Portfolio optimization is an important topic in Finance. Modern portfolio theory (MPT) states that investors are risk averse and given a level of risk, they will choose the portfolios that offer the most return. To do that we need to optimize the portfolios.

To perform the optimization we will need

  • To download the price data of the assets
  • Calculate the mean returns for the time period
  • Assign random weights to the assets and then use those to build an efficient frontier

So lets begin

Downloading data

First lets load our packages

library(tidyquant) # To download the data
library(plotly) # To create interactive charts
library(timetk) # To manipulate the data series

Next lets select a few stocks to build our portfolios.

We will choose the following 5 stocks

  • Apple Inc (AAPL)
  • Amazon (AMZN)
  • Netflix (NFLX)
  • Exxon Mobil (XOM)
  • AT&T (T)

Lets download the price data.

tick <- c('AMZN', 'AAPL', 'NFLX', 'XOM', 'T')

price_data <- tq_get(tick,
                     from = '2014-01-01',
                     to = '2018-05-31',
                     get = 'stock.prices')

Next we will calculate the daily returns for these stocks. We will use the logarithmic returns.

log_ret_tidy <- price_data %>%
  group_by(symbol) %>%
  tq_transmute(select = adjusted,
               mutate_fun = periodReturn,
               period = 'daily',
               col_rename = 'ret',
               type = 'log')

Lets look at the first few rows.

head(log_ret_tidy)
## # A tibble: 6 x 3
## # Groups:   symbol [1]
##   symbol date            ret
##   <chr>  <date>        <dbl>
## 1 AMZN   2014-01-02  0      
## 2 AMZN   2014-01-03 -0.00385
## 3 AMZN   2014-01-06 -0.00711
## 4 AMZN   2014-01-07  0.0111 
## 5 AMZN   2014-01-08  0.00973
## 6 AMZN   2014-01-09 -0.00227

As you can see that this data is in tidy format. We will use the spread() function to convert it to a wide format. And we will also convert it into a time series object using xts() function.

log_ret_xts <- log_ret_tidy %>%
  spread(symbol, value = ret) %>%
  tk_xts()
## Warning in tk_xts_.data.frame(data = data, select = select, date_var =
## date_var, : Non-numeric columns being dropped: date
## Using column `date` for date_var.
head(log_ret_xts)
##                    AAPL         AMZN          NFLX             T
## 2014-01-02  0.000000000  0.000000000  0.0000000000  0.0000000000
## 2014-01-03 -0.022210680 -0.003851917  0.0007714349 -0.0043010588
## 2014-01-06  0.005438071 -0.007113316 -0.0097694303  0.0045870163
## 2014-01-07 -0.007177294  0.011115982 -0.0574349094 -0.0002859575
## 2014-01-08  0.006313210  0.009725719  0.0043791809 -0.0072749741
## 2014-01-09 -0.012852410 -0.002266707 -0.0116217990 -0.0206557460
##                     XOM
## 2014-01-02  0.000000000
## 2014-01-03 -0.002409058
## 2014-01-06  0.001506219
## 2014-01-07  0.014048946
## 2014-01-08 -0.003270542
## 2014-01-09 -0.009775460

This is better for our purpose.

Next lets calculate the mean daily returns for each asset.

mean_ret <- colMeans(log_ret_xts)
print(round(mean_ret, 5))
##     AAPL     AMZN     NFLX        T      XOM 
##  0.00085  0.00127  0.00173  0.00015 -0.00004

Next we will calculate the covariance matrix for all these stocks. We will annualize it by multiplying by 252.

cov_mat <- cov(log_ret_xts) * 252

print(round(cov_mat,4))
##        AAPL   AMZN   NFLX      T    XOM
## AAPL 0.0523 0.0238 0.0268 0.0089 0.0127
## AMZN 0.0238 0.0869 0.0489 0.0078 0.0129
## NFLX 0.0268 0.0489 0.1759 0.0081 0.0133
## T    0.0089 0.0078 0.0081 0.0260 0.0110
## XOM  0.0127 0.0129 0.0133 0.0110 0.0340

Before we apply our methods to thousands of random portfolio, let us demonstrate the steps on a single portfolio.

To calculate the portfolio returns and risk (standard deviation) we will us need

  • Mean assets returns
  • Portfolio weights
  • Covariance matrix of all assets
  • Random weights

Lets create random weights first.

wts <- runif(n = length(tick))
print(wts)
## [1] 0.8147560 0.5657122 0.8132951 0.1917444 0.8166054
print(sum(wts))
## [1] 3.202113

We created some random weights, but the problem is that their sum is more than 1. We can fix this as shown below.

wts <- wts/sum(wts)
print(wts)
## [1] 0.25444323 0.17666840 0.25398702 0.05988058 0.25502078
sum(wts)
## [1] 1

Next we will calculate the annualized portfolio returns.

port_returns <- (sum(wts * mean_ret) + 1)^252 - 1

Next we will calculate the portfolio risk (Standard deviation). This will be annualized Standard deviation for the portfolio. We will use linear algebra to calculate our portfolio risk.

port_risk <- sqrt(t(wts) %*% (cov_mat %*% wts))
print(port_risk)
##           [,1]
## [1,] 0.1878822

Next we will assume 0% risk free rate to calculate the Sharpe Ratio.

# Since Risk free rate is 0% 

sharpe_ratio <- port_returns/port_risk
print(sharpe_ratio)
##          [,1]
## [1,] 1.317836

Lets put all these steps together.

# Calculate the random weights
wts <- runif(n = length(tick))
wts <- wts/sum(wts)

# Calculate the portfolio returns
port_returns <- (sum(wts * mean_ret) + 1)^252 - 1

# Calculate the portfolio risk
port_risk <- sqrt(t(wts) %*% (cov_mat %*% wts))

# Calculate the Sharpe Ratio
sharpe_ratio <- port_returns/port_risk

print(wts)
## [1] 0.20421152 0.15279108 0.25758987 0.29417104 0.09123649
print(port_returns)
## [1] 0.2396506
print(port_risk)
##           [,1]
## [1,] 0.1778445
print(sharpe_ratio)
##          [,1]
## [1,] 1.347529

We have everything we need to perform our optimization. All we need now is to run this code on 5000 random portfolios. For that we will use a for loop.

Before we do that, we need to create empty vectors and matrix for storing our values.

num_port <- 5000

# Creating a matrix to store the weights

all_wts <- matrix(nrow = num_port,
                  ncol = length(tick))

# Creating an empty vector to store
# Portfolio returns

port_returns <- vector('numeric', length = num_port)

# Creating an empty vector to store
# Portfolio Standard deviation

port_risk <- vector('numeric', length = num_port)

# Creating an empty vector to store
# Portfolio Sharpe Ratio

sharpe_ratio <- vector('numeric', length = num_port)

Next lets run the for loop 5000 times.

for (i in seq_along(port_returns)) {
  
  wts <- runif(length(tick))
  wts <- wts/sum(wts)
  
  # Storing weight in the matrix
  all_wts[i,] <- wts
  
  # Portfolio returns
  
  port_ret <- sum(wts * mean_ret)
  port_ret <- ((port_ret + 1)^252) - 1
  
  # Storing Portfolio Returns values
  port_returns[i] <- port_ret
  
  
  # Creating and storing portfolio risk
  port_sd <- sqrt(t(wts) %*% (cov_mat  %*% wts))
  port_risk[i] <- port_sd
  
  # Creating and storing Portfolio Sharpe Ratios
  # Assuming 0% Risk free rate
  
  sr <- port_ret/port_sd
  sharpe_ratio[i] <- sr
  
}

All the heavy lifting has been done and now we can create a data table to store all the values together.

# Storing the values in the table
portfolio_values <- tibble(Return = port_returns,
                  Risk = port_risk,
                  SharpeRatio = sharpe_ratio)


# Converting matrix to a tibble and changing column names
all_wts <- tk_tbl(all_wts)
## Warning in tk_tbl.data.frame(as.data.frame(data), preserve_index,
## rename_index, : Warning: No index to preserve. Object otherwise converted
## to tibble successfully.
colnames(all_wts) <- colnames(log_ret_xts)

# Combing all the values together
portfolio_values <- tk_tbl(cbind(all_wts, portfolio_values))
## Warning in tk_tbl.data.frame(cbind(all_wts, portfolio_values)): Warning: No
## index to preserve. Object otherwise converted to tibble successfully.

Lets look at the first few values.

head(portfolio_values)
## # A tibble: 6 x 8
##    AAPL   AMZN  NFLX     T      XOM Return  Risk SharpeRatio
##   <dbl>  <dbl> <dbl> <dbl>    <dbl>  <dbl> <dbl>       <dbl>
## 1 0.302 0.160  0.206 0.130 0.203     0.231 0.175       1.32 
## 2 0.192 0.0303 0.140 0.362 0.276     0.130 0.143       0.906
## 3 0.196 0.227  0.158 0.235 0.184     0.209 0.163       1.28 
## 4 0.194 0.214  0.142 0.304 0.146     0.199 0.158       1.26 
## 5 0.256 0.106  0.197 0.195 0.245     0.196 0.164       1.20 
## 6 0.328 0.320  0.178 0.173 0.000823  0.293 0.192       1.52

We have the weights in each asset with the risk and returns along with the Sharpe ratio of each portfolio.

Next lets look at the portfolios that matter the most.

  • The minimum variance portfolio
  • The tangency portfolio (the portfolio with highest sharpe ratio)
min_var <- portfolio_values[which.min(portfolio_values$Risk),]
max_sr <- portfolio_values[which.max(portfolio_values$SharpeRatio),]

Lets plot the weights of each portfolio. First with the minimum variance portfolio.

p <- min_var %>%
  gather(AAPL:XOM, key = Asset,
         value = Weights) %>%
  mutate(Asset = as.factor(Asset)) %>%
  ggplot(aes(x = fct_reorder(Asset,Weights), y = Weights, fill = Asset)) +
  geom_bar(stat = 'identity') +
  theme_minimal() +
  labs(x = 'Assets', y = 'Weights', title = "Minimum Variance Portfolio Weights") +
  scale_y_continuous(labels = scales::percent) 

ggplotly(p)

As we can observe the Minimum variance portfolio has no allocation to Netflix and very little allocation to Amazon. The majority of the portfolio is invested in Exxon Mobil and AT&T stock.

Next lets look at the tangency portfolio or the the portfolio with the highest sharpe ratio.

p <- max_sr %>%
  gather(AAPL:XOM, key = Asset,
         value = Weights) %>%
  mutate(Asset = as.factor(Asset)) %>%
  ggplot(aes(x = fct_reorder(Asset,Weights), y = Weights, fill = Asset)) +
  geom_bar(stat = 'identity') +
  theme_minimal() +
  labs(x = 'Assets', y = 'Weights', title = "Tangency Portfolio Weights") +
  scale_y_continuous(labels = scales::percent) 

ggplotly(p)

Not surprisingly, the portfolio with the highest sharpe ratio has very little invested in Exxon Mobil and AT&T. This portfolio has most of the assets invested in Amazon, Netflix and Apple. Three best performing stocks in the last decade.

Finally lets plot all the random portfolios and visualize the efficient frontier.

p <- portfolio_values %>%
  ggplot(aes(x = Risk, y = Return, color = SharpeRatio)) +
  geom_point() +
  theme_classic() +
  scale_y_continuous(labels = scales::percent) +
  scale_x_continuous(labels = scales::percent) +
  labs(x = 'Annualized Risk',
       y = 'Annualized Returns',
       title = "Portfolio Optimization & Efficient Frontier") +
  geom_point(aes(x = Risk,
                 y = Return), data = min_var, color = 'red') +
  geom_point(aes(x = Risk,
                 y = Return), data = max_sr, color = 'red') +
  annotate('text', x = 0.20, y = 0.42, label = "Tangency Portfolio") +
  annotate('text', x = 0.18, y = 0.01, label = "Minimum variance portfolio") +
  annotate(geom = 'segment', x = 0.14, xend = 0.135,  y = 0.01, 
           yend = 0.06, color = 'red', arrow = arrow(type = "open")) +
  annotate(geom = 'segment', x = 0.22, xend = 0.2275,  y = 0.405, 
           yend = 0.365, color = 'red', arrow = arrow(type = "open"))
  

ggplotly(p)

In the chart above we can observe all the 5000 portfolios. As mentioned above, a risk averse investor will demand a highest return for a given level of risk. In other words he/she will try to obtain portfolios that lie on the efficient frontier.