, verz-kapitaal, invoerdat, oorzaak-type, / / / / / / / / / / / / / / / / / / / / / / / reden-van-afwijzing, type-lening, ao-percentage, ao-vaststelling-, / / / / / / / / / / / / / / / / / / / / / / / / juridisch-proces-date, reopened-date, / / / / / / / / / / / / / / / / / / / / / / / / uitkeringspercentage, pemsa-insuredid, pemsa-coverage-type, coverage-duration, pemsa-product-code 8 / 17

slide-17
SLIDE 17

Change-point detection in insurance portfolios Dr Patrick J. Laub Introduction Problem Data ML Options

All Columns − Unused/Unknown/NaN

Contracts table Claims table

policy-nr, coverage-type, insured-id, product-code, broker-nr, agency-nr, network-nr, / / / / / / / / / / / / / / / certificate-nr, professional-status, / / / / / / / / / / / / / / / / / contract-status, coverage-status, last-coverage-status-change, premium-type, / / / / / / / / / / / / / / / / / / / / premium-schema, / / / / / / / / / / / / / / / / / / / / / / / commission-schema, / / / / / / / / / / / / / / / / / / / / commission-type, birth-date, gender, effect-date, dm, currency, / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / percentage-premium-increase, claim-waiting-period, / / / / / / / / / / / / / / type-of-loan, / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / percentage-annuity-decreasing, insured-amount–capital, / / / / / / / / / / / / / / / / / / / / / / / / insured-amount-type, fiscality, / / / / / / / / / / / / / / / / / / / / premium-increase, / / / / / / / / / / / / / / / / / / / / / / / incasso-machtiging, / / p, / / / / pm,/ / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / percentage-chosen-commission, / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / percentage-chosen-commission-y1, / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / percentage-chosen-commission-y2, / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / percentage-chosen-commission-y3, smoking, / / / / / / / / / / / / / / / / / / / / / / / / death-coverage-type, ao-coverage-type, ao-vaststelling, / / / / / / / / / / / / / / / / / / / / / / / / / / / / claim-payment-duration, / / / / / / / / / / / / / / / / child-addition, / / / / / / / / / / / / / / repatriation, dm1, premium-dm1, / / / / / / / / / / / / / / / / premium-dm2, payment-received, last-paid-billing-date, / / / / / / / / / / / / / single-com,/ / / / / / / / / / rev-com, / / / / / / / / / / / / / / / / surrenderdate, invoerdatum,/ / / / / / / / / / / / / / / / / / / / discount-amount, / / / / / / / / / / / / / / / discount-dm,/ / / / / / / / / / / / / / / / discount-type, / / / / / / / / / / / / / / / / / / / / / / coverageendreason, / / / / / / tariff,/ / / / / / / / / / / / / pawntaker, / / / / / / / / / / / / /

  • ccupation, /

/ / / / / / / / / / / / / / / / / /

  • ccupation-code,

/ / / / / / / / / / / / / / / / / / /

  • ccupation-class

loss-number, application-code, gender, birth-date, status-eng, policy-number, prod-code, product, insured-risk, insured-amount, start-date-contract, policy-duration, end-date-contract,/ / / / / / / / / / / / / / / / / / / / end-date-benefit, date-of-loss-occurence, date-of-loss-declaration, date-loss-was-closed, type-of-loss, franchise-abs, / / / / / / / / / decision, / / / / / / / / / / / / / / / / / / / / / toekenningsreden, / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / reason-commercial-coulance, / / / / / / / / / / / / / / / / / / / / / / / / coulance-monitoring, / / / / / / / / beroep, total-amount-paid-euro, first-payment, benefit-period-of-first-payment, latest-payment, benefit-period-of-last-payment, number-of-paid-periods,/ / / / / / / / / / / / / / / / / / relatienummer, verz-kapitaal, invoerdat, oorzaak-type, / / / / / / / / / / / / / / / / / / / / / / / reden-van-afwijzing, type-lening, ao-percentage, ao-vaststelling-, / / / / / / / / / / / / / / / / / / / / / / / / juridisch-proces-date, reopened-date, / / / / / / / / / / / / / / / / / / / / / / / / uitkeringspercentage, pemsa-insuredid, pemsa-coverage-type, coverage-duration, pemsa-product-code 8 / 17

slide-18
SLIDE 18

Change-point detection in insurance portfolios Dr Patrick J. Laub Introduction Problem Data ML Options

All Columns − Unused/Unknown/NaN/Dates/$’s

Contracts table Claims table

policy-nr, coverage-type, insured-id, product-code, broker-nr, agency-nr, network-nr, / / / / / / / / / / / / / / / certificate-nr, professional-status, / / / / / / / / / / / / / / / / / contract-status, coverage-status, last-coverage-status-change, / / / / / / / / / / / / / / / / premium-type, / / / / / / / / / / / / / / / / / / / / premium-schema, / / / / / / / / / / / / / / / / / / / / / / / commission-schema, / / / / / / / / / / / / / / / / / / / / commission-type, birth-date, gender, / / / / / / / / / / / / effect-date, dm, currency, / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / percentage-premium-increase, claim-waiting-period, / / / / / / / / / / / / / / type-of-loan, / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / percentage-annuity-decreasing, insured-amount–capital, / / / / / / / / / / / / / / / / / / / / / / / / insured-amount-type, / / / / / / / / / fiscality, / / / / / / / / / / / / / / / / / / / / premium-increase, / / / / / / / / / / / / / / / / / / / / / / / incasso-machtiging, / / p, / / / / pm,/ / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / percentage-chosen-commission, / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / percentage-chosen-commission-y1, / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / percentage-chosen-commission-y2, / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / percentage-chosen-commission-y3, smoking, / / / / / / / / / / / / / / / / / / / / / / / / death-coverage-type, ao-coverage-type, ao-vaststelling, / / / / / / / / / / / / / / / / / / / / / / / / / / / / claim-payment-duration, / / / / / / / / / / / / / / / / child-addition, / / / / / / / / / / / / / / repatriation, dm1,/ / / / / / / / / / / / / / / / / premium-dm1, / / / / / / / / / / / / / / / / premium-dm2, / / / / / / / / / / / / / / / / / / / / / payment-received, / / / / / / / / / / / / / / / / / / / / / / / / last-paid-billing-date, / / / / / / / / / / / / / single-com,/ / / / / / / / / / rev-com, / / / / / / / / / / / / / / / / surrenderdate, / / / / / / / / / / / / / / / invoerdatum,/ / / / / / / / / / / / / / / / / / / / discount-amount, / / / / / / / / / / / / / / / discount-dm,/ / / / / / / / / / / / / / / / discount-type, / / / / / / / / / / / / / / / / / / / / / / coverageendreason, / / / / / / tariff,/ / / / / / / / / / / / / pawntaker, / / / / / / / / / / / / /

  • ccupation, /

/ / / / / / / / / / / / / / / / / /

  • ccupation-code,

/ / / / / / / / / / / / / / / / / / /

  • ccupation-class

loss-number, application-code, gender, birth-date, status-eng, policy-number, prod-code, product, insured-risk, insured-amount,/ / / / / / / / / / / / / / / / / / / / / / / start-date-contract, policy-duration, / / / / / / / / / / / / / / / / / / / / / end-date-contract,/ / / / / / / / / / / / / / / / / / / / end-date-benefit, / / / / / / / / / / / / / / / / / / / / / / / / / / date-of-loss-occurence, / / / / / / / / / / / / / / / / / / / / / / / / / / / / date-of-loss-declaration, / / / / / / / / / / / / / / / / / / / / / / / / date-loss-was-closed, type-of-loss, franchise-abs, / / / / / / / / / decision, / / / / / / / / / / / / / / / / / / / / / toekenningsreden, / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / reason-commercial-coulance, / / / / / / / / / / / / / / / / / / / / / / / / coulance-monitoring, / / / / / / / / beroep, / / / / / / / / / / / / / / / / / / / / / / / / / / / total-amount-paid-euro, / / / / / / / / / / / / / / / / first-payment, / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / benefit-period-of-first-payment, / / / / / / / / / / / / / / / / / / latest-payment, / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / benefit-period-of-last-payment, / / / / / / / / / / / / / / / / / / / / / / / / / / / / number-of-paid-periods,/ / / / / / / / / / / / / / / / / / relatienummer, verz-kapitaal, / / / / / / / / / / / invoerdat, oorzaak-type, / / / / / / / / / / / / / / / / / / / / / / / reden-van-afwijzing, type-lening, ao-percentage, ao-vaststelling-, / / / / / / / / / / / / / / / / / / / / / / / / juridisch-proces-date, / / / / / / / / / / / / / / / / / reopened-date, / / / / / / / / / / / / / / / / / / / / / / / / uitkeringspercentage, pemsa-insuredid, pemsa-coverage-type, coverage-duration, pemsa-product-code 8 / 17

slide-19
SLIDE 19

Change-point detection in insurance portfolios Dr Patrick J. Laub Introduction Problem Data ML Options

All Columns − Unused/Unknown/NaN/Dates/$’s

Contracts table Claims table

policy-nr, coverage-type, insured-id, product-code, broker-nr, agency-nr, network-nr, / / / / / / / / / / / / / / / certificate-nr, professional-status, / / / / / / / / / / / / / / / / / contract-status, coverage-status, last-coverage-status-change, / / / / / / / / / / / / / / / / premium-type, / / / / / / / / / / / / / / / / / / / / premium-schema, / / / / / / / / / / / / / / / / / / / / / / / commission-schema, / / / / / / / / / / / / / / / / / / / / commission-type, birth-date, gender, / / / / / / / / / / / / effect-date, dm, currency, / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / percentage-premium-increase, claim-waiting-period, / / / / / / / / / / / / / / type-of-loan, / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / percentage-annuity-decreasing, insured-amount–capital, / / / / / / / / / / / / / / / / / / / / / / / / insured-amount-type, / / / / / / / / / fiscality, / / / / / / / / / / / / / / / / / / / / premium-increase, / / / / / / / / / / / / / / / / / / / / / / / incasso-machtiging, / / p, / / / / pm,/ / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / percentage-chosen-commission, / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / percentage-chosen-commission-y1, / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / percentage-chosen-commission-y2, / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / percentage-chosen-commission-y3, smoking, / / / / / / / / / / / / / / / / / / / / / / / / death-coverage-type, ao-coverage-type, ao-vaststelling, / / / / / / / / / / / / / / / / / / / / / / / / / / / / claim-payment-duration, / / / / / / / / / / / / / / / / child-addition, / / / / / / / / / / / / / / repatriation, dm1,/ / / / / / / / / / / / / / / / / premium-dm1, / / / / / / / / / / / / / / / / premium-dm2, / / / / / / / / / / / / / / / / / / / / / payment-received, / / / / / / / / / / / / / / / / / / / / / / / / last-paid-billing-date, / / / / / / / / / / / / / single-com,/ / / / / / / / / / rev-com, / / / / / / / / / / / / / / / / surrenderdate, / / / / / / / / / / / / / / / invoerdatum,/ / / / / / / / / / / / / / / / / / / / discount-amount, / / / / / / / / / / / / / / / discount-dm,/ / / / / / / / / / / / / / / / discount-type, / / / / / / / / / / / / / / / / / / / / / / coverageendreason, / / / / / / tariff,/ / / / / / / / / / / / / pawntaker, / / / / / / / / / / / / /

  • ccupation, /

/ / / / / / / / / / / / / / / / / /

  • ccupation-code,

/ / / / / / / / / / / / / / / / / / /

  • ccupation-class

loss-number, application-code, gender, birth-date, status-eng, policy-number, prod-code, product, insured-risk, insured-amount,/ / / / / / / / / / / / / / / / / / / / / / / start-date-contract, policy-duration, / / / / / / / / / / / / / / / / / / / / / end-date-contract,/ / / / / / / / / / / / / / / / / / / / end-date-benefit, / / / / / / / / / / / / / / / / / / / / / / / / / / date-of-loss-occurence, / / / / / / / / / / / / / / / / / / / / / / / / / / / / date-of-loss-declaration, / / / / / / / / / / / / / / / / / / / / / / / / date-loss-was-closed, type-of-loss, franchise-abs, / / / / / / / / / decision, / / / / / / / / / / / / / / / / / / / / / toekenningsreden, / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / reason-commercial-coulance, / / / / / / / / / / / / / / / / / / / / / / / / coulance-monitoring, / / / / / / / / beroep, / / / / / / / / / / / / / / / / / / / / / / / / / / / total-amount-paid-euro, / / / / / / / / / / / / / / / / first-payment, / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / benefit-period-of-first-payment, / / / / / / / / / / / / / / / / / / latest-payment, / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / benefit-period-of-last-payment, / / / / / / / / / / / / / / / / / / / / / / / / / / / / number-of-paid-periods,/ / / / / / / / / / / / / / / / / / relatienummer, verz-kapitaal, / / / / / / / / / / / invoerdat, oorzaak-type, / / / / / / / / / / / / / / / / / / / / / / / reden-van-afwijzing, type-lening, ao-percentage, ao-vaststelling-, / / / / / / / / / / / / / / / / / / / / / / / / juridisch-proces-date, / / / / / / / / / / / / / / / / / reopened-date, / / / / / / / / / / / / / / / / / / / / / / / / uitkeringspercentage, pemsa-insuredid, pemsa-coverage-type, coverage-duration, pemsa-product-code

Demographic data: birth date, gender, smoking, employed/self-employed/unemployed

8 / 17

slide-20
SLIDE 20

Change-point detection in insurance portfolios Dr Patrick J. Laub Introduction Problem Data ML Options

Extract time series from claims table

9 / 17

slide-21
SLIDE 21

Change-point detection in insurance portfolios Dr Patrick J. Laub Introduction Problem Data ML Options

Extract time series from claims table

9 / 17

slide-22
SLIDE 22

Change-point detection in insurance portfolios Dr Patrick J. Laub Introduction Problem Data ML Options

Extract time series from claims table

9 / 17

slide-23
SLIDE 23

Change-point detection in insurance portfolios Dr Patrick J. Laub Introduction Problem Data ML Options

Extract time series from claims table

9 / 17

slide-24
SLIDE 24

Change-point detection in insurance portfolios Dr Patrick J. Laub Introduction Problem Data ML Options

Extract time series from claims table

9 / 17

slide-25
SLIDE 25

Change-point detection in insurance portfolios Dr Patrick J. Laub Introduction Problem Data ML Options

Extract time series from claims table

9 / 17

slide-26
SLIDE 26

Change-point detection in insurance portfolios Dr Patrick J. Laub Introduction Problem Data ML Options

Extract time series from claims table

Number of active contracts in the portfolio

Need to somewhat guess the start/end dates for when the contract is in force 9 / 17

slide-27
SLIDE 27

Change-point detection in insurance portfolios Dr Patrick J. Laub Introduction Problem Data ML Options

Extract time series from claims table

Average age of contract-holders in the portfolio

9 / 17

slide-28
SLIDE 28

Change-point detection in insurance portfolios Dr Patrick J. Laub Introduction Problem Data ML Options

Extract time series from claims table

Number of past claims made / number of contracts in the portfolio

9 / 17

slide-29
SLIDE 29

Change-point detection in insurance portfolios Dr Patrick J. Laub Introduction Problem Data ML Options

Extract time series from claims table

Average claim sizes scaled by number of contracts with claims & number of years contract has been in force

9 / 17

slide-30
SLIDE 30

Change-point detection in insurance portfolios Dr Patrick J. Laub Introduction Problem Data ML Options

Can throw these traditional time-series into change-point algorithms

Truonga et al. (2019), Selective review of offline change point detection methods, arXiv:1801.00718v2

10 / 17

slide-31
SLIDE 31

Change-point detection in insurance portfolios Dr Patrick J. Laub Introduction Problem Data ML Options

Supervised learning

If we had for each (Ct, Ct+1) pair an expert-generated label yt =

  • 1

if there was a significant change at time t

  • therwise

then it would be possible to directly apply neural network classifier to the task.

11 / 17

slide-32
SLIDE 32

Change-point detection in insurance portfolios Dr Patrick J. Laub Introduction Problem Data ML Options

Using simulated data

Make some simulated data:

1 Generate ‘standard’ instance {C0, C1 . . . , CT}. 2 Choose a random time for a change-point, τ 3 Choose some scenario, e.g. the mean of a column gets

bigger

4 Modify Cτ, Cτ+1, . . . , CT according to the selected

scenario

5 Set yτ = 1 and yt = 0 for the rest.

Train a neural network classifier on this...

12 / 17

slide-33
SLIDE 33

Change-point detection in insurance portfolios Dr Patrick J. Laub Introduction Problem Data ML Options

Temporal clustering

Anderlucci et al. (2017), The Importance of Being Clustered — Uncluttering the Trends of Statistics from 1970 to 2015, ArXiv Preprint 1709.03563v1.

13 / 17

slide-34
SLIDE 34

Change-point detection in insurance portfolios Dr Patrick J. Laub Introduction Problem Data ML Options

Using autoencoders

Lee et al. (2018), Time series segmentation through automatic feature learning.

14 / 17

slide-35
SLIDE 35

Change-point detection in insurance portfolios Dr Patrick J. Laub Introduction Problem Data ML Options

Data science process

Wikipedia, Cross-industry standard process for data mining.

15 / 17

slide-36
SLIDE 36

Change-point detection in insurance portfolios Dr Patrick J. Laub Introduction Problem Data ML Options

The problem with skipping to the fun parts...

“A frequent mistake that many inexperienced data scientists make is to focus their efforts on the modeling stage of the CRISP-DM and to rush through the other stages... However, data science veterans will spend more time on ensuring that the project has a clearly defined focus and that it has the right data. For a data science project to succeed, a data scientist needs to have a clear understanding of the business need that the project is trying to solve. So the business understanding stage of the process is really important.

Kelleher and Tierney (2018), Data Science, MIT Press, Cambridge MA, p. 65.

16 / 17

slide-37
SLIDE 37

Change-point detection in insurance portfolios Dr Patrick J. Laub Introduction Problem Data ML Options

Business & data understanding

What are some example scenarios where this kind of algorithm would have been useful? What is “overall risk exposure” from before? A precise definition is required. What is the current approach to solve this problem?

17 / 17