Traditional portfolio optimization (often called modern portfolio theory, or mean-variance optimization) balances expected portfolio return with expected portfolio variance. You input how opposed you are to portfolio variance (your risk tolerance), then you build a portfolio that gives you the best return given your risk tolerance.
Goals-based investing, by contrast, defines “risk” as the probability of not achieving your goal. We assume you want to minimize that risk, so that is the only portfolio objective of a goals-based investor. As you will see, variance and returns are inputs into that equation, but they are not the equation (I rant more about that here).
For the most part, goals-based optimization produces portfolios that are on the mean-variance efficient frontier, but not always. I won’t go into why that is in this post (here is why), but I will demonstrate when the two methods diverge.
Fundamentally, there are three basic steps to optimizing a goal-based portfolio:
- Determine your goal variables: time horizon, amount of wealth dedicated to the goal today, and future required wealth value.
- Develop capital market expectations for your investment universe: correlations, return expectations, and volatility.
- Run a standard optimizer with a goal-based utility function.
This post is all about how to optimize a goal based portfolio in R.
First, we need to understand the goal, what is it you want to do with the money? To keep things simple, let’s say you need $1,000 in 10 years, and you have $750 dedicated to it today. We organize this into a goal vector with the goal’s value, the required funding value ($1,000), and the time horizon (10 years). Note that the goal’s value is only relevant when optimizing your current wealth across your goals, which this post does not cover.
pool <- 750 # Total amount dedicated to this goal
goal_vector <- c(1, 1000, 10) # c(Goal value, goal funding requirement, time horizon)
Ok. Step 1 is done.
Second, we need to develop capital market expectations for our investment universe. This topic is so big numerous books have been written on it because it is a very important step. Better forecasts yield better results. Since this post isn’t about building CMEs, let’s just input something simple.
- Stocks: 9% average return with 15% volatility
- Bonds: 4.5% average return with 5% volatility
- Gamble: -1% average return with 80% volatility
- Cash: 0.5% average return with 0.01% volatility
# Asset names
assets <- c('Stocks', 'Bonds', 'Gamble', 'Cash')
# Capital Market expectations - each is a vector in order of assets
cme <- data.frame( 'Return_Forecast' = c(0.09, 0.045, -0.01, 0.005),
'Volatility Forecast' = c(0.15, 0.05, 0.80, 0.001) )
# Correlation matrix, in order of assets both sideways and vertically
correlations <- matrix( c(1.00, 0.10, 0.00, 0.00,
0.10, 1.00, 0.00, 0.00,
0.00, 0.00, 1.00, 0.00,
0.00, 0.00, 0.00, 1.00),
nrow = length(assets), ncol = length(assets), byrow=T )
Note the “gamble” asset–we are going to have fun with that in a minute! Step 2 is complete.
Finally, now that we have our human-based inputs, let’s proceed with the algorithm. Load our required libraries.
library(tidyverse)
library(Rsolnp) # this is the optimizer solnp()
And build the functions we will use.
# Required Functions
# This function converts the covariance table and weight vector into a
# portfolio standard deviation.
sd.f = function(weight_vector, covar_table){
covar_vector = 0
for(z in 1:length(weight_vector)){
covar_vector[z] = sum(weight_vector * covar_table[,z])
}
return( sqrt( sum( weight_vector * covar_vector) ) )
}
# This function will return the expected portfolio return, given the
# forecasted returns and proposed portfolio weights
mean.f = function(weight_vector, return_vector){
return( sum( weight_vector * return_vector ) )
}
# This function will return the probability of goal achievement, given
# the goal variables, allocation to the goal, expected return of the
# portfolio, and expected volatiltiy of the portfolio
phi.f = function(goal_vector, goal_allocation, pool, mean, sd){
required_return = (goal_vector[2]/(pool * goal_allocation))^(1/goal_vector[3]) - 1
if( goal_allocation * pool >= goal_vector[2]){
return(1)
} else {
return( 1 - pnorm( required_return, mean, sd, lower.tail=TRUE ) )
}
}
# For use in the optimization function later, this is failure probability,
# which we want to minimize.
optim_function = function(weights){
1 - phi.f(goal_vector, allocation, pool,
mean.f(weights, return_vector),
sd.f(weights, covar_table) )
}
# For use in the optimization function later, this allows the portfolio
# weights to sum to 1.
constraint_function = function(weights){
sum(weights)
}
Since we input correlations and volatilities, we need to build a covariance table. This uses the
form (covariance of asset i to j equals the correlation of i and j times the volatility of i times the volatility of j).
# Convert correlations to covariances
covariances <- matrix( nrow = length(assets), ncol = length(assets) )
for(i in 1:length(assets)){
for(j in 1:length(assets)){
covariances[j,i] <- cme[i,2] * cme[j,2] * correlations[i,j]
}
}
All that is left is to do the optimization
# Optimization
return_vector <- cme$Return_Forecast
covar_table <- covariances
allocation <- 1
starting_weights <- rep(0.25, length(assets)) # start with 25% weights
result <- solnp( starting_weights, # Initialize weights
optim_function, # The function to minimize
eqfun = constraint_function, # The constraint function ( sum(weights) )
eqB = 1, # Constraint function must equal 1
LB = rep(0, length(assets)), # Lower bound of constraint, weight >= 0
UB = rep(1, length(assets)) ) # Upper bound of constraint, weight <= 1
# Results
optimal_weights <- data.frame( 'Assets' = assets,
'Optimal Weights' = round(result$pars, 2) )
The solnp function from the Rsolnp package is quite powerful. Plus, when you are running it on complicated problems, the output makes me feel like a hacker, which is always a plus!
And we find that our optimal weights are 31% stocks, 69% bonds.
So What’s Different About Goals-Based Investing?
Now that you’ve got the basics of goals-based portfolio optimization, we may ask what is so different about GBI? Well, let’s find out!
To illustrate, let’s build allocations for various levels of starting wealth.
pool_seq <- seq(50, 1000, 50)
And empty lists to hold the various allocation results
# Empty lists of allocation to hold results
stock_allocation <- 0
bond_allocation <- 0
gamble_allocation <- 0
cash_allocation <- 0
Then iterate through each starting wealth value and determine the optimal investment allocation (this code assumes you’ve run the code in the previous section).
# Loop through each level of starting wealth to determine optimal allocation
for(i in 1:length(pool_seq)){
pool <- pool_seq[i]
result <- solnp( starting_weights,
optim_function,
eqfun = constraint_function,
eqB = 1,
LB = rep(0, length(assets)),
UB = rep(1, length(assets)) )
# Store results for this iteration
stock_allocation[i] <- result$pars[1] %>% round(digits = 2)
bond_allocation[i] <- result$pars[2] %>% round(digits = 2)
gamble_allocation[i] <- result$pars[3] %>% round(digits = 2)
cash_allocation[i] <- result$pars[4] %>% round(digits = 2)
}
Since I am using ggplot, I’ll now need to store the results in a long form data frame.
seq_results <- data.frame( 'Weight' = c(stock_allocation,
bond_allocation,
gamble_allocation,
cash_allocation ),
'Asset' = c( rep('Stock', length(stock_allocation)),
rep('Bond', length(bond_allocation)),
rep('Gamble', length(gamble_allocation)),
rep('Cash', length(cash_allocation))),
'Wealth' = rep(pool_seq, length(assets)) )
And, finally, visualize the results.
ggplot( seq_results, aes(x = Wealth, y = Weight, fill = Asset))+
geom_area( linetype=1, size=0.5, color='black')+
xlab('Initial Wealth')+
ylab('Investment Weight')+
labs(fill = 'Asset',
title = 'Optimal Allocation for Various Levels of Starting Wealth',
subtitle = 'Goals-based optimization')
As you can see, goals-based portfolios will lean on high-variance, low return investments when your starting wealth is small enough. Technically speaking, GBI portfolios begin allocating to lottery-like investments whenever the return required to hit the goal is greater than the return offered by the mean-variance efficient frontier. In traditional mean-variance optimization, the optimizer will maintain exposure to the endpoint of the frontier, or 100% stock allocation, in our example.
As a comparison, here is the mean-variance optimizer result. As you can see, the “gamble” asset is eliminated from consideration.
Because of this, goal-based portfolios yield higher probabilities of goal achievement than mean-variance portfolios (mean-variance portfolios are stochastically dominated by goals-based portfolios). This was Exhibit 5 in my recent paper on this subject.
For all of these reasons (and more), if you have goals to achieve then you should be using goals-based portfolio theory. I hope this post helped you understand how to implement the basic framework!
Goals-based portfolio theory is also concerned with how you allocate your wealth across your goals, which I have not covered in this post, but I will cover that in a later discussion.
Keywords: goals-based investing, goal based investing, investing, portfolio theory, portfolio optimization.
Franklin J. Parker, CFA 
I cut an pasted your code and got the following solution which disagrees with the solution in you post
solnp–> Completed in 2 iterations
optimal weights Assets Optimal.Weights
1 Stocks 0.31
2 Bonds 0.69
3 Gamble 0.00
4 Cash 0.00
LikeLike
That is odd!
I did what you did and copy/pasted the code into R and got the 0.31/0.69.
I must have changed something in my editing of the post.
At any rate, good catch and thanks for letting me know! I’ve edited the post to reflect the change.
LikeLike
I have some more comments about your post.
First, I found that your use of ‘volatility’ was ambiguous. Variances of standard deviations?
Second, I think that ‘allocation’ and ‘pool’ unnecessarily complicate the formulation. I know there may have been a reason you have them, but using “initial portfolio value’ and ‘target portfolio value’ (or equivalent) would be clearer.
Third, your formulation just checks the the probability of achieving the required return at one period of time. Although the mean return stays constant over each period, the portfolio variance grows proportional to time. Why not find the mean return and the variance at the end of the (say) 10 periods, and find the probability that the final portfolio value is less than the target?
LikeLike
Yes, the verbiage is a little unclear. That is largely because it is out of context of the broader theory and practice of the thing. But I’ll agree that adjusting them for a stand-alone post would be much clearer!
Typically, volatility in finance refers to standard deviation, but often variance and standard deviation are used interchangeably.
Now, you make an excellent point about inter-temporal optimization. In fact, I had the same observation in one of my papers. Even using something like a Geometric Brownian Motion model would address your point.
The reason it is broken into a “this year” problem is twofold: (1) investment return distributions become less and less normal through time (check a 10-year histogram of a stock index!).
(2) Practically speaking, most investment managers build capital market expectations for the year in front of them. It may be difficult to estimate means and variances 10-years in the future.
Thanks for the interest and input, btw.
LikeLike
Thanks for your comments. I computed the histograms for 1 year and 10 year returns for SPY and VFITX and take your point about non Gaussian distributions. (I would have attached them to this post if i could.) As a MIT PhD in rocket science, but self taught in finance, I would have assumed they would be more Gaussian (central limit theorem and all that), That’s why I should have gotten the CFA too! I also take your point about using annual required return for the annual outlook reason.
We have a monthly Silicon Valley AAII meeting (now Zoomed) of about 30-50 people interested in CIMI (computer investing, mechanical investing). A lot of us do monthly trading in tactical asset allocation strategies. I think your GBI article will be of interest to the group, so I am giving a short talk on October 1 to introduce GBI to the group. I also plan to present what happens when GBI is used monthly to recover from the big draw down of 03/23/2020.
I’ll send the slides to you first for comments if you don’t mind looking at them
LikeLiked by 1 person
Very cool. I’d love to take a look. I’ll email you.
LikeLike
could this model be desidgned to optimize for a DB pension fund portifolio that has initially more surplus, to be invested in 2 asset classes (stocks and bonds), than the the required payout benefit in one year. Although it is intresting to note that ultimately the aggregate benefit paid out by time, t, (mortality=1, everyone is dead) will have surpassed the “initial wealth”.
LikeLike
To be fair, I know enough about actuarial methods to know that I don’t know near enough to comment on the funding nature of pension schemes! My intuition says that, yes, this approach can be married to a pension asset pool, though exactly how, I don’t know. If you figure it out, I’d be happy to co-author a paper with you!
LikeLike