A

nudge is any aspect of the choice architecture

that alters peoples behaviour in a predictable way
without forbidding any op9ons.

-Richard Thaler

Professor of Behavioural Economics

University of Chicago

Big picture
Be=air is an online marketplace
Match odds, bet sizes are unknown a priori
Nudging user behaviour can be useful

Big picture
Be=air is an online marketplace
Match odds, bet sizes are unknown a priori
Nudging user behaviour can be useful

Examples of nudges

Informing (passive) users that their bets are likely to lapse

Giving opLons to (acLve) users who want to cancel their bets
Using a user skill database to incenLvize selected (skilled) users
All of the above need predicLve capabiliLes

Two key quesLons:

What is the probability of any given bet being cancelled or lapsed?

What is the probability of an individual user winning a transacLon?


The goal of our analysis is to develop models to answer these quesLons.

Problem: Way too much data!

Which bets are likely to be cancelled?

S
.19 .01 .80 .00
100%
yes

BET_SIZE >= 1961

no

S
.43 .01 .56 .00
16%
BET_SIZE >= 19e+3

C = Cancelled

L = Lapsed

C
.69 .01 .30 .00
2%

S
.39 .01 .60 .00
14%

S
.15 .01 .84 .00
84%

S = SeWled

V = Void

Rattle 2015Oct17 14:40:27 rohit

Which bets are likely to lapse?

150

Mean bet price/match
Summed over all matches

Sum of mean Bet Price across all matches

AggregaLon Approach

100

STATUS_ID
C
L
S

50

Thickness indicates frequency

0

Status of Bet

Cancelled bets correlate with bet size and whether they are made in-play.

Lapsed bets correlate with mean bet price.

Possible confounders?

Cancelled bets correlate with bet size and whether they are made in-play.

Lapsed bets correlate with mean bet price.

Possible confounders?

Second approach look at users
Separated users into Regular (Total bets < 1 million AUD), Power (Total bets < 10
million AUD) and Super (Total bets > 10 million AUD) categories.

~18000 Regulars (R), ~1650 Power (P) and ~350 Super (S) users idenLed.

Cancelled bets correlate with bet size and whether they are made in-play.

Lapsed bets correlate with mean bet price.

Possible confounders?

Second approach look at users
Separated users into Regular (Total bets < 1 million AUD), Power (Total bets < 10
million AUD) and Super (Total bets > 10 million AUD) categories.

~18000 Regulars (R), ~1650 Power (P) and ~350 Super (S) users idenLed.

Key ndings - Mean Bet sizes per user ~370 (R), ~1250 (P) and ~2700 (S) (all AUD)

Mean transacLons per user ~50 (R), ~500 (P) and ~4450 (S) !

Big players bet far more regularly but are they beWer?

Regular

Total Prots/losses divided by total bets

MulLple modes = Lost everything

0
5

0
Power

25

20

Density

15

USER_TYPE
Regular
Power

10

Super

Power users do beWer than

regular ones

0
0.3

0.0

0.3
Super

50

0.6

0.9

40

Super users do the best

- Right-shijed curve

Trickle up eect?

30

20

10

0
0.2

0.1

0.0

Ratio of profits/losses to Total Bets

0.1

 Model 1 Lapsed/cancelled bet classier

Methodology:
Separated rst 40 match (train) and QF data (test)
Used Random Forest algorithm along with feature selecLon

 Model 1 Lapsed/cancelled bet classier

Methodology:
Separated rst 40 match (train) and QF data (test)
Used Random Forest algorithm along with feature selecLon
Results:
92.47 % accuracy (CI: 91.94 92.98)
High sensiLvity and specicity for cancelled bets.

LimitaLons
Poor performance with predicLng lapsed bets
Feature selecLon can be improved

fit

INPLAY_BET

INPLAY_BET

BET_PRICE

BET_SIZE

BET_SIZE

BET_PRICE

BID_TYP

BID_TYP

50 150 250 350

MeanDecreaseAccuracy

500
1500
MeanDecreaseGini

 Model 2 Individual user win/lose classier

Methodology:
Separated data into blocks, employed Lme series approach to each user
Comprehensive feature engineering introduced including cumulaLve wins at
sample, prior average bet size, prior prot/loss, number of prior bets etc
Used a logisLc regression classier to predict win/loss

 Model 2 Individual user win/lose classier

Methodology:
Separated data into blocks, employed Lme series approach to each user
Comprehensive feature engineering introduced including cumulaLve wins at
sample, prior average bet size, prior prot/loss, number of prior bets etc
Used a logisLc regression classier to predict win/loss

Results:
It performs reasonably well.
LimitaLons:
Future plans to try alternaLve approaches including random forest and neural
networks to improve accuracy

Conclusions

Suggested mechanism - nudging customers

Requires predicLve analyLcs

Feature design for classicaLon was informed by extensive data analysis

Two classier models were developed for bet status and individual user win/loss

The approaches show promise