CoCalc Logo Icon
StoreFeaturesDocsShareSupportNewsAboutSign UpSign In

Real-time collaboration for Jupyter Notebooks, Linux Terminals, LaTeX, VS Code, R IDE, and more,
all in one place.

| Download
Views: 3376
Image: ubuntu2004
Tweet
Share
Share
Kernel: Python 3 (system-wide)

Performance portfolio management

Camilo A. Garcia Trillos 2020

In this notebook

We retrieve the same database available to us when choosing efficient portfolios, but now focus on how to manage an existing portfolio to improve its performance using the ideas of risk allocation and Sharpe ratio.



%matplotlib inline
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import scipy.stats as stats
from math import ceil

Reading data

We read again the database we used on the exercise on efficient frontiers. If in doubt of any on the commands below, go back to that notebook.

data = pd.read_csv('~/Data/basket_20171201_10y.csv')
data['date']=pd.to_datetime(data['date'], yearfirst=True)
P = data.pivot(index='date', columns='ticker', values='adj_close')
P = P.resample('W-FRI').last()
P.head()

logR=np.log(P).diff()
logR.drop(logR.index[0],inplace=True) # We drop the first line which produces NAN
logR.head()

Improving performance of a portfolio

Assume that we have a long only portfolio. Assume for example that we have an equally weighted portfolio

π=[19,19,…,19].\mathbf{\pi} = [\frac 1 9, \frac 1 9 , \ldots, \frac 1 9].

Assume, for the exercise, that we have this portfolio on 2011-12-31. We will use the risk allocation technique and Sharpe ratio to progressively improve the performance of the portfolio.

First, let us calculate the mean and variance of the portfolios using the 9 years prior to the date we have fixed.

logR_sub = logR[logR.index< '2016-12-01']
mu = logR_sub.mean()
Sigma = logR_sub.cov()
Rho = logR_sub.corr()

print('Mean:', mu)
print('Variances', np.diag(Sigma))
print('Var-cov:', Sigma)
print('Correlation', Rho)
Mean: ticker BAC -0.001473 C -0.004167 GE 0.000337 JNJ 0.001638 MSFT 0.001850 PG 0.001107 T 0.001307 WMT 0.001236 XOM 0.000838 dtype: float64 Variances [0.00622047 0.00899841 0.00185316 0.00045822 0.00133477 0.00052352 0.00078705 0.00068348 0.00085506] Var-cov: ticker BAC C GE JNJ MSFT PG T \ ticker BAC 0.006220 0.006109 0.002279 0.000734 0.000961 0.000650 0.000910 C 0.006109 0.008998 0.002623 0.000790 0.001084 0.000612 0.001161 GE 0.002279 0.002623 0.001853 0.000390 0.000583 0.000399 0.000500 JNJ 0.000734 0.000790 0.000390 0.000458 0.000347 0.000286 0.000308 MSFT 0.000961 0.001084 0.000583 0.000347 0.001335 0.000332 0.000386 PG 0.000650 0.000612 0.000399 0.000286 0.000332 0.000524 0.000320 T 0.000910 0.001161 0.000500 0.000308 0.000386 0.000320 0.000787 WMT 0.000594 0.000557 0.000330 0.000255 0.000344 0.000307 0.000325 XOM 0.000912 0.000951 0.000549 0.000341 0.000493 0.000331 0.000427 ticker WMT XOM ticker BAC 0.000594 0.000912 C 0.000557 0.000951 GE 0.000330 0.000549 JNJ 0.000255 0.000341 MSFT 0.000344 0.000493 PG 0.000307 0.000331 T 0.000325 0.000427 WMT 0.000683 0.000298 XOM 0.000298 0.000855 Correlation ticker BAC C GE JNJ MSFT PG T \ ticker BAC 1.000000 0.816495 0.671209 0.434770 0.333541 0.360150 0.411355 C 0.816495 1.000000 0.642261 0.389076 0.312820 0.281768 0.436317 GE 0.671209 0.642261 1.000000 0.422911 0.370724 0.405148 0.414015 JNJ 0.434770 0.389076 0.422911 1.000000 0.443119 0.584024 0.512466 MSFT 0.333541 0.312820 0.370724 0.443119 1.000000 0.396635 0.376291 PG 0.360150 0.281768 0.405148 0.584024 0.396635 1.000000 0.499200 T 0.411355 0.436317 0.414015 0.512466 0.376291 0.499200 1.000000 WMT 0.288191 0.224767 0.293084 0.455463 0.360466 0.513583 0.443629 XOM 0.395248 0.342841 0.436346 0.545207 0.461577 0.494555 0.520472 ticker WMT XOM ticker BAC 0.288191 0.395248 C 0.224767 0.342841 GE 0.293084 0.436346 JNJ 0.455463 0.545207 MSFT 0.360466 0.461577 PG 0.513583 0.494555 T 0.443629 0.520472 WMT 1.000000 0.389416 XOM 0.389416 1.000000

We can now calculate the Sharpe ratio of each asset and the Sharpe ratio of the whole portfolio. For this exercise, we are going to assume that only risky assets are available. 

sharpe_ea = mu/np.diag(Sigma)**0.5
print('Sharpe ratio for each individual asset',sharpe_ea)
Sharpe ratio for each individual asset ticker BAC -0.018674 C -0.043931 GE 0.007824 JNJ 0.076541 MSFT 0.050625 PG 0.048376 T 0.046578 WMT 0.047262 XOM 0.028652 dtype: float64 
# Define a function that returns the Sharpe ratio of a portfolio on the given assets

def sharpe_portfolio (pi, mmu=mu,msigma=Sigma ):
    
    mu_portfolio = pi @ mu
    var_portfolio = (Sigma@pi)@pi   # This is equivalent to pi^T Sigma pi
    
    return mu_portfolio/var_portfolio**0.5

# Test
np.testing.assert_allclose(sharpe_portfolio(np.array([1,0,0,0,0,0,0,0,0])),sharpe_ea[0])


ref_pi = np.ones(9)/9
print('The Sharpe ratio of the equally weighted portfolio is ',sharpe_portfolio( ref_pi ))
The Sharpe ratio of the equally weighted portfolio is 0.00948921913278367

A naive way to try to improve the performance is to increment the investment in those assets whose Sharpe ratio are larger than the overall portfolio Sharpe ratio, while reducing it on the ones that fall below.

A better way is to make a risk allocation following the Euler per unit allocation, and calculate the relative performance of each asset. Recall that when we use standard deviation to estimate risk, we obtain κi=cov(Ri,Rπ)sd(Rπ)\kappa_i = \frac{\mathrm{cov}(R^i,R^\pi)}{ \mathrm{sd}(R^\pi)}

def risk_allocation(pi, mmu=mu,msigma=Sigma ):
    
    cov_ea = Sigma @ pi   # The covariance of asset i would be e_i^T Sigma pi, where e_i is the vector with one only in the ith position. Here we calculate all simultaneously
    var_portfolio = (Sigma@pi)@pi
    
    return cov_ea/var_portfolio**0.5

With the above function, we can calculate the relative Sharpe (also printing again the individual Sharpe for comparison)

print('Absolute Sharpe')
print(sharpe_ea)

print('Relative Sharpe')
print(mu/risk_allocation(ref_pi))
Absolute Sharpe ticker BAC -0.018674 C -0.043931 GE 0.007824 JNJ 0.076541 MSFT 0.050625 PG 0.048376 T 0.046578 WMT 0.047262 XOM 0.028652 dtype: float64 Relative Sharpe ticker BAC -0.021408 C -0.051269 GE 0.009976 JNJ 0.118024 MSFT 0.088798 PG 0.082878 T 0.071791 WMT 0.094162 XOM 0.045743 dtype: float64 
def update_portfolio(pi, step, epsilon , mmu=mu,msigma=Sigma ):
    '''
    The function returns an improvement on the long-only portfolio 'pi' such that changes are proportional 
    to how different are the overall Sharpe portfolio and the relative Sharpe portfolio calculated with the 
    per unit risk allocation. The function must satisfy that a minimum investment of epsilon must be kept on each 
    asset.
    
    '''
    
    # We calculate the difference between relative Sharpe and Sharpe of the portfolio
    aux_positive = np.clip(mu/risk_allocation(pi) - sharpe_portfolio(pi),0,None)
    aux_negative = np.clip(sharpe_portfolio(pi)-mu/risk_allocation(pi),0,None)

    
    # Determine the entries in the portfolio above the threshold
    
    aux_negative[pi<epsilon+1e-12] = 0    # The additional 1e-12 is to avoid roundup errors in this comparison.
    aux_positive[pi<epsilon+1e-12] = 0
    
    # We cannot modify those entries below the threshold, or those with the same Sharpe as the overall portfolio.  
    # The following if tests if there is nothing else to improve within the constraints
    
    
    if aux_positive.sum() ==0 or aux_negative.sum() == 0 :
        return pi
    
    # Now, we calculate the maximum step we can take without violating the minimal position constraint
    
    f_change = aux_positive/aux_positive.sum() - aux_negative/aux_negative.sum()
    aux_v = (epsilon-pi)/f_change    # When f_change is zero, this generates some infinities, but this does not affect the following operation
    aux_step = aux_v[aux_v>0 ].min()
    
    # Calculate effective change and return
    
    change = f_change*min(step, aux_step)
    return pi+change
    
    
    
    
    

pi_new = update_portfolio(ref_pi,0.3,0.01)
print(pi_new)

new_sharpe = sharpe_portfolio(pi_new)
print(new_sharpe)
ticker BAC 0.059693 C 0.010000 GE 0.111278 JNJ 0.148317 MSFT 0.138298 PG 0.136269 T 0.132468 WMT 0.140137 XOM 0.123539 dtype: float64 0.04343099977491054

Note that there is already a big improvement: the performance increased by an order of magnitude. We can repeat the process to improve even further until we cannot improve any more without violating our constraints.


k=1
while k<1000:
    print('Iteration', k)
    pi_new = update_portfolio(pi_new,0.3,0.01)
    print(pi_new)
    old_sharpe = new_sharpe
    new_sharpe = sharpe_portfolio(pi_new)
    print(new_sharpe)
    if abs(old_sharpe-new_sharpe)<1e-12:
        break;
    
    k+=1
Iteration 1 ticker BAC 0.010000 C 0.010000 GE 0.087606 JNJ 0.175204 MSFT 0.152373 PG 0.147521 T 0.142151 WMT 0.154487 XOM 0.120657 dtype: float64 0.05584163800337831 Iteration 2 ticker BAC 0.010000 C 0.010000 GE 0.010000 JNJ 0.227240 MSFT 0.170425 PG 0.158960 T 0.150907 WMT 0.172125 XOM 0.090344 dtype: float64 0.06322960787509502 Iteration 3 ticker BAC 0.010000 C 0.010000 GE 0.010000 JNJ 0.287353 MSFT 0.181435 PG 0.161034 T 0.149848 WMT 0.180329 XOM 0.010000 dtype: float64 0.06668083717555018 Iteration 4 ticker BAC 0.010000 C 0.010000 GE 0.010000 JNJ 0.517049 MSFT 0.200847 PG 0.047951 T 0.010000 WMT 0.184152 XOM 0.010000 dtype: float64 0.07090709706485519 Iteration 5 ticker BAC 0.010000 C 0.010000 GE 0.010000 JNJ 0.722536 MSFT 0.010000 PG 0.059241 T 0.010000 WMT 0.158223 XOM 0.010000 dtype: float64 0.07109427388931822 Iteration 6 ticker BAC 0.010000 C 0.010000 GE 0.010000 JNJ 0.722536 MSFT 0.010000 PG 0.059241 T 0.010000 WMT 0.158223 XOM 0.010000 dtype: float64 0.07109427388931822

After 6 iterations, we found a portfolio whose Sharpe portfolio cannot be improved: it invest the minimal allowed amount on all assets except JNJ, PG, and WMT. Note that it does not focus on just the asset with the largest Sharpe ratio, although, due to the constraints, the obtained Sharpe is slightly worse than investing in JNJ alone... but if this investment is put together with the constraints on minimal portfolio, we would produce a less performing portfolio.

Indeed,

another_pi = 0.01*np.ones(9) 
another_pi[3] += 1-0.09
print(another_pi.sum())

sharpe_portfolio(another_pi   )
1.0
0.07087780201105662

This is smaller. The interaction between the two assets helps to obtain a more performant portfolio within the constraints.

The above discussion is all 'static', that is, we are doing all changes as if we could go back and modify our portfolio several times. The idea of using such incremental steps is to slowly adjust the portfolio as time passes by. To do so, let us make evolve our portfolio and run again the improvement, but updating the information.

A good tool for this purpose is the exponentially weighted moving average (EWMA). This is a way to smoothly decrease the influence of old observations: the idea is as follows: let λ>0\lambda>0. We update a time average of asset ii by posing mean1i=x1mean^i_1=x_1 and

meant+1i=λmeanti+(1−λ)xt+1imean^i_{t+1} = \lambda mean^i_t + (1-\lambda) x^i_{t+1}

This is a simple and common way of slowly 'forgetting' the information in the distant past. We use this to update our mean and standard deviation estimations. For one element in the variance-covariance matrix, we can use

covt+1i,j=λcovti,j+(1−λ)(xt+1i−meant+1i)(xt+1j−meant+1j)cov^{i,j}_{t+1} = \lambda cov^{i,j}_t + (1-\lambda) (x^i_{t+1}- mean^i_{t+1} )(x^j_{t+1}-mean^j_{t+1})

Then, we can simultaneously update our estimation for these parameters and the best portfolio. This is the purpose of the following code

#Inititalise the mean and variance from historical observations

mu = logR_sub.mean()
Sigma = logR_sub.cov()

mlambda = 519/520   # Roughly, one new observation counts like one week in 10 years (52 weeks x 10 years = 520 ). Also, we called it mlambda because lambda is a reserved kewyword
logR_2 = logR[logR.index>= '2016-12-01']
pi_new = ref_pi    # The initial equal weight portfolio

sharpe = []


out_dataframe = logR_2.copy()   # We are going to save here the investment protfolios for each date

for date in logR_2.index:
    lr = logR_2.loc[date]
    mu = mlambda*mu + (1-mlambda)*lr   # Calculate the updated estimation of the mean
    lr_hat = np.array(lr - mu)
    Sigma = mlambda*Sigma + (1-mlambda)*lr_hat.reshape((-1,1))@ lr_hat.reshape((1,-1))    # Calculate the updated estimation of the covariance
    
    
    #Use our algorithm to update the portfolio. Play with the step parameter and observe how the evolution of portfolios   and Sharpe ratio evolves
    pi_new = update_portfolio(pi_new,0.2,0.01, mmu=mu,msigma=Sigma )    
    new_sharpe = sharpe_portfolio(pi_new)   # Calculate the Sharpe ratio
    
    sharpe.append(new_sharpe)    #Save in a list for later plotting
    
    out_dataframe.loc[date]=pi_new   # Save the portfolio composition
        
     
plt.plot(logR_2.index,sharpe)
plt.title('Evolution of Sharpe ratio')

plt.figure()
out_dataframe.plot()
plt.title('Evolution of portfolio composition')

print(pi_new)
ticker BAC 0.010000 C 0.010000 GE 0.010000 JNJ 0.799978 MSFT 0.010000 PG 0.010000 T 0.010000 WMT 0.130022 XOM 0.010000 dtype: float64
Image in a Jupyter notebook
<Figure size 432x288 with 0 Axes>
Image in a Jupyter notebook

The figure shows that the strategy in this period becomes more and more concentrated on only one asset. Note, though that this process is not monotone, and sometimes the proportion changes. If the data starts supporting a more diversified portfolio.

Before finishing, let us check that our estimation (through the use of the EWMA estimator ) is close to the one we would have obtained by using the whole data. This occurs since we chose a factor that roughly gives every new data the same weight as one week in a period of 10 years (which is how it works within the whole data series).

print(logR.mean(), mu)
print(logR.cov(),Sigma)
ticker BAC -0.000786 C -0.003253 GE -0.000627 JNJ 0.001889 MSFT 0.002297 PG 0.001198 T 0.001146 WMT 0.001717 XOM 0.000750 dtype: float64 ticker BAC -0.000756 C -0.003210 GE -0.000691 JNJ 0.001899 MSFT 0.002320 PG 0.001197 T 0.001135 WMT 0.001753 XOM 0.000752 dtype: float64 ticker BAC C GE JNJ MSFT PG T \ ticker BAC 0.005745 0.005614 0.002074 0.000669 0.000883 0.000584 0.000844 C 0.005614 0.008234 0.002391 0.000721 0.000992 0.000550 0.001076 GE 0.002074 0.002391 0.001784 0.000361 0.000518 0.000362 0.000483 JNJ 0.000669 0.000721 0.000361 0.000442 0.000322 0.000264 0.000280 MSFT 0.000883 0.000992 0.000518 0.000322 0.001245 0.000303 0.000346 PG 0.000584 0.000550 0.000362 0.000264 0.000303 0.000495 0.000300 T 0.000844 0.001076 0.000483 0.000280 0.000346 0.000300 0.000770 WMT 0.000534 0.000500 0.000273 0.000237 0.000314 0.000281 0.000287 XOM 0.000839 0.000873 0.000506 0.000314 0.000450 0.000297 0.000390 ticker WMT XOM ticker BAC 0.000534 0.000839 C 0.000500 0.000873 GE 0.000273 0.000506 JNJ 0.000237 0.000314 MSFT 0.000314 0.000450 PG 0.000281 0.000297 T 0.000287 0.000390 WMT 0.000672 0.000264 XOM 0.000264 0.000797 ticker BAC C GE JNJ MSFT PG T \ ticker BAC 0.005726 0.005594 0.002066 0.000667 0.000880 0.000581 0.000842 C 0.005594 0.008203 0.002382 0.000719 0.000988 0.000548 0.001073 GE 0.002066 0.002382 0.001785 0.000360 0.000515 0.000360 0.000483 JNJ 0.000667 0.000719 0.000360 0.000442 0.000321 0.000263 0.000278 MSFT 0.000880 0.000988 0.000515 0.000321 0.001241 0.000301 0.000344 PG 0.000581 0.000548 0.000360 0.000263 0.000301 0.000494 0.000299 T 0.000842 0.001073 0.000483 0.000278 0.000344 0.000299 0.000770 WMT 0.000532 0.000498 0.000269 0.000236 0.000313 0.000280 0.000286 XOM 0.000836 0.000870 0.000505 0.000313 0.000449 0.000296 0.000389 ticker WMT XOM ticker BAC 0.000532 0.000836 C 0.000498 0.000870 GE 0.000269 0.000505 JNJ 0.000236 0.000313 MSFT 0.000313 0.000449 PG 0.000280 0.000296 T 0.000286 0.000389 WMT 0.000673 0.000263 XOM 0.000263 0.000794

Exercises

  1. Download 5 years of history for the price of 10 of your favorite stocks (use, for example yahoo finance). Set, a priori, a long-only portfolio that you would constitute in those shares and a minimal proportion investment on each asset. then reproduce the above discussion to tune progressively your portfolio.

  1. Assume you start with 100 pounds to invest. Moreover, assume that you have to pay a fee of 0.001 times the value of any operation you make in the market (for example, if you sell 50 pounds worth of one asset to buy 50 of another, you pay a fee of 0.0001*(50+50) = 1 cent).

Modify your code above to include the effect of this fee in your calculations (ideally, code it with a variable for the transaction cost)