Algorithmic Trading With Python using Zipline

Setup

Follow the instructions on the zipline tutorial for installation. You need python 3.5 or 3.6 for it to work. Installing with conda seemed to be most seamless.

Make sure that you set up a new python project with the conda environment as the project interpreter, or else you won't have access to zipline.

Once it's installed check out the algo.py file and copy/paste it into your local directory.

The Basics

Now is where our algorithm starts. The initialize method is run once per backtest, so when we run our algorithm we basically tell it what stocks we want to lookup and the symbol ids. Symbol ids are important for the complicated reason that sometimes ticker symbols change, but symbol ids are unique to each asset.

STOCKS = ['AMD', 'CERN', 'COST', 'DELL', 'GPS', 'INTC', 'MMM']

def initialize(algo):
    algo.stocks = STOCKS
    algo.symbol_ids = [algo.symbol(asset) for asset in STOCKS]

Here we tell it that we want our algorithm to run on these stocks.
Also, the algo keyword might be a bit confusing, but it basically acts as a self parameter for the algorithm.

Now to buy and sell stocks we need to tell the algorithm how much of each stock to buy and sell when we execute an order.

The make_weights function is an easy way for us to tell the algorithm that we want to order a percent of the stock based on how much the short moving average is above the long moving average.

def make_weights(algo, data):
    recent = data.history(assets=algo.symbol_ids, 
                          fields='close', 
                          bar_count=60, 
                          frequency='1d')
    short_mean = recent[-10:].mean()
    long_mean = recent[-50:].mean()
    weights = (short_mean - long_mean) / long_mean
    norm_weights = weights / weights.abs().sum()
    return norm_weights

The history method tells the algorithm that we want to consider the last 60 days of price. And the short/long mean variables are the mean price of the last 10 and 50 days respectively. We calculate the normalized percent change between these values to determine how large of a percent of our portfolio's value to invest in each asset each day.

Now the handle data method just executes the trades that we told the algorithm to make in the make_weights method.

def handle_data(algo, data):
    weights = make_weights(algo, data)
    for security in algo.symbol_ids:
        order_target_percent(security, weights[security])

To run the algorithm we simply execute the run_algorithm method.

start = pd.Timestamp('2016-1-1', tz='utc')
end = pd.Timestamp('2017-1-1', tz='utc')
results = run_algorithm(start=start, end=end, capital_base=1000, 
                        initialize=initialize, handle_data=handle_data)

Note that the initialize and handle_data methods are required as it definines our algorithm's behavior.
Any other function we create doesn't need to be included in run_algorithm.

To get usable results from this backtest we run the following pyfolio function to unpack results.

returns, positions, orders = pf.utils.extract_rets_pos_txn_from_zipline(results)

Some simple functions help us visualize our model's performance. The plots are too big to be effectively displayed on the screen so saving them to a png is usually easier.
This function lets us view our returns.

def plot_returns():
    pf.plot_rolling_returns(returns)
    plt.xlabel("Date")
    plt.ylabel("Returns")
    plt.savefig("returns")
    plt.show()

Check out some of the plots in the repository. Nothing is stellar since this is about as simple of an algorithm that you can make, but now you're well on your way to understanding the intracies of more complex trading algorithms.