In a previous post I posed the following bet:
Suppose you have $100 and are offered a gamble involving a series of coin flips. For each flip, heads will increase your wealth by 50%. Tails will decrease it by 40%. Flip 100 times.
The changes in wealth under a sequence of flips of this nature is “nonergodic”, as the expected value of the bet does not converge with its timeaverage growth rate. The bet has a positive expected value, 5% of the bettor’s wealth per flip, and the ensemble average across a large enough population will approximate this expected value in growth in overall wealth. But, the timeaverage growth rate for an individual is approximately a loss of 5% of their wealth with each flip. Most individuals will experience a loss, and in the longrun everyone will. (To understand why this is so, see my primer post on ergodicity economics.)
That many people decline bets of this nature suggests that there may be some wisdom in our decision making process. But what is that process?
Are we risk averse?
As I noted in that previous post, economists have a readily available explanation for the rejection of this bet. People are risk averse expected utility maximisers. As I wrote there:
A risk averse person will value the expected outcome of a gamble lower than the same sum with certainty.
Risk aversion can be represented through the concept of utility, where each level of wealth gives subjective value (utility) for the gambler. If people maximise utility instead of the value of a gamble, it is possible that a person would reject the bet.
For example, one common utility function to represent a risk averse individual is the logarithm of their wealth. If we apply the log utility function to the gamble above, the gambler will reject the offer of the coin flip. [The maths here is simply that the expected utility of the gamble is 0.5ln(150) + 0.5ln(60)=4.55, which is less than the utility of the sure $100, ln(100)=4.61.]
The concept of a risk averse expected utility maximiser with a utility function such as the logarithmic has been a staple explanation for many decisions. The St Petersberg Paradox is one such problem, with that series of bets rarely valued above $10 despite the infinite expected value of the bet. (It is another nonergodic system.)
But do we need an expected utility function to provide us with such risk aversion? Would a more parsimonious explanation for the rejection of the bet simply be that the person is seeking to maximise the growth rate of their wealth. With that objective and a timeaverage growth rate of minus 5%, rejection is the obvious thing to do. There is no need for an expected utility function. Rather, the person simply needs a way of deciding whether accepting the bet will maximise the growthrate of their wealth.
An interesting alignment of economic history and ergodicity economics occurs here. One of the most commonly used expected utility functions is the logarithm (as noted above). People maximise utility by maximising the expected logarithm of their wealth.
Yet, the way to maximise the geometric growth rate of your wealth when facing a multiplicative bet is also to maximise the logarithm of your wealth. The calculations of the expected utility maximiser with a logarithmic utility function and of the timeaverage growthrate maximiser are the same.
As Ole Peters and Alexander Adamou write in their ergodicity economics lecture notes:
[E]xpected utility theory as we have presented it above is consistent with growth rate optimisation, provided a suitable pair of dynamic and utility function is used. For multiplicative dynamics, the necessary utility function is the logarithm. That this is the most widely used utility function in both theory and practice is a psychological fluke in the classic mindset; from our perspective it indicates that our brains have evolved to produce growthoptimal decisions in a world governed by multiplicative dynamics, i.e. where entities produce more of themselves.
This effectively means that many of the “puzzles” that expected utility maximisation has been used to solve can also be “solved” by growthrate optimisation. For instance, insurance or the St Petersberg puzzle provide a challenge for expected wealth optimisation, but are equivalently solved by assuming an expected log utility maximiser or a growthrate optimiser.
That these two concepts overlap raises a conundrum. An expected log utility maximiser looks much like a growthrate maximiser in their behaviour (noting that log utility is only one of many functional forms an expected utility maximiser could theoretically have). If we would expect to see the same decision under both expected log utility and growthrate maximisation in multiplicative dynamics, how can we differentiate the two?
Additive dynamics
Before I answer that question, I am going to detour into the world of additive dynamics. What if I offered you the following bet?
Suppose you have $100 and are offered a gamble involving a series of coin flips. For each flip, heads will increase your wealth by $50. Tails will decrease it by $40. Flip 100 times.
You can see the tweak from the original bet, with dollar sums rather than percentages. The first flip is effectively identical, but future bets will be additive on that result and always involve the same shift of $50 up or $40 down. In contrast, the earlier bet was multiplicative, in that the bettor’s wealth was multiplied by a common factor. As a result, the multiplicative bet scales up and down with wealth.
An important feature of this second series of flips is that the system is ergodic. The expected value of each flip is $5 (0.5*$500.5*$40=$5). The timeaverage growth rate is also $5.
Let’s simulate as we did for multiplicative bets in the ergodicity economics primer post, with 10,000 people starting with $100 and flipping the coin 100 times. The below plot shows the average wealth of the population, together with the paths of the first 20 of the 10,000 people (in red).
Average wealth of population and path of first 20 people
The individual growth paths cluster on ether side of the population average. After 100 periods, the mean wealth is $593 and the median $600. 86% of the population has gained in wealth. The wealthiest person has $2220, or 0.04% of the total wealth of the population. After 1000 rounds (not plotted here), the mean wealth is $5,095 and the median $5,100. All except one person has gained. The wealthiest person has $11,130, or 0.02% of the total wealth of the population. This alignment between the mean and median wealth, and the relatively equal distribution of wealth, are characteristic of an ergodic system.
Now for a wrinkle, which can be seen in the plotted figure. Of those first 20 people plotted on the chart, 12(!) had their wealth go into the negative over those 100 periods. An additional two of those first 20 go into the negative over the subsequent 900 periods. This is reflected across the broader population, with 5,439 dropping below zero in those first 100 periods. 5,684 drop below zero across the full 1000.
To the extent zero wealth is ruinous at the time it occurs (e.g. death, you cannot continue to play), that event is serious. If you only incur the consequences of your final position, the bet is somewhat less likely to result in ruin, but still presents a real threat of catastrophe.
So what would an expected utility maximiser do here? For a person with log utility, any probability of ruin over the course of the flips would lead them to reject the series of gambles. The log of zero is negative infinite, so that outweighs all other possible outcomes, whatever their magnitude or probability.
The growthrate maximiser would, if they didn’t fear ruin, accept the bet. The timeaverage growth of $5 per flip would pull them in. If ruin was feared and consequential, then they might also reject.
Risk and loss aversion in the two different worlds
To the title of my post, what light does this shed on risk or loss aversion?
Let us suppose humans are growthrate maximisers. In a multiplicative world, people would exhibit what is by definition risk averse behaviour – they prefer a certain sum to a gamble with the same expected value. This is a consequence of maximising the growth rate by maximising the expected logarithm of their wealth. This, however, has a different underlying rationale to explanations of log utility based on either psychology or the diminishing utility of wealth.
What of loss aversion, the concept that losses loom larger than gains? Risk aversion results in a phenomena that looks like loss aversion, in that losses are weighted more heavily due to the diminishing utility of additional wealth. However, loss aversion is a dislike of losses over and above that. It involves a “kink” in the utility curve, so should be observed for small amounts and result in a greater aversion to bets than risk aversion alone would predict.
The growthrate maximisation model would not lead us to predict loss aversion. Whatever their wealth, growthrate maximisation does not produce a marked difference between gains and losses beyond that induced by risk aversion. There is no “kink” at the reference point at which losses hurt more than gains are enjoyed. Are there any phenomena described as loss aversion which this theory would suggest are actually growthrate maximising behaviour? Not that I can think of.
In the additive world, things are more interesting. Growthrate maximisation is equivalent to wealth maximisation. People aren’t risk averse. (In fact, assuming only growth rate maximisation in an additive environment leaves much about their risk tolerance unspecified.) They simply take the positive value bets.
Here the broader evidence across experimental economics and psychology places a question mark over the claim (experiment described below excepting). People regularly reject positive value additive bets. There are ways to attempt to reconcile growthrate maximisation with these rejections. For instance, we could argue that these people are in a multiplicative world, of which the bet is only a small part, so the bet described as additive is actually part of a multiplicative dynamic. We know little about their broader circumstances. But even then, the rejected additive bets are often so favourable that even a growthmaximiser in a multiplicative dynamic would generally accept them.
Loss aversion is also not a prediction of growthrate maximising behaviour in the additive world. There is not only not any kink at the reference point. Losses and gains have the same weight no matter their scale.
We could add loss aversion to the growthrate maximiser in the additive environment by introducing an absorbing state at zero. The path to ruin can be quicker in an additive world than in a multiplicative as the bet sizes don’t scale down with diminished wealth, plus there is the possibility of losing absolutely everything. But what is the agent’s response to this potential for ruin? We would need to add some assumptions additional to that provided by a simple growthrate maximisation approach.
Ergodicity and behavioural economics
A short note here, because ergodicity economics in the twittersphere has been noted as the behavioural economics killer. I’ve already noted loss aversion, but I will state here that many behavioural phenomena remain to be explained even if we accept the foundational ergodicity concepts.
A core group of these behavioural phenomena involve framing, whereby presentation of effectively the same choice can result in different decisions. Status quo bias, the reflection effect, default effects, and the like, remain. So while ergodicity economics gives a new light to shine on decision making under uncertainty, it hasn’t suddenly solved the raft of behavioural puzzles that have emerged over the last seventy years.
Part of that is unsurprising. Much of the behavioural critique of expected utility theory is that our decisions don’t look like expected log utility maximisation decision making (or other similar functions). If that’s the case, those puzzles remain for a growthrate maximiser that maximises their expected log wealth in a multiplicative environment.
Distinguishing expected utility from growthrate maximisation: an experiment
Now to return to an earlier question. If we would expect to see the same decision under both expected utility and growthrate maximisation in multiplicative dynamics, how can we differentiate the two?
A group led by Oliver Hulme ran an experiment that sheds some interesting light on this question (branded the Copenhagen experiment in the twittersphere). The preprint reporting the experimental results is available on arXiv, with supporting materials and data on GitHub. Despite some of my questions below, this is a innovative and well thoughtout experiment.
The concept behind the experiment was to differentiate between three possible models of human decision making:
 Prospect theory, which includes features such as different risk aversion parameters in the gain and loss domains, and loss aversion.
 Isoelastic utility, a classic model of expected utility, of which log utility is a special case
 Time optimal utility, where changes in utility are determined by linear utility under additive dynamics and by logarithmic utility under multiplicative dynamics.
The third could be differentiated from the other two if the utility function effectively changes when the environment changes between additive and multiplicative dynamics.
The test this, the experimental procedure ran as follows.
Eighteen experimental subjects (actually 20, but two were excluded from analysis) participated in a series of gambles over two days. On one day they were exposed to a series of additive bets. The other day involved multiplicative bets. The order of the days was switched for some subjects. They were not directly informed of the nature of each day.
Each day consisted of a passive session, followed by an active session.
At the beginning of the passive session each experimental subject was endowed with 1000 Danish Krone (approx $150USD). They then watched a series of fractal images, each of which had an effect on their wealth. Their task through the passive session was to learn the effect of each of these fractals on their wealth. Some examples of the images are below:
Example fractal images
There were 9 fractals on each day, each with a unique effect on wealth. For the multiplicative day, the fractals changed wealth by the following factors: 0.447, 0.546. 0.668. 0.818, 1, 1.223, 1.496, 1.830, 2.239. On the additive day, the changes were: 428kr, 321kr, 214kr, 107kr, 0, 107kr, 214kr, 321kr, 428kr. The subjects observed a sequence of 333 fractals in the passive sequence (each fractal would appear 37 times), giving them a good feel for the effect of each fractal on their wealth.
As the outcome from an equal distribution of these fractals is no growth, each participant ended the 333rd passive fractal with their original wealth of 1000kr. They were then given one additional random fractal to set the level of wealth that they would commence the active session with.
During the active session, they would be shown two sets of two fractal images, as in Figure 3 below. The left and the right side represent two possible bets, each of which would involve a 50:50 probability of getting one of the two fractals and the effect of that fractal on their wealth. They were asked to choose one of the bets. For example, choosing the left two images would give them a 50:50 chance for each of those two images on the left.
Example choice
Hulme and friends then analysed these experimental results to determine which model best described the observed decisions.
The headline finding is well illustrated in the following chart, which provides the posterior probability of the parameter of risk aversion for each experimental subject. The blue distribution is that for the additive day, the red for the multiplicative day.
Posterior probability of the parameter of risk aversion
A risk aversion parameter of 0 gives us linear utility. A parameter of 1 is logarithmic utility. On that basis, the time optimal utility of ergodicity economics comes out looking strong. There is a clear change in risk aversion across most participants as they changed between the ergodic and nonergodic environments.
Hulme and friends also calculated the posterior probability of each model for each participant, with time optimal (the growth rate maximiser) generally having the stronger probability.
Posterior model probabilities
The authors write:
[T]o approximate time optimal behavior, different dynamics require different ergodicity mappings. Thus, when an agent faces a different dynamic, this should evoke the observation of a different utility function. This was observed, in that all subjects showed substantial changes in their estimated utility functions … Second, in shifting from additive to multiplicative dynamics, agents should become more risk averse. This was also observed in all subjects. Third, the predicted increase in risk aversion should be, in the dimensionless units of relative risk aversion, a step change of +1. The mean step change observed across the group was +1.001 (BCI_{95%}[0.829,1.172]). Third, to a first approximation, most (not all) participants modulated their utility functions from ~linear utility under additive dynamics, to ~logarithmic utility under multiplicative dynamics (Fig. 3d). Each of these utility functions are provably optimal for growing wealth under the dynamical setting they adapted to, and in this sense they are reflective of an approximation to time optimality. Finally, Bayesian model comparison revealed strong evidence for the time optimal model compared to both prospect theory and isoelastic utility models, respectively. The latter two models provide no explanation or prediction for how risk preferences should change when gamble dynamics change, and even formally preclude the possibility of maximising the time average growth rate when gamble dynamics do change. Congruent with this explanatory gap, both prospect theory and isoelastic utility models were relatively inadequate in predicting the choices of most participants.
My major question about the experiment concerns the localised nature of the growthrate maximisation. These people have lives outside of the experiment and existing wealth (of an unknown level). Yet the behaviour we observed in the multiplicative world was maximisation of the growth rate within the experiment. They effectively maximised the log utility of the inexperiment wealth.
If any of these subjects had any material wealth outside of the experiment and were general growthrate maximisers, their utility function within this experiment should be closer to linear, despite the multiplicative dynamics. The log function has material curvature for small wealth changes near zero. Once you are further up the logarithmic function (higher wealth), a short section of the function is approximately linear. Even though the stakes of this experiment are described as large (~$150USD with potential to win up to ~$600USD), they are likely not large within the context of the subjects’ broader wealth.
This point forms one of the central planks of the criticism of expected utility theory emerging from behavioural economics. People reject bets that, if they had any outside wealth, would be “nobrainers” for someone with log utility. Most of this evidence is gathered from experiments with additive dynamics, but there is also little evidence of linear utility in such circumstances.
Why did the Copenhagen experiment subjects adopt this narrow frame? It’s not clear, but the explanation will likely have to call on psychology or and an understanding of the experimental subjects’ broader circumstances.
Another line of critique comes from Adam Goldstein, who argues that “the dynamic version of EUT, multiperiod EUT, predicts the same change in risk aversion that EE predicts in a simplified model of CE [the Copenhagen experiment].”
Goldstein is right that EUT predicts a reduction in measured risk aversion in an additive environment. But Goldstein’s analysis depends on people being able to observe each flip and their change in wealth, and then changing their behaviour accordingly. If they could take the bets flip by flip, the first bet on its own is unattractive for a risk averse utility maximiser. But it is possible (indeed likely) for them to reach a level of wealth where a single bet is attractive (in this case, above a wealth of $200), in which case they can continue to accept. Conversely, if they head toward ruin, they can start to reject bets.
The possibility of getting up to a level of wealth where the bet becomes attractive can lead an expected logarithmic utility maximiser to accept the first bet due to the potential utility from later bets. The way to determine whether they will do this uses a technique called dynamic programming, which involves working from the last bet backward to work out the expected utility of each single bet.
However, I am not convinced this critique applies to the experiment by Hulme and friends. The experimental subjects never got to observe the changes in their wealth during the active session (although they might weakly infer the likely direction based on the favourability of the bets they had been exposed to). As a result, I’m not convinced that you would see the change in risk aversion observed in the experiment under an expected utility framework.
That inability to observe outcomes also makes the experiment a weaker examination of dynamics over time than it might otherwise be. It is in some ways a single period game where all outcomes are realised at the same time, multiplying or adding at that point. The absence of seeing how subjects act given a change in wealth removes the ability to see some of the distinguishing phenomena. For instance, the time optimal utility maximiser would not increase risk aversion after losing in the additive environment, whereas in that same environment the traditional utility maximiser would be more likely to reject when the bets become a larger proportion of their wealth. The prospect theory decision maker may become risk seeking if they perceived themselves to be in the domain of losses. The authors note that the lack of update is because they want to avoid mental accounting, but that is, of course, a feature of prospect theory (if I understand their use of the term mental accounting). (I should also say that I understand the lack of updating given the multiple purposes of the experiment, but it would be great to see that relaxed in future iterations.)
Goldstein also raised a second possible driver of the reduced risk aversion in the additive scenario. Experimental subjects were paid on the based on a random draw of 10 of their active gambles. If the final wealth for an experimental subject from those 10 gambles was negative, they would be given a new draw of 10 gambles. In effect, they were protected from the most severe risks in the additive case, which would reduce risk aversion.
One possible mitigant of this effect is that the experimental subjects were not explicitly told they could not lose (although they would likely have inferred they could not suffer loss). I am also not convinced that the strength of this effect would be enough to result in purely linear utility as was observed, but it should be accounted for. [Update: I simulated the potential payments of the participants based on the choices they actually made. Only around 4% of the potential payments involved a loss, which would have triggered the redraw. That affirms my view that, while it should be accounted for, it is unlikely to explain the experimental result.]
A related point is that the paths involving negative wealth were removed from the passive session on the additive day. This means that the subjects were not conditioned to see these negative potential consequences. In simulations I conducted of the additive passive day (see code below), around 90% of the simulations breach that zero lower bound. Twenty five percent breach the 5000kr upper bound (some of them the same paths that went below zero), leaving only 2% of the trials that could be provided to subjects on the passive day. In contrast, passive multiplicative paths could not be excluded for going below zero, despite providing a hairraising ride. Around 30% of the passive multiplicative paths that I simulated involve wealth dropping to less than 1kr (a 99.9% loss) and 60% to less than 10kr (a 99% loss). Then at the top end, 90% of the passive multiplicative paths went above $5000kr, leading to their exclusion.
The result is that the subjects were conditioned on a limited subset of additive paths that excluded the most negative moments (although also the 25% highest), and on multiplicative paths that excluded the most positive moments. This is a large asymmetry. Obviously, each person saw only one path, and they might have ended up with that combination anyhow, but the systematic conditioning of subjects with benign passive paths and harrowing multiplicative paths should be considered a potential factor in the response of subjects to those fractals.
It could be argued that despite the difference in paths, people are simply learning the effect of the fractals that they bet on. However, I am not convinced that experimental subjects would be unaffected by seeing the potential cumulative effect of these bets.
As a result, my preliminary view on this experiment is that it provides potential evidence that the dynamics of the environment can influence our model of decision making. However, the experimental results involve behaviour that don’t seem to be accounted for by any of the models, and it involves a conditioning process that I’m not completely sold on.
Some of my other observations on the experiment include:
 The experiment involved a number of “discrepant trials”, where linear utility should have generated one choice and log utility another. These trials generated moderate evidence against the hypothesis of linear utility under additive dynamics. You can also see in Figure 4 above (and in other parts of the paper) that the experimental subjects had mild risk aversion in the additive environment. Similarly, the coefficient of risk aversion in the multiplicative environment seems slightly greater than one – indicating more risk aversion than log utility. (Saying this, I wouldn’t read too much into these particular numbers.)

Although there is a consistent shift for most subjects between the two environments, there is a lot of variation in their degree of risk aversion. This could be due to outside factors, such as total wealth, but raises the question of how much idiosyncracy there is between people in their approaches to growthrate maximisation (or whatever else it is they are maximising).

I’m not convinced that the experiment had a design with the strength necessary to elicit a loss aversion parameter of prospect theory (assuming it exists). Every bet involved a choice between a two gambles involving a gain and a loss, rather than having a mix of gaingain and gainloss options that might highlight loss aversion. Shifting between those frames would also provide more power to tease out the risk aversion coefficients in the loss and gain domains. (I should note that I’m not confident that the experiment doesn’t have the necessary strength – I use the words “I’m not convinced” deliberately.)

There was a required choice between the gambles, which eliminates status quo effects, an arguable driver of many behavioural dynamics (as argued by David Gal).

The elicitation of preferences where people need to learn the probabilities through experience is one of the experimental circumstances where loss aversion has generally not been shown to occur (see this literature review by Yechiam and Hochman (pdf)). This provides another reason we might not not expect to elicit loss aversion in this experiment.

The set up is complicated. The subjects need to learn fractal relationships. Their payout is based on a random selection of 10 of their bets. The multiplicative environment harder to learn. Does uncertainty drive some of the increase in risk aversion?
Summary
Where does this leave us? I take the following lessons from ergodicity economics and the experimental evidence to date:
 The concept that simple growthrate maximisation results in the same observed behaviour as expected logarithmic utility maximisation in a multiplicative environment (possibly the world we live in), yet possibly provides a more parsimonious explanation, is important. This deserves much more research, including the question of whether this is a model on which we could build the broader decision making architecture. Would prospect theory look different if built on this foundation?

We don’t need to throw everything out of the window. Maximising expected utility through using the logarithm of wealth is equivalent to maximising the growth rate in a multiplicative environment. We can continue to use this functional form in much economics work, but should consider a different interpretation on its use.

For decisionmaking under uncertainty, there is a case for placing greater weight on the logarithmic “utility function” over other more highlyspecified utility models that do not maximise the growth rate. On this point, Paul Samuelson led a somewhat acrimonious debate about whether an investment strategy using the Kelly criterion – which maximises the geometric growth rate (discussed in my ergodicity economics primer post) – was an appropriate investment strategy. I’ll cover that debate in more detail in a future post, but one of Samuelson’s central points was that Kelly criterion investments are only optimal for an expected log utility maximiser, not for people with other utility functions. The ergodicity economics approach attempts to circumvent this debate by suggesting that our utility function is growth rate maximisation.

A behavioural response to possible absorbing states (i.e. ruin, death) would seem to require an addition to the growthrate maximisation model, rather than being directly derived from it. The growthrate maximisation model also says little about riskreturn tradeoffs, particularly in an additive environment. (This was also a point raised by Samuelson in the debate about the Kelly criterion, as growthrate maximisation over finite time horizons can result in catastrophic loss.)

There are a lot of decisionmaking phenomena that would require substantial additions to the ergodicity economics framework if they were to be incorporated. Examples include status quo bias, framing effects, nonlinear probability weighting, and rejection of many bets that would seem to maximise a person’s time average growth rate if accepted (or that require an inordinate amount of storytelling to justify it). (Peters and friends have some papers on the application of ergodicity economics to discounting that I’ll deal with in another post.)
On that final point, Peters often mentions that expected utility theory was an attempt to rescue the failure of expected wealth maximisation to capture decision dynamics. One of the benefits of his model is that the need for psychological explanations is removed.
However, an attempt to remove psychology from decision models will leave a lot of behaviour unexplained. There is a fair questions about “what psychology?” is required, and whether this is the psychology of behavioural economics, ecological decision making, resource rationality or something else (see my critical behavioural economics and behavioural science reading list for a flavour of this). But in many situations people do not appear to maximise the growth rate of wealth.
My other posts on loss aversion can be found here:
 Kahneman and Tversky’s debatable loss aversion assumption
 What can we infer about someone who rejects a 50:50 bet to win $110 or lose $100? The Rabin paradox explored
 The case against loss aversion
 Ergodicity economics – a primer
 Ergodicity economics – Do we need risk or loss aversion to explain our failure to accept some gambles? (this post)
Code
Below is the R code used for the simulations described above and generation of the supporting figures.
Load the required packages:
library(ggplot2) library(scales) #use the percent scale later
Create a function for running of the bets.
bet < function(p, pop, periods, start=100, gain, loss, ergodic=FALSE, absorbing=FALSE){ #p is probability of a gain #pop is how many people in the simulation #periods is the number of coin flips simulated for each person #start is the number of dollars each person starts with #if ergodic=FALSE, gain and loss are the multipliers #if ergodic=TRUE, gain and loss are the dollar amounts #if absorbing=TRUE, zero wealth ends the series of flips for that person params < as.data.frame(c(p, pop, periods, start, gain, loss, ergodic, absorbing)) rownames(params) < c("p", "pop", "periods", "start", "gain", "loss", "ergodic", "absorbing") colnames(params) < "value" sim < matrix(data = NA, nrow = periods, ncol = pop) if(ergodic==FALSE){ for (j in 1:pop) { x < start for (i in 1:periods) { outcome < rbinom(n=1, size=1, prob=p) ifelse(outcome==0, x < x*loss, x < x*gain) sim[i,j] < x } } } if(ergodic==TRUE){ for (j in 1:pop) { x < start for (i in 1:periods) { outcome < rbinom(n=1, size=1, prob=p) ifelse(outcome==0, x < xloss, x < x+gain) sim[i,j] < x if(absorbing==TRUE){ if(x<0){ sim[i:periods,j] < 0 break } } } } } sim < rbind(rep(start,pop), sim) #placing the starting sum in the first row sim < cbind(seq(0,periods), sim) #number each period sim < data.frame(sim) colnames(sim) < c("period", paste0("p", 1:pop)) sim < list(params=params, sim=sim) sim }
Simulate 10,000 people who accept a series of 1000 50:50 bets to win $50 or lose $40 from a starting wealth of $100.
set.seed(20200203) ergodic < bet(p=0.5, pop=10000, periods=1000, gain=50, loss=40, ergodic=TRUE, absorbing=FALSE)
Create a function for plotting the path of individuals in the population over a set number of periods.
individualPlot < function(sim, periods, people){ basePlot < ggplot(sim$sim[c(1:(periods+1)),], aes(x=period)) + labs(y = "Wealth ($)") for (i in 1:people) { basePlot < basePlot + geom_line(aes_string(y = sim$sim[c(1:(periods+1)),(i+1)]), color = 2) #need to use aes_string rather than aes to get all lines to print rather than just last line } basePlot }
Plot both the average outcome and first twenty people on the same plot.
jointPlot < function(sim, periods, people) { individualPlot(sim, periods, people) + geom_line(aes(y = rowMeans(sim$sim[c(1:(periods+1)),2:(sim$params[2,]+1)])), color = 1, size=1) } ergodicPlot < jointPlot(sim=ergodic, periods=100, people=20) ergodicPlot
Create a function to generate summary statistics.
summaryStats < function(sim, period=100){ meanWealth < mean(as.matrix(sim$sim[(period+1),2:(sim$params[2,]+1)])) medianWealth < median(as.matrix(sim$sim[(period+1),2:(sim$params[2,]+1)])) num99 < sum(sim$sim[(period+1),2:(sim$params[2,]+1)]<(sim$params[4,]/100)) #number who lost more than 99% of their wealth numGain < sum(sim$sim[(period+1),2:(sim$params[2,]+1)]>sim$params[4,]) #number who gain num100 < sum(sim$sim[(period+1),2:(sim$params[2,]+1)]>(sim$params[4,]*100)) #number who increase their wealth more than 100fold winner < max(sim$sim[(period+1),2:(sim$params[2,]+1)]) #wealth of wealthiest person winnerShare < winner / sum(sim$sim[(period+1),2:(sim$params[2,]+1)]) #wealth share of wealthiest person print(paste0("mean: $", round(meanWealth, 2))) print(paste0("median: $", round(medianWealth, 2))) print(paste0("number who lost more than 99% of their wealth: ", num99)) print(paste0("number who gained: ", numGain)) print(paste0("number who increase their wealth more than 100fold: ", num100)) print(paste0("wealth of wealthiest person: $", round(winner))) print(paste0("wealth share of wealthiest person: ", percent(winnerShare))) }
Generate summary statistics for the population and wealthiest person after 100 and 1000 periods.
summaryStats(sim=ergodic, period=100) summaryStats(sim=ergodic, period=1000)
Determine how many people experienced zero wealth or less during the simulation.
numZero < function(sim, periods=1000){ numZero < sim$params[2,]  sum(sapply(ergodic$sim[1:periods,2:(ergodic$params[2,]+1)], function(x) all(x>0))) numZero } numZero(sim=ergodic, periods=1000) numZero(sim=ergodic, periods=100)
Determine the mimimum wealth experienced by any person.
minWealth < function(sim, periods=1000){ minWealth < min(sim$sim[1:periods,2:(ergodic$params[2,]+1)]) minWealth } minWealth(ergodic, 1000)
Simulate the passive paths for the Copenhagen experiment.
passiveSim < function(type="additive", people=10000, start=1000){ #parameters used in Copenhagen experiment add < c(428, 321, 214, 107, 0, 107, 214, 321, 428) mult < c(0.447, 0.546, 0.668, 0.818, 1, 1.223, 1.496, 1.830, 2.239) add333 < rep(add,37) mult333 < rep(mult, 37) gamblePath < cbind(rep(start, people), matrix(data = NA, nrow = people, ncol = 333)) if(type=="additive"){ for (i in 1:people){ gamble < sample(add333, size=333, replace=FALSE) for (j in 1:333){ gamblePath[i, j+1] < gamblePath[i, j]+gamble[j] } } } if(type=="multiplicative"){ for (i in 1:people){ gamble < sample(mult333, size=333, replace=FALSE) for (j in 1:333){ gamblePath[i, j+1] < gamblePath[i, j]*gamble[j] } } } gamblePath } addSim < passiveSim(type="additive") multSim < passiveSim(type="multiplicative")
Examine how many simulated paths conform to the required range.
#function to output number below lower limit, above upper limit, and within the range of the two numRange < function(sim, lower=0, upper=5000, people=10000){ low < people  sum(apply(sim, 1, function(x) all(x>lower))) up < people  sum(apply(sim, 1, function(x) all(x<upper))) range < sum(apply(sim, 1, function(x) all(x>lower & x<upper))) print(low) print(up) print(range) } numRange(addSim) numRange(multSim)
Simulate the payments to participants based on their actual choices.
library("tidyverse") set.seed(20200321) #Import the data on the choices made for (i in 1:19){ importData < read.csv(paste0("https://raw.githubusercontent.com/olliehulme/ergodicitybreakingchoiceexperiment/master/data/TxtFiles_additive/", i, "_2.txt"), sep="")[1:312,] #limit to 312 entries as subject 3 has 314 importData < select(importData, earnings, KP_Final, Gam1_1, Gam1_2, Gam2_1, Gam2_2) assign(paste0("subject_data_", i), importData) } payment < data.frame(matrix(NA, nrow=10, ncol=18)) #Simulate 1000 payments for each participant for (i in c(1:19)){ for (j in 1:1000){ subject_data < get(paste0("subject_data_", i)) subject_data % mutate(Gam1 = case_when(KP_Final==9 ~ Gam1_1, KP_Final==8 ~ Gam2_1)) %>% mutate(Gam2 = case_when(KP_Final==9 ~ Gam1_2, KP_Final==8 ~ Gam2_2)) %>% mutate(result = mapply(function(x,y){sample(c(x,y),1)}, x=Gam1, y=Gam2)) #Payment is the initial endowment from the passive phase plus a draw of 10 gambles payment[j,i] < subject_data$earnings[1] + sum(sample(subject_data$result, 10)) #starting money plus random draw of 10 } } colnames(payment) < c(1:19) #remove subject 5 from analysis as excluded in paper payment % select(5) #Determine how many participants made a loss sum(payment>0, na.rm=TRUE) sum(payment<0, na.rm=TRUE) sum(payment<(1000)) sum(is.na(payment))
Hi Jason,
A couple of comments:
(1) “However, I am not convinced this critique applies to the experiment by Hulme and friends. The experimental subjects never got to observe the changes in their wealth during the active session (although they might weakly infer the likely direction based on the favourability of the bets they had been exposed to). As a result, Iâ€™m not convinced that you would see the change in risk aversion observed in the experiment under an expected utility framework.”
I guess you didn’t buy the argument I used my response to your comment on my Medium post, so let me try a slightly different one.
We know that CE (Copenhagen Experiment) assumes test subjects behave in the active phase as if they’re perpetually in the initial state with known initial wealth (that’s how Meder et al. estimate the risk aversion parameter). We also know that for a true singleperiod gamble there’s no difference between additive and multiplicative dynamics (in my paper I use the example of a gamble that can be viewed as either +1/1 or +25%/25%). So, here’s my question: why do the test subjects behave differently in this perpetualinitialstate scenario, where clearly behavior changes with learned dynamics, than they would in a true singleperiod gamble where’s there’s no difference between additive and multiplicative?
The answer is clear: test subject learn the rules of the game by observing the wealth dynamics in a scenario where they are aware of how gambles affect wealth changes. Once they’ve learned the game rules during the passive phase, they infer an optimal strategy and then behave using this same strategy during the active phase. My paper shows that, in a simplified version of CE, this optimal strategy is to have zero risk aversion in the initial state, which is exactly how they behave. They continue to have zero risk aversion as the active phase progresses because they continue to behave as though they’re perpetually in the initial state (same assumption Meder et al. use)
(2) “Goldstein also raised a second possible driver of the reduced risk aversion in the additive scenario. Experimental subjects were paid on the based on a random draw of 10 of their active gambles. If the final wealth for an experimental subject from those 10 gambles was negative, they would be given a new draw of 10 gambles. In effect, they were protected from the most severe risks in the additive case, which would reduce risk aversion.
One possible mitigant of this effect is that the experimental subjects were not explicitly told they could not lose (although they would likely have inferred they could not suffer loss). I am also not convinced that the strength of this effect would be enough to result in purely linear utility as was observed, but it should be accounted for.”
This wasn’t in my researchers.one paper, but in my second Medium piece I pointed out that the “implied put option” given to the test subjects was in fact much more valuable than you describe. They were guaranteed a 1,000 DKK payment at the end of the 2 days even if they lost every gamble! They were given 1,000 DKK of gambling money at the start of each day, so initial wealth “within the game” can be viewed as 2,000 DKK and they were guaranteed a 1,000 DKK payout at the end. This “put option” with a 1,000 DKK strike price is very valuable compared to the 1,000 DKK of gambling money they were given at the start of each day. It has a huge effect on observed risk aversion using additive dynamics and, importantly, no effect at all on multiplicative dynamics.
I’ve done some more analysis on the value of this implied put option and will write it up shortly. Suffice it to say for now that this put option has a dramatic effect on CE. Even if the minimum payout was much smaller, say 100 DKK, even that would have been enough to cause zero risk aversion with additive dynamics.
Regards,
Adam
Hi Adam, thanks for the response.
On 1), my question mark doesn’t change with that different framing. The optimal strategy in your paper requires knowledge of the state at each point in time so that they can change from accept/reject depending on the path. Assuming you’re always at the initial state as you describe above doesn’t give you the information to execute that strategy.
On 2), I accidentally omitted the link to the Medium post – now added. But on that, the 1000DKK guaranteed is irrelevant except to the extent that it changes their total wealth (not that Meder et al considered total wealth). They received it regardless of the outcome – the “exercise of the option” does not affect that sum. On the analysis in the Medium post, that again (if I understand correctly) is based on a simulation where the agent has the ability to change bets depending on the path to that point. What do they do when they don’t know the path at that moment and can’t use dynamic programming to derive the optimal choice at each step?
I’m pulling together some simulations on these two points. I’ll put them up when I’ve got my head around them, but in effect an expected utility maximiser in your scenario but without feedback on outcomes would simply consider ll the possible outcomes (at the final step), their probabilities, and calculate expected utility based on that. A floor as in the CE would lift the EU (and affect the measure of risk aversion), but by how much? I’m playing with that now.
Hi Jason,
Actually the analysis is much simpler if we assume no knowledge of each gamble outcome and we’re always in the initial state, and it behaves much like my dynamic programmingbased analysis showed. Here’s a simple example:
Let’s say initial wealth w0=3, minimum payout (floor)=1 and we’re deciding between two gambles: (a) +/ 1 with win probability p=0.6, and (b) +/ 2 with same p=0.6. Let’s assume the agent maximizes log utility of final wealth. If he’s offered a single gamble, he’ll choose (a) because 0.6log(4) + 0.4log(2) > 0.6log(5) + 0.4log(1).
But what if he’s offered the choice between (a) and (b) and the gamble is repeated N times with the outcomes added together? For large enough N and log utility of final wealth with a floor, the agent will always choose the gamble with higher expected value (EV). (Note: a more general version of this statement is proven in this paper https://www.sciencedirect.com/science/article/pii/S0022053185710769). Since only EV determines the gamble choice and nothing else, the agent displays zero risk aversion in the limit of large N.
Note that in this formulation the agent decides in the initial state which gamble he’s going to choose, knowing that he’s going to repeat that decision N times.
In the example I gave above, once N is at least 2 the agent chooses (b). For N=2, 0.36log(7) + 0.48log(3) + 0.16log(1) > 0.36log(5) + 0.48log(3) + 0.16log(1), so agent chooses (b).
A good paper to read on this topic is “Adding Risks” by Ross: https://www.researchgate.net/publication/227406273_Adding_Risks_Samuelson's_Fallacy_of_Large_Numbers_Revisited
Regards,
Adam
I am aware of that argument – it’s effectively the scenario I’m testing in the simulation. It seems a different argument, however, to that in your paper.
That said, for the CE N=10 and the floor is some distance from the starting wealth. I’m not convinced that the small sample and the relative size of the gambles are sufficient to drive risk aversion to zero – but I guess those simulations will expose whether my instinct is right or not.
Thinking out loud – another interesting test would be to take the discrepant trials from the CE that should generate different decisions for a growthrate maximiser and EU maximiser, and consider whether they would result in different decisions if we think of each as being repeated 10 times with a floor of a loss of 1000DKK. A crude approximation, but would be revealing…..
>The bet has a positive expected value
My math says it doesn’t have a positive expected value
what is happenening
On Tue, Feb 18, 2020 at 2:01 AM Jason Collins blog wrote:
> Jason Collins posted: “In a previous post I posed the following bet: > Suppose you have $100 and are offered a gamble involving a series of coin > flips. For each flip, heads will increase your wealth by 50%. Tails will > decrease it by 40%. Flip 100 times. The changes in wealt” >
Every flip: 0.50.5+0.5(0.4)=0.05
Website wouldn’t let me reply to your latest comment, so I’ll put it here.
Yes, I agree this is a different (but very related) argument compared to what’s in my paper. Ollie has told me they plan to reveal gamble outcomes and update wealth immediately in new versions of the experiment, so my paper applies more directly to that future scenario. It also directly addresses Peters’ claims that “EUT predicts dynamics have no effect on gamble choices”, which is clearly incorrect.
Within the game, starting wealth was 2,020 DKK and minimum wealth at the end is 1,020 DKK. The largest gamble outcome is 428 DKK, so it doesn’t seem too hard to hit the minimum. I guess you’re right about N=10 though, I was thinking in terms of total trials in the experiment, which was around 300. Keep in mind, EE only makes predictions in the limit of very long time horizon. So if N=10 isn’t large enough to show the effect I’m discussing for EUT, I’d argue it’s also not large enough for EE to be valid. Also: their experiment didn’t show risk aversion falls to literally zero, just that switching to additive dynamics causes measured eta to become very small.
You mentioned realworld wealth in your post, I’ve discussed that with Ollie as well. I think that has a significant effect on additive dynamics but no effect whatsoever on multiplicative. If realworld wealth is W0, it raises initial wealth to W0+2,020 and minimum final wealth to W0+1,020 DKK for the additive dynamics case.
For multiplicative dynamics, only the initial 1,020 DKK of gambling money is multiplied by the gamble outcomes, so the minimum value is never hit and the “put option” is irrelevant. In this case, maximizing utility “within the game” is equivalent to maximizing realworld final wealth
Probably hit the thread limit…
Agree on those early points, particularly around the claim of dynamics having no effect on gamble choices.
On the effect of real world wealth, here’s my concern. If you’re maximising expected log utility in the multiplicative world, you should be more likely to accept a particular bet the higher your external wealth. Once you’re at large wealth, the within experiment behaviour should look pretty much like expected value maximisation.
This point is the same for both EUT and timeoptimal – so it doesn’t affect the ability to differentiate. However, a person with any outside wealth would have a calculated risk aversion coefficient (much) less than one if it was calculated ignoring that outside wealth (as was done in the experiment). That this wasn’t the case in the experiment – the coefficient was around one – is a challenge to both EUT and growthrate maximisation and suggests some form of mental accounting / localised maximisation is going on.
Yes, I agree. As I thought more about my comment on realworld wealth having no effect on multiplicative dynamics, something didn’t seem right so I just took another look at it. Clearly it should affect measured risk aversion in the EUT model. Something related is puzzling me: multiplicative dynamics no longer seems to result in a “myopic” multiperiod strategy when there’s realworld wealth.
I’ll have to think about this stuff some more, but why do you say this point is the same for EUT and timeoptimal (EE)? Seems to me like this corroborates EE but is a problem for EUT. They’re saying log(.) is just a nonlinear transformation to make the wealth changes ergodic, it’s not really a risk aversion parameter. So the required ergodic transformation should be independent of realworld wealth according to their theory, right?
“So the required ergodic transformation should be independent of realworld wealth according to their theory, right?”
I also need to think about that some more – but that seems right to me. That has some interesting consequences. Suppose one person maximises the growth rate of realworld wealth. The other maximises the growthrate within the experiment. Both are consistent with EE, but would make different decisions. How do you predict someone’s decisions without adding additional assumptions about the nature of the transformation?
Jason … can we apply this more broadly to postinternet economics where everyone should be a winner but in fact a small proportion are doing very well and the rest of the population are in gradual decline, ahead of a big bust where everyone loses?
If you take the model to its limits, its a possibility. But if those people at the top are able to (and do) bet the Kelly optimum amount or less, that big bust for the winners wonâ€™t necessarily come.
I was probably being more philosophical than scientific – the two scenarios just bear remarkable similarities and we need to find a solution for the reallife wealth disparity one! Unless a virus kills us all first.