Course goals¶
- Get exposure to quantitative investment strategies
- Learn the types of risk adjustments people do to analyze all investment strategies - information ratios, attribution analysis, ...
- Learn to execute python code to backtest and analyze strategies
I am not promising you will be proficient at writing python code by the end of the course, but ChatGPT can help.
Course materials¶
- Slides, assignments, and links to notebooks at mgmt638.kerryback.com
- Submit assignments on Canvas
- Three versions of slides: html, pdf, and Jupyter notebook
- The notebook is more inclusive. It contains all of the code to do all of the analysis that is presented in the html and pdf slides.
- Notebooks are set to open on Google Colab.
Grading¶
- Grades will be based on individual weekly assignments (80%) and class participation (20%).
- We will do work similar to the assignments in class each week, so there will be an opportunity for coaching.
Predictors¶
- Past prices
- moving averages, support and resistance, ...
- momentum and reversals
- pairs trading (error correction)
- Corporate fundamentals: ratios and growth rates
- Corporate actions (earnings, dividends, 8ks)
- Corporate insider trades
- Other trades - institutional, retail, short selling
- Social media sentiment
- Proprietary data (satellite images, phone location data, ...)
Backtesting¶
- Idea + data $\rightarrow$ backtest
- Can we reasonably backtest the strategy "buy electric car companies whose name starts with T?"
- We can backtest a more general strategy with two parameters: type of company, first letter of name
- We can backtest in a loop, updating once per year for example:
- Find the type of company and the first letter of name that did best in the past $n$ years
- Buy that company and hold for a year
- Update each year: find the new best company/first letter and hold it for a year
Tests and other considerations¶
- Past average return
- Sharpe ratio
- CAPM alpha and information ratio
- Fama-French alpha and information ratio
- Attribution analysis
- Maximum drawdown
- Tracking error relative to a benchmark
- Correlation with other strategies
- Turnover and ransactions costs (including shorting fees)
Universe of stocks¶
- Large cap or small cap or mid cap or some of all?
- Industries: do we want to bet on industries or match industry weights to a benchmark?
- Value vs growth, etc.
- Our goal could be to find the best possible strategy without any constraints or we might be constrained to find the best strategy within mid-cap energy, for example.
- Different strategies may work better or worse depending on the universe of stocks we can consider.
Example for today¶
- Do moving average strategies work?
- Get adjusted closing prices from Yahoo Finance
- Adjusted for splits and dividends
- % change is total return, including dividends
In [2]:
data.head()
Out[2]:
| closeadj | |
|---|---|
| Date | |
| 2010-06-29 | 1.592667 |
| 2010-06-30 | 1.588667 |
| 2010-07-01 | 1.464000 |
| 2010-07-02 | 1.280000 |
| 2010-07-06 | 1.074000 |
Compute and plot moving averages¶
- Compute average of adjusted closing price over prior $n$ days
- Do 10 day and 50 day as illustration
- Plot from 2020 on only so we can see detail better
In [4]:
_ = data[["ten", "fifty"]].loc["2020-01-01":].plot()
Compute returns¶
- Buy and hold return = percent change in adjusted closing price
- Moving average strategy:
- Long all money in account when 10 day > 50 day
- Zero position (and no interest for simplicity) otherwise
In [6]:
data.head()
Out[6]:
| closeadj | ten | fifty | buy_hold | long | mvg_avg | |
|---|---|---|---|---|---|---|
| Date | ||||||
| 2010-09-09 | 1.381 | 1.351 | 1.322 | -0.009 | True | -0.009 |
| 2010-09-10 | 1.345 | 1.357 | 1.318 | -0.026 | True | -0.026 |
| 2010-09-13 | 1.381 | 1.360 | 1.313 | 0.027 | True | 0.027 |
| 2010-09-14 | 1.408 | 1.366 | 1.311 | 0.019 | True | 0.019 |
| 2010-09-15 | 1.465 | 1.375 | 1.314 | 0.041 | True | 0.041 |
What tests do we want to do?¶
- Start by calculating average returns - multiply by 252 to annualize
- Then look at plot of compound returns - log scale works better for long time period
- Compute Sharpe ratios
- CAPM alphas, ...
Mean returns¶
In [7]:
buy_hold = 252*data.buy_hold.mean()
mvg_avg = 252*data.mvg_avg.mean()
print(f"buy and hold mean return is {buy_hold:.2%} annualized")
print(f"moving average mean return is {mvg_avg:.2%} annualized")
buy and hold mean return is 54.43% annualized moving average mean return is 37.46% annualized
Compound return plots¶
- We will plot the compound return (how much your money grows to starting from $1).
- First in a normal scale and then in a log scale.
In [8]:
_ = (1+data[["buy_hold", "mvg_avg"]]).cumprod().plot()
In [9]:
_ = (1+data[["buy_hold", "mvg_avg"]]).cumprod().plot(
logy=True
)
Sharpe ratios¶
- Sharpe ratio is expected return minus risk-free rate / standard deviation
- We'll skip the risk-free rate
- Annualize mean return by multiplying by $252$
- Annualize variance by multiplying by $252$
- $\Rightarrow$ annualize standard deviation by multiplying by $\sqrt{252}$
- $\Rightarrow$ annualize Sharpe ratio by multiplying by $\sqrt{252}$
In [11]:
print(f"Buy and hold Sharpe ratio is {buy_hold_sharpe:.2%} annualized")
print(f"Moving average Sharpe ratio is {mvg_avg_sharpe:.2%} annualized")
Buy and hold Sharpe ratio is 95.92% annualized Moving average Sharpe ratio is 88.76% annualized
Multiple stocks¶
- We can get data for multiple stocks from Yahoo Finance by passing a list of tickers.
- Here is an example.
In [13]:
data.head(7)
Out[13]:
| closeadj | ||
|---|---|---|
| date | ticker | |
| 2000-01-03 | CVX | 17.508469 |
| F | 13.405162 | |
| MSFT | 36.205597 | |
| PG | 28.428749 | |
| WMT | 43.717701 | |
| 2000-01-04 | CVX | 17.508469 |
| F | 12.957256 |
- Then we can do everything we did for a single ticker by running the code in a groupby object.
- Portfolio returns:
- Instead of looking at returns for each stock individually, we can compare portfolios.
- We will equal weight each day for simplicity.
- Neither of the strategies is buy and hold - we have to trade each day to get back to equal weights.
- For the moving average strategy, we will equal weight the stocks for which the 10 day > 50 day moving average (which could be no stocks or all stocks or anything in between).
In [16]:
rets.head()
Out[16]:
| eq_wtd | mvg_avg | |
|---|---|---|
| date | ||
| 2000-01-03 | NaN | 0.0 |
| 2000-01-04 | -0.024771 | 0.0 |
| 2000-01-05 | -0.001449 | 0.0 |
| 2000-01-06 | 0.013458 | 0.0 |
| 2000-01-07 | 0.051980 | 0.0 |
In [17]:
rets.tail()
Out[17]:
| eq_wtd | mvg_avg | |
|---|---|---|
| date | ||
| 2023-10-16 | 0.010296 | 0.004630 |
| 2023-10-17 | 0.004664 | 0.002299 |
| 2023-10-18 | 0.000899 | 0.000413 |
| 2023-10-19 | -0.004877 | 0.000946 |
| 2023-10-20 | -0.006354 | -0.005492 |
Mean returns¶
In [18]:
eq_wtd = 252*rets.eq_wtd.mean()
mvg_avg = 252*rets.mvg_avg.mean()
print(f"equally weighted mean return is {eq_wtd:.2%} annualized")
print(f"moving average mean return is {mvg_avg:.2%} annualized")
equally weighted mean return is 10.60% annualized moving average mean return is 5.65% annualized
Compound returns¶
In [19]:
_ = (1+rets).cumprod().plot()
In [20]:
_ = (1+rets).cumprod().plot(logy=True)
Sharpe ratios¶
In [22]:
print(f"Equally weighted Sharpe ratio is {eq_wtd_sharpe:.2%} annualized")
print(f"Moving average Sharpe ratio is {mvg_avg_sharpe:.2%} annualized")
Equally weighted Sharpe ratio is 35.35% annualized Moving average Sharpe ratio is 29.48% annualized
Exercise¶
Look at different sets of stocks and different moving averages and test strategies.