How Robo Advice changes individual investor behavior Andreas Hackethal (Goethe University) February 16, 2018 OEE, Paris Financial support by OEE of presented research studies is gratefully acknowledged
General definition of automation in financial advice ( robo advice ) (Joint Committee of the European Supervisory Authorities 2015) Input: individual information Production: tool and algorithm Output: recommendation and reporting Objective data, e.g., age, job, monthly income, number of children Subjective data, e.g., investment goal, financial condition, risk tolerance, knowledge and experience Automated tool operated by the consumer Portfolio management ranges from pure buyand-hold to risk mitigation and liabilitybased strategies Without (or with very limited) human intervention Guidance 1 on asset allocation and product selection Transactions on the push of one button Reporting on portfolio perspective and full cost transparency 1 The provision of information, generic advice and/or a general recommendation supporting customers in making their own investment decisions which does not (in and of itself) involve a personal recommendation 2
Robo Advisors are mushrooming not only in Germany Some 30+ FinTechs Most online brokerages Most large incumbents such as Deutsche Bank, Savings Bank Group, Co-operative Banks Insurance Companies are testing waters Surpassing 1bn AUM 3
OEE support sparked off three papers on Robo Advice 1) Portfolio Efficiency 2) Savings Plans 3) Participation Loos, Previtero, Scheurle, Hackethal (2018): Robo Advice and Investor Behavior Main findings: Introduction of Robo Advice improves portfolio efficiency of retail investors and increases trading frequency Contribution: First study on Robo Advice with multiple identification strategies (within-subject plus DiD) Novel results on Robo impact Bräuer, Hackethal, Scheurle (2018). Fund Savings Plan Choices with and without Robo-Advice Main findings: In fund savings plans, robo-advice increases the share of ETFs (+24.3%) and the conditional choice of lowest-cost ETFs (+12.6%) Contribution: Introduce the role of defaults and guided decision making to a general savings domain Augmented methodology compared to existing default literature (panel setting, controls) Scheurle and Hackethal (2017): Can Robo-Advice Invitations Spur Stock Market Participation? Main findings: Inviting individuals to use robo-advice increases propensity to enter the stock market by a factor of 1.8 Contribution: Extend stock market participation literature to effects of marketing and convenience Study novel type of advice 4
Robo Advisor in question enables guided (FCA 2016) investment decisions in passive and active funds Low Medium Input field Default choice Step 1 Step 2 Step 3 Choose investment type Savings plan Lump sum 35% 5% 60% List of default ETFs/ active funds from broad set of issuers (diversified, low cost) Please choose your desired risk-level Low Medium High Equity Bonds Commodities Equity Blackrock Europe 600 Allianz S&P 500 Bonds Please choose your desired investment horizon 5 Years Please choose fund type ETF Additional fund universe sorted by total expense ratio (TER) Equity Europe TER Please choose your investment amount/ contribution rate 10k/3k/100 Euro Mutual fund Mix Blackrock Commerzbank Deutsche Bank ishares 0.09 0.10 0.11 0.12 5
Unique dataset from a large German retail bank covering more than 200,000 clients (mostly non-participators) Large German online bank with full range of retail banking products (e.g., checking/savings accounts, credit/debit cards, credit, trading, personal advice) Bank started robo-adviser in May 2014: 5,084 clients have joined the service in the following 13 months 56% were bank clients before We observe demographics, login activities, checking / savings accounts, portfolios and trades from 01/2003-12/2015 for four types of clients (data update to 12/2017) clients that are not investing (non-participators) do-it-yourself investors who do not seek advice Clients who seek human advice on investments robo-advised clients The bank has conducted four large marketing campaigns to existing clients We observe when and how clients were contacted for each campaign the bank selected target clients, based on internally-generated customer groups...... and a control group of clients who were not contacted we can investigate the behaviors of clients before and after joining with zero PF holdings before joining the same client across different accounts 6
Descriptive Statistics by Investor Type before launch of Robo- Advice (January 2012 - December 2013) younger less wealth less risk 7
Question 1: Do Robo Advisors increase financial risk-taking? Financial advisors have influence over their clients risk-taking: Gennaioli, Shleifer and Vishny (2015) hypothesize that advisors can reduce client anxiety and promote higher financial risk-taking Chalmers and Reuter (2015): Plan participants who invest through a broker hold portfolios with higher total risk Robo Advisors might promote more risk-taking: By providing easier access to low cost equity investments (e.g., ETFs) By promoting a diversified portfolio perspective They might not: The lack of a human advisor could limit the benefits associated with hand-holding and anxiety reduction 8
Question 2: Do Robo Advisors mitigate behavioral biases? Advised clients exhibit many of the well-known investor mistakes: Linnainmaa, Melzer and Previtero (2017 wp) find that advisers themselves share these same behaviors in their own personal portfolio Hoechle, Ruenzi, Schaub, and Schmid (2017 RoF) find that advisers reduce investment biases, but hurt trading performance even more Robo Advisors might reduce behavioral biases by providing access to cost-effective, well-diversified, passive investments They might not (as much as hoped): If Robo Advisers elicit lower level of client compliance than traditional advisors Bhattacharya, Hackethal, Kaesler, Loos, and Meyer (2012 RFS) highlight that advice is useless if not followed Bhattacharya, Hackethal, Loos, and Meyer (2017 RoF): investors using ETFs do not improve their portfolio performance 9
Pre and post analysis for Robo Advice users: Descriptives Higher AuM in funds Higher share of investments Higher turnover Better diversification Note: An investor to enter into this table has to be present for at least 6 months before and after taking Robo Advice in the period from 2012-2015. 10
Pre and post analysis: regressions confirm descriptive results wrt. share of portfolio holdings (vs deposits) 10%pts higher investments Note: All the errors are double-clustered at the investor and month level. An investor to enter into this regression has to be present for at least 6 months before and after taking Robo Advice in the period from 2012-2015 11
Pre and post analysis: regressions confirm descriptive results wrt. diversification and biases >10%pts less domestic equity >20%pts more index products Note: All the errors are double-clustered at the investor and month level. An investor to enter into this regression has to be present for at least 6 months before and after taking Robo Avice in the period from 2012-2015 12
Endogeneity concerns: Investors joining directly after campaigns less likely to adapt behavior irrespective of Robo Advice Note: All the errors are double-clustered at the investor and month level. An investor to enter into this regression has to be present for at least 6 months before and after taking robo-advice in the period from 2012-2015 13
Second paper zooms in on change in savings plan usage from Robo Advice Step 1 Choose investment type Please choose your desired risk-level 5 Savings plan Lump sum Low Medium High Please choose your desired investment horizon Years Please choose your investment amount/ contribution rate We compare savings plans characteristics for Robo clients to a matched sample of other DIY investors Matching conducted on demographic and ex ante PF characteristics About half of clients with Robo savings plans have also other savings plans running Robo Non-Robo # Investors 1,653 16,993 # Observed SPs 3,761 42,566 # Observed Robo SPs 1,728 - Mean contribution (EUR) 260.28 253.88 Mean contribution (% of total AUM) 6.2% 3.6% 10k/3k/100 Euro 14
Robo associated with more funds, diversification into bond funds and overall higher passive share and lower cost Matched Robo users Matched Robo non-users Analyses are based on a matched sample of robo-advice users (treatment group) and non-users (control group). Matching is based on a nearest-neighbor matching using the Coarsened Exact Matching method (Iacus et al. 2012). As matching variables we use pre-treatment means of the variables gender, age, relationship length, total assets under management, income proxy, number of logins as well as savings plan contribution rate, passive share, number of funds and equity share. 15
Robo defaults associated with higher ex ante Sharpe Ratios of Robo users 90% before 90% before 90% after 90% after Analyses are based on a matched sample of robo-advice users (treatment group) and non-users (control group). Matching is based on a nearest-neighbor matching using the Coarsened Exact Matching method (Iacus et al. 2012). As matching variables we use pre-treatment means of the variables gender, age, relationship length, total assets under management, income proxy, number of logins as well as savings plan contribution rate, passive share, number of funds and equity share. 16
Participation Study: customers are invited to use robo-advice identification is based on random timing of treatment and on random exclusion of control group There are 3 marketing channels for robo-advice invitations Electronic message for which customers are randomly selected n customers 2 Selection 1 Random treatment (absolute) Physical letter mailing Phone call Criteria vary by campaign: min. EUR 0 AuM min. EUR 5,000 AuM min. EUR 2,000 in savings or on checking account min EUR 1,000/ min. EUR 2,000 monthly income or on checking account A B Treated over time Not treated Staggered invitations for roboadvice Selected and then randomly excluded from treatment 58.361 896 1 Further inclusion criteria in my analysis: Customer has not been approached for personal advice campaigns before introduction of robo-advice, customer has not acquired personal advice from the bank or external advice, AuM is positive 2 Customers who were selected for robo-advice invitations, have been customer at the introduction of robo-advice (June 2014) and have not participated in the stock market one month before introduction of the robo-advice tool 17
Cox semiparametric hazard model applied to measure impact of robo-advice invitations on participation hazard Empirical model specification: Cox proportional hazards model Survivor function Hazard function Failure variable Explanatory variables (Z) Advantages of hazard model Advantages of Cox model S(t) / Probability of surviving beyond time t conditional on no failure before time t h $ & ' = h " $ exp (& ' -) First active stock market participation of customers without security transfer to other bank Ø Invitation to use robo-advice Ø Indicators for 1-10 days, 11-20 days, 21-45 days, 46-60 days, 61-90 days, 90+ days postinvitation Ø Comparison with invitations to use personal advice Ø Controls: Ø Demographic data, account value, account activity Ø German stock market (DAX) return and volatility Ø Year and cohort (year-of-birth cluster) fixed effects Ø Other investment marketing campaigns Ø Combination of separate binary-outcome analyses at failure times, taking into account dependency of person-day observations Ø Observation relative to time, accounting for left- and right-censoring Ø Straight-forward interpretability of hazard ratios as factors on baseline probability Ø No parameterization on the baseline hazard rate h " ($) Ø Only assumption: proportional hazards assumption, i.e. equal relationship length and coefficients imply same risk of participation 18
Hazard ratios: the effect is stronger for robo-advice campaigns than for personal advice campaigns Factor of 2.42 increase in propensity to enter stock market 1-10 days postinvitation Weighted factor of 1.86 until 2 months post-invitation Hazard ratios of invitations to advice on stock market participation The figure depicts hazard ratios and their 95% confidence intervals 10, 20, 45, 60, 90, and 91+ days after an invitation to use robo-advice or personal advice. 19
Concluding Remarks Key finding Implications 1 Investor Behavior Increase in (risky) portfolio holdings, better diversification and higher turnover Robo Advice helps debiasing retail investors 2 Savings Plans and role of defaults Increase in number of funds and flows into low-cost ETFs Robo Advice associated with higher AuM Defaults are powerful can be scrutinized by supervisors 3 Participation Increase in propensity to enter the stock market after Robo Advice invitations especially for less wealthy Robo Advice reduces perceived participation cost (convenience, guidance, transparency) 20
BACKUP 21
BACKUP Behavior: Outcome Variables by Investor Type before launch of Robo-Advice (January 2012 - December 2013) 22
BACKUP Participation: Three identification strategies to deal with endogeniety concerns 1. A pre and post analysis, using investor fixed effects Controlling for non time-varying unobserved heterogeneity at the investor level Using timing of marketing campaigns to account for the endogeneity in joining 2. A generalized difference-in-differences approach Relying on different matched control groups (self-directed clients and clients with human advisor) Using late joiners as controls for early joiners 3. A (quasi) randomized controlled trial Comparing the effects on the treatment group (106,000 in five campaigns) vs. the control groups (some 7,000) From Intention to Treat (ITT) to Treatment on the Treated (TOT) 23
The dataset allows to observe the degree to which a bank relationship is the main bank for a customer Year 2014 Evidence for main bank Checking account existence At least half of allowance for capital income tax exemption 68.9% 24.2% >=2 logins monthly 64.5% No security transfer to or from other bank (conditional on security holding; full observation period) 57.1% 24
BACKUP Participation: Results for interaction with wealth consistent with robo-advice lowering participation cost Cox semiparametric hazard model Failure: First stock market participation Included customers: Selected for robo-advice invitation and/or personal advice invitation Controls: Demographic, Wealth, Account activity, stock market, Year FE, Cohort FE Reported: hazard ratios and t- statistics Standard errors are linear estimates from a stratified sampling approach and clustered by individual N=76,927 ***,**,* denote significance at 10%, 1%, and 0.1% respectively Product invitations (days) I:Robo invite 45 i,t #I:AuM_Eurok 0-25 i,t-30 1.7493*** (11.70) I:Robo invite 45 i,t #I:AuM_Eurok 25-50 i,t-30 2.0915*** (5.86) I:Robo invite 45 i,t #I:AuM_Eurok 50-100 i,t-30 2.6871*** (5.25) I:Robo invite 45 i,t #I:AuM_Eurok 100-150 i,t-30 2.5076** (2.33) I:Robo invite 45 i,t #I:AuM_Eurok 150- i,t-30 1.0456 (0.05) I:Personal adv invite 45 i,t #I:AuM_Eurok 0-25 i,t-30 1.0880 (0.49) I:Personal adv invite 45 i,t #I:AuM_Eurok 25-50 i,t-30 0.5886** (-2.04) I:Personal adv invite 45 i,t #I:AuM_Eurok 50-100 i,t-30 0.7739 (-0.77) I:Personal adv invite 45 i,t #I:AuM_Eurok 100-150 i,t-30 1.2594 (0.48) I:Personal adv invite 45 i,t #I:AuM_Eurok 150- i,t-30 1.8718 (1.25) Model F-ratio 112.27 (1) 25
Male, younger and stock market participating individuals are slightly overrepresented in the dataset compared to the German population As of end of year 2014 Sample size Demographic data Dataset HFCS 1 Germany 2 n 202,478 4,461 82.1mn Male 64.7% 47.8% 49.0% Male age (years) 46.0 52.1 42.9 Female age (years) 43.7 52.7 45.6 Investment Stock market participation (total) 43.5% Observed participation switch 2003-2015 conditional on prior nonparticipants 20.6% 18.1% 4 13.1% 5 1 Eurosystem Household Finance and Consumption Network (2016) 2 Federal Institute for Population Research (2014) 4 Includes bond mutual funds 5 DAI - Deutsches Aktieninstitut (2015); fraction of adult population 26