Soft Regulation for Financial Advisors

2.1 Introduction to the Game

This section first explains the setup of the Baseline treatment and then further elaborates on the treatment variations of the Transparency and Fee treatments. In each of the 18 rounds of the game, the advisor will initially be assigned a project type. Projects differ in their quality type and we refer to low quality projects as ‘unfavorable’ and to high quality projects as ‘favorable’. Favorable projects are defined as having a high probability of being successful while unfavorable projects are defined as having a low probability of being successful. The project type is private information of the advisor. Advisors are then randomly assigned to an investor in each round of the game. If the investor invests in a project that was recommended by the advisor, the advisor will receive a commission. The investor can decide to invest in the project of the advisor, invest herself or not invest at all. The incentives of the two parties are not aligned when the advisor receives an unfavorable project.

At the beginning of the game, advisors are divided into high and low ability types, where high ability types are more likely to receive the favorable project. Investors receive information on the past number of successful and failed projects of the advisor.^[1] Investors receive this information along with the advisor recommendation before choosing whether to invest with the advisor. This is meant to represent the most basic information that investors have access to, such as the past performance of a managed fund for example. This can give the investor some indication about the ability of the advisor. However, a random element remains as returns are stochastic, meaning that the advisor ability type is never fully revealed. This is meant to emulate a real life setting in which the ability of the advisor can be estimated after looking at his history of investment projects but is never truly known.

This setup leads to two problems for the regulator. On the one hand, malicious advisors are able to exploit their informational advantage to the detriment of investors by recommending the unfavorable project. On the other hand, investors are unaware of the true project type, meaning that they may decline the opportunity to invest even if it would be in their best interest. This leads to the over-funding of unfavorable projects, under-funding of favorable projects and a net loss of expected utility for the investor from a regulatory perspective.

2.2 Experimental Overview

The experiment was conducted through the WZB laboratory in Berlin, Germany in September of 2021. Participants were allocated to a random treatment and all instructions for the game were provided before the game began. For the Baseline and Fee treatment there are 1728 observations (8 sessions), for the Transparency treatment there are 1512 (7 sessions). One session in the Transparency treatment in which we did not have the full participant number was dropped. This amounts to a total of 23 sessions with 12 participants per session, leading to a total of 276 participants. Half of the participants are investors, while the other half are advisors.

The experiment was programmed on o-tree (Chen et al. 2016) while participants were recruited using ORSEE (Greiner 2015). On average, the payout of the experiment was 15.04 Euros while sessions lasted between 45 and 55 minutes. After the main game, subjects additionally completed a trust game (Berg et al. 1995), a multiple price list procedure (Holt and Laury 2002) as well as a Big 5 measure and a survey. The final payoffs were determined after the completion of the experiment and paid to subjects via pay-pal. Table 1 provides a summary of the observations and payouts of the experiment.

2.3 Players and Timing of the Game

Before the game starts, participants are split into advisors (A) and investors (I) wherein half the participants are investors and half are advisors. Advisors are further split equally into two types: low ability (A _L ) and high ability (A _H ). Subjects are informed that they are going to play between 15 and 20 rounds, while the actual round number is 18.^[2] This was done to avoid end game effects. In the Baseline treatment each round t has the following timing, which is again shown in Figure 1 below:

1. Each investor I is randomly matched to an advisor A. Nature assigns A a project pro _t  ∈ {Fav, UnFav} where A _H is assigned a favorable project (Fav) with high probability p _H  = 0.6 and A _L is assigned a favorable project with low probability p _L  = 0.37. Both ability type and project type are private information of the advisor.

2. The advisor decides whether to recommend (1) or not recommend (0) the project, knowing its type (r _t  ∈ {1, 0}). If there was a recommendation, the investor gets information on the number of past successes and failures of the advisor.

3. The investor makes her investment decision, inv _t  ∈ {Adv, Self, Not}. In cases of a recommendation the investor can invest in the project of the advisor (Adv), invest herself (Self) or not invest (Not). In cases without a recommendation the investor can invest herself or not invest.

4. The success (1) or failure (0) of the project, suc _t  ∈ {1, 0} is stochastically determined according to the corresponding success probabilities. Payoffs are subsequently determined according to the outcome of the game.

Figure 1:

Baseline timing of the game.

An experimental session consists of 12 participants: 6 investors and 6 advisors. Participants are matched randomly in each round with repeated re-matching. The probability of being re-matched with the same participants again in the next round is thus 1 6 for any of the future rounds.

2.4 Payoffs and Probabilities

The payoff function of the advisor is as follows:

The advisor receives a base salary of 2000 points and additionally receives a commission of 3500 points if the investor decides to invest in his project. Since the investor can only invest after a recommendation, the advisor always has a financial incentive to recommend the project no matter what type it is. The payoff function of the investor is the following:

If the investor does not invest, she receives a constant payoff of 3000 points. In cases of investment, either with the advisor or by herself, the payoff depends on whether the project is successful. A successful project rewards the investor with 5500 points whereas an unsuccessful project only rewards her with 2000 points. Investors are thus best off if they invest in a successful project and worst off if they invest in an unsuccessful project.

High ability advisors have a high probability of getting the favorable project (p _H  = 0.6), whereas low ability advisors have a low probability of receiving the favorable project (p _L  = 0.37). The probability of success given the advisor and project type is given in Figure 2 below. The success probability in a self investment is constant at 0.5. This means that investors are strictly better off investing in a favorable project than investing themselves but worse off investing in an unfavorable project than investing themselves. Advisors thus know that they are harming the investor if they recommend the unfavorable project. Misconduct is then defined as recommending the unfavorable project.

Figure 2:

Assignment of projects to advisor types.

Given the probabilities of the advisor types, investors are better off investing with a high ability advisor compared to self investing but worse off investing with a low ability advisor compared to self investing. All players are made aware of this and are given the probabilities for each advisor type to receive a successful and unsuccessful project.^[3] This way, investors do not need to make complex calculations about each advisor type.

This complex setting is meant to emulate real life investment scenarios in which retail investors also have to form intuitive beliefs about their investment strategies. The goal of our experimental instructions was to give participants an intuitive understanding of the game rather than expecting them to perform exact calculations. Probabilities and payoffs were explicitly provided in the instructions and in each round giving participants the opportunity to perform exact calculations. However, participants were primarily encouraged to recognize that more successful projects compared to unsuccessful ones would imply a greater probability that the advisor is a high ability type. The experimental results indicate that participants understood this.^[4]

At the end of the game, 3 randomly selected rounds are paid out to the subjects where 1000 points amounted to one Euro.

2.5 Transparency Treatment

A Transparency treatment and a Fee treatment were introduced in order to test changes in misconduct and investment behavior across different regulatory environments. In the Transparency treatment advisors are able to pay a small sum of 100 points in order to provide a track record of their honesty after the first round of the game. This track record is shown through a ‘truthfulness rate’ which will be provided to investors before they make their investment decision if advisors decide to use the transparency option. If advisors do not use the transparency option, investors will not be informed about the truthfulness rate. The rate is an imperfect measure of past misconduct behavior that is meant to represent a voluntary external audit of the advisor by a third party. It is calculated as follows:

where N represents the number of past rounds in which the specified set up occurred. This measure captures the alignment between recommendations of advisors and the actual project outcomes. Intuitively, the truthfulness rate rises when the advisor either recommends a project that turns out to be successful or if he does not recommend a project that ends up being unsuccessful. Favorable projects are likely to be successful while unfavorable projects are unlikely to be successful. For this reason, the truthfulness rate rises on average if the advisor recommends a favorable project or if he does not recommend an unfavorable project. All players in this treatment were given an intuitive explanation of how the truthfulness rate works. They were asked to primarily understand that, on average, a higher truthfulness rate is indicative of past truthfulness.^[5]

The transparency option was only available after the first round, as no prior decisions have been made in the first round. Investors are aware that advisors are given this transparency option. If the advisor does not use this option for any given round other than the first round, investors are told that the advisor chose to not use this option for the current round.

This additional reputational component gives the advisor an incentive to tell the truth as well as let the investor know about the truthfulness of the advisor in the past, allowing her to select the investments with advisors that have a better reputation. This imperfect measure of truthfulness was chosen as it best reflects a real world situation in which external parties like auditors have limited information about advisor behavior. In such a setting, they are not able to provide investors with a perfect measure of misconduct. Transparency may for example represent a voluntary advisor paid audit that is used to increase transparency as well as provide incentives for truth telling through reputation. The advisor decides whether to pay the transparency payment after recommending the project in step 3 (see Section 2.3). If the advisor does not recommend the project, there is no way to present the track record for the advisor, i.e. the transparency option is not available as investors cannot invest with the advisor. Figure 3 below shows the timing of the Transparency treatment.

Figure 3:

Transparency timing of the game.

2.6 Fee Treatment

In the Fee treatment investors decide whether to pay a small fee of 200 points to the advisor in order to receive information about the project. This fee was deducted from the final payoff of investors. If an investor does not pay the fee, the advisor cannot make a recommendation and investors can only self invest or not invest.

The mechanism is designed to foster goodwill or trust in those advisor-investor pairs in which the fee was paid and thereby hopes to reduce advisor misconduct. This treatment is meant to represent a fee based business model, in which advisors receive a fixed payment for their service in addition to the income from their commission. Advisors are aware that investors have to pay the fee and that they will not be given the option to recommend their project if the fee is not paid. They are also specifically told again that investors paid their fee in each round if this was the case. This is meant to foster reciprocity and potentially induce advisors to act more truthfully. The timeline of the Fee treatment when the fee is not paid is shown in Figure 4 below.

Figure 4:

Fee treatment timing when the fee is not paid.

If the fee is not paid, advisors are unable to receive a commission since investors cannot invest with the advisor. If the fee is paid, the timing can be seen in Figure 5 below. This timing is now similar to the Baseline game except that the fee is paid before the recommendation, which is meant to influence both advisor and investor decisions.

Figure 5:

Fee treatment timing when the fee is paid.

4.1 Data Overview

Our data shows that the distribution of favorable and unfavorable projects is very similar across the treatments,^[7] which allows us to look at the efficiency of each treatment and underscores the effective randomization of each treatment. The allocation of favorable projects across advisor ability is also similar to the predicted ratios, as high ability advisors receive favorable projects in 62 % of all cases while low ability advisors receive favorable projects in 36 % of all cases.

Next, we provide an overview of the Fee and Transparency treatments individually to highlight some key insights that are relevant for the rest of the analysis. For the Fee treatment, investor fees were paid only approximately 50 % of the time, meaning that the number of projects that investors could invest in was cut in half. Furthermore, high ability advisors had their fee paid in 59 % of all cases while low ability advisors only had their fee paid in 40 % of all cases. As a result, the number of project recommendations for the Fee treatment is significantly reduced, especially for low ability advisors.

We then shift to the Transparency treatment and examine the use of the transparency option. Our data show that high ability advisors were 9 percentage points more likely to use the transparency option.^[8] The transparency option can also be linked to greater honesty. On average, a recommendation of a project was favorable in 68 % of all cases when the advisor used the transparency option. If the transparency option was not used, only 57 % of projects were favorable. This highlights that advisors who use the transparency option tend to be more truthful and that the use of the transparency option signals truthfulness to investors.

As a final point in our overview, we analyze investment rates of investors after a recommendation by the advisor. The investment rate of the Fee treatment stands out at 92 % compared to 57 % in the Baseline and 54 % in the Transparency treatment. These results indicate that investors almost always invest after a recommendation in the Fee treatment. This can be explained by two reasons. First, completely rational investors would only pay the fee if they plan to invest after a recommendation. The choice to invest should thus have already been made before the fee was paid and the recommendation was even possible. Second, the reason behind the high investment rate could involve investors acting irrationally or in a behavioral manner. Investors might, for instance, fail to regard the fee they have already paid as a sunk cost. Alternatively they may begin trusting the advisor more once they paid the fee, expecting some form of reciprocity and therefore lower misconduct. Either way, the findings imply that a recommendation in the Fee treatment is very likely to receive investment, which would also mean that misconduct is likely to receive investment. The investment rates in the Baseline and Transparency treatments are very similar.

4.2 Misconduct by Advisor Ability

We initially investigate whether high ability advisors engage in less misconduct compared to low ability advisors as conjectured in hypothesis H1. Misconduct is a dummy variable which is equal to 1 if the advisor recommended the unfavorable project and 0 otherwise. The descriptive statistics are in line with H1, since high ability advisors misconduct with a probability of around 40 % while low ability advisors misconduct with a probability of around 50 %. This misconduct gap in ability types is largest in the Transparency treatment at 13 percentage points, compared to around 9 percentage points in the Baseline treatment and 8 percentage points in the Fee treatment.^[9]

In order to take the panel structure (repeated decisions of the same subject) and the possible correlation within sessions into account, we will (in this and the following subsections) employ multilevel models with clustering at both the subject- and session-level. Regression Table 2 provides the results of a multilevel logit model using misconduct as the dependent variable. This regression shows that advisor ability has a statistically significant effect on the likelihood of misconduct. In the regression, the advisor ability dummy is significant at the 10 percent level without controls and becomes significant at 5 percent when adding controls. This indicates a negative relationship between high ability advisors and the likelihood of misconduct. The most important significant controls are the age of the participants, which is negatively linked to misconduct, and the round number which is positively linked to misconduct. Therefore there is more misconduct by younger participants and in later rounds of the game.

The average partial effect of advisor ability on misconduct in the regression model (1) without controls is around −0.1. This negative effect becomes slightly larger (−0.124) when participant characteristics and round numbers are added in model (2). Finally, when also controlling for a risk measure in model (3), the average partial effect of advisor ability is even stronger (−0.138). These effects imply that being classified as a high ability advisor reduces the propensity to engage in misconduct by approximately 10–14 percentage points. Considering that high ability advisors misconduct approximately 40 % of the time, these reductions represent a substantial relative decrease in the likelihood of misconduct.

To check the robustness of these findings, several other regression models were run, in particular logit models with standard errors clustered at the subject-level^[10] (see regression Table 2 in Section A.2 of the Online Appendix), models with random-effects at the subject-level and models with random-effects at the session-level (these models are additionally available on request). All models provide similar findings concerning both statistical significance and magnitude of the coefficients.

4.3 Favorable Investments

In our next step, we evaluate the efficiency of each treatment by analyzing the distribution of favorable and unfavorable investments. We conjecture in H2 that the treatments lead to an increase in the number of favorable investments and a decrease in the number of unfavorable investments. For the analysis of favorable investments, we focus exclusively on instances where the project assigned is favorable. We define a binary variable for favorable investment that takes the value 1 if a favorable project is recommended and receives investment, and 0 otherwise. The same methodology applies to unfavorable investments, where we consider only unfavorable projects and define the binary variable accordingly.

The fraction of favorable investments is 60 % in Baseline, 58 % in Transparency and 48 % in the Fee treatment, highlighting a significant reduction of such investments in the Fee treatment. To formally test this hypothesis we regressed this favorable investment indicator on the treatment dummies as well as controls.

Regression Table 3 shows the results from a multilevel logit model to estimate the treatment effects for favorable investments; once again, clustering occurs at both the subject and session level. These regression results clearly indicate that favorable investment rates are lower in the Fee treatment compared to the Baseline treatment.^[11] Additionally, the round number as well as the number of successful projects of the advisor influence the favorable investment choice. The number of successful projects is positively linked to favorable investments while the round number is negatively linked to favorable investment. The average partial effects for the Fee treatment in regression Table 3 are −0.105, −0.174, and −0.185 for models (1), (2) and (3) respectively. This implies that the Fee treatment led to a reduction in favorable investments of around 10–18 percentage points compared to Baseline. From a regulatory perspective this outcome highlights a considerable loss in efficiency. This reduction in efficiency is of course strongly driven by the lower amount of project recommendations in the Fee treatment, as the fee was paid only in around half of all cases. Nevertheless, this resulted in a lower total amount of favorable projects that received investment, which would make investors worse off. Robustness checks using models with clustered standard errors at the subject level, as well as models with individual random effects for subject and session levels, confirm these results (results are available upon request). The regression results for the model with clustered standard errors at the subject level are shown in regression Table 3 in Section A.2 of the Online Appendix.

Unfavorable investments are almost identical in the different treatments, which highlights the inability of the Transparency and Fee treatments to increase efficiency. The Baseline treatment had 23 % investment in unfavorable projects, whereas the Transparency and Fee treatments had unfavorable investments of 21 % and 20 % respectively. Various regression analyses all indicate that there is no significant difference between the treatments in terms of unfavorable investments.^[12]

The inability of the Fee treatment to decrease unfavorable investments is particularly interesting as investors were already limiting the number of available projects by only paying the fee in approximately half of all rounds. Nonetheless, the number of investments in unfavorable projects remains almost the same. A possible drawback of the Fee treatment is that advisors who get their fee paid are very likely to receive investment after engaging in misconduct due to the high investment rates after recommendation. In the Fee treatment, investors invested in an unfavorable project that was recommended in 90 % of all cases compared to just 52 % in the Baseline treatment. This means that high ability advisors with low behavioral or moral lying cost may be very successful at engaging in misconduct, as they have their fee paid more often. In real life scenarios, this could lead to self-selection of such advisors into a fee based structure and therefore result in increasing investment in unfavorable projects. Thus, malicious high ability advisors with low moral lying costs are able to exploit a fee based environment to increase their successful misconduct behavior to the detriment of investors.

4.4 Treatment Differences in Misconduct

Lastly we examine misconduct rates across treatments in order to evaluate whether the treatments are able to reduce misconduct as hypothesized in H3. Surprisingly, misconduct rates are very similar across the treatments and actually slightly higher in the Transparency and Fee treatments. The Baseline treatment shows a misconduct rate of 44 %, while the Transparency and Fee treatments have rates of 48 % and 49 %, respectively. This suggests that neither treatment significantly reduces misconduct and may even slightly increase it. A multitude of regression analyses confirm that there are no statistically significant differences between the misconduct rates of the Baseline treatment and the misconduct of the other treatments at the 5 % significant level for any of our regressions.^[13] This is to be expected given that the changes in overall misconduct rates are rather small.

Regression Table 4 demonstrates this result for a multilevel logit model using misconduct as its dependent variable. Once again, clustering occurs at both the subject and session level. We confirm this result in various other regression specifications. Most importantly, logit models with standard errors clustered at the subject level as well as random effects models with random effects at the subject and session level individually. Regression Table 4 in Section A.2 the Appendix shows the specification using clustered standard errors at the subject level.

In order to gain further insight into why there were no changes in misconduct we start by splitting the treatment data by round number and advisor ability. While the Baseline and Fee treatments have minor increases in misconduct in the last rounds, the misconduct increase in the Transparency treatment is drastic. This increase in misconduct in the last rounds can be clearly seen in Figure 6 using a 3 round split of the data. The increase in misconduct in the Transparency treatment is equal to 28 percentage points, whereas the increase in the Baseline and Fee treatment is 7 and 9 percentage points respectively. Other splits show the same trend. While increases in the Baseline and Fee treatment can be attributed to decreased sensitivity as well as learning effects, the comparatively drastic increase in the Transparency treatment points to other factors playing a role.

Figure 6:

Misconduct increase in later rounds. Note: This figure illustrates the increase in misconduct from rounds 7–12 to rounds 13–18 of the game. Each bar represents the proportion of the unfavorable projects that were recommended by advisors for the various treatments.

Various regression analyses reveal that the increase in misconduct from periods 7–12 to periods 13–18 is only statistically significant in the Transparency treatment and not in the other two treatments. This result is obtained by splitting the round numbers into three separate groups (1–6, 7–12 and 13–18) and entering corresponding dummy variables into the regression model. Regression Tables 5 and 6 in Section A.2 of the Appendix show models with interaction terms between the round number and the treatment dummies. Subsequent Wald test of significance between rounds 7–12 and 13–18 for each treatment reveal that there is only a significant difference at the 5 % level for the Transparency treatment (p-value < 0.01).

A possible explanation for this is that advisors strategically exploit their truthfulness rate in order to engage in misconduct in the final periods. We next look specifically into the Transparency treatment and examine the misconduct of advisors with above average truthfulness rates, who would be seen as being more truthful by investors. For these advisors, misconduct rates are constant at roughly 35 % for the first 12 rounds, however, misconduct rises drastically for the later rounds of the game. For the rounds 13–18 these advisors increase their misconduct to 61 % (see Table 1 in the Appendix Section A.2). It can thus be seen that advisors who were particularly honest in their initial periods almost double their misconduct in the last periods. Our interpretation of these results is that advisors used the transparency option in order to build up a reputation for additional investments in the final rounds and thereby ‘game the system’. This essentially undermines the effectiveness of the identifiability mechanism, as misconduct numbers drastically increase in later rounds.

Next, we examine the Fee treatment individually and find that a form of negative reciprocity might be at play which leads to greater misconduct numbers among advisors that did not receive their fee frequently. Figure 7 shows average misconduct ratios in the Fee treatment, segmented by round number and whether the frequency of fees paid in the first nine rounds was above or below the overall participant average for those rounds. The figure shows that advisors who received less than the average amount of fee payments in the first half of the experiment had a more pronounced increase in misconduct compared to their counterparts who were paid more frequently. Despite similar levels of misconduct in the early rounds, this later divergence suggests a form of negative reciprocity from advisors receiving lower fees. Such a response undermines the intended effect of the voluntary fee mechanism, indicating that a failure to ‘meet the expectations of advisors’ in terms of fees can lead to increased misconduct behavior.

Figure 7:

Misconduct by fee paid in the fee treatment. Note: Figure 7 shows average misconduct ratios in the fee treatment, segmented by round number and whether the frequency of fees paid in the first nine rounds was above or below the overall participant average for those rounds. The bars on the left indicate the number of project recommendations of unfavorable projects (misconduct) for advisors who had their fee paid above average for the first 9 rounds while the bars on the right show the same for those advisors who had their fee paid less than average in the first 9 rounds. It can be seen that the initial recommendations are almost identical. However, those who had their investor fee paid less than average strongly increased their misconduct in later rounds compared to those who had their fee paid more frequently.

We can thus see that both the identifiability and the reciprocity mechanisms experience significant shortcomings in settings with different advisor types. Also, the incentives or behavioral influence factors of advisors may be altered depending on the advisor type. This leads to the current situation in which neither mechanism is able to reduce misconduct.