In the previous lesson, you learned how to query and process data so that we can analyze it with Alphalens tear sheets. In this lesson, you will experience a few iterations of the alpha discovery phase of the quant workflow by analyzing those tear sheets.
In this lesson, we will:
create_information_tear_sheet().create_returns_tear_sheet().The following code expresses an alpha factor based on a company's net income and market cap, then creates an information tear sheet for that alpha factor. We will start analyzing the alpha factor by looking at it's information coefficient (IC). The IC is a number ranging from -1, to 1, which quantifies the predictiveness of an alpha factor. Any number above 0 is considered somewhat predictive.
The first number you should look at is the IC mean, which is an alpha factor's average IC over a given time period. You want your factor's IC Mean to be as high as possible. Generally speaking, a factor is worth investigating if it has an IC mean over 0. If it has an IC mean close to .1 (or higher) over a large trading universe, that factor is probably exceptionally good. In fact, you might want to check to make sure there isn't some lookahead bias if your alpha factor's IC mean is over .1
Run the cell below to create an information tear sheet for our alpha factor. Notice how the IC Mean figures (the first numbers on the first chart) are all positive. That is a good sign!
In [ ]:
from quantopian.pipeline.data import factset
from quantopian.pipeline import Pipeline
from quantopian.research import run_pipeline
from quantopian.pipeline.factors import CustomFactor, SimpleMovingAverage
from quantopian.pipeline.filters import QTradableStocksUS
from alphalens.tears import create_information_tear_sheet
from alphalens.utils import get_clean_factor_and_forward_returns
def make_pipeline():
# 1 year moving average of year over year net income
net_income_moving_average = SimpleMovingAverage(
inputs=[factset.Fundamentals.net_inc_af],
window_length=252
)
# 1 year moving average of market cap
market_cap_moving_average = SimpleMovingAverage(
inputs=[factset.Fundamentals.mkt_val],
window_length=252
)
average_market_cap_per_net_income = (market_cap_moving_average / net_income_moving_average)
# the last quarter's net income
net_income = factset.Fundamentals.net_inc_qf.latest
projected_market_cap = average_market_cap_per_net_income * net_income
return Pipeline(
columns = {'projected_market_cap': projected_market_cap},
screen = QTradableStocksUS() & projected_market_cap.notnull()
)
pipeline_output = run_pipeline(make_pipeline(), '2010-1-1', '2012-1-1')
pricing_data = get_pricing(pipeline_output.index.levels[1], '2010-1-1', '2012-2-1', fields='open_price')
factor_data = get_clean_factor_and_forward_returns(pipeline_output, pricing_data)
create_information_tear_sheet(factor_data)
Alphalens is useful for identifying alpha factors that aren't predictive early in the quant workflow. This allows you to avoid wasting time running a full backtest on a factor that could have been discarded earlier in the process.
Run the following cell to express another alpha factor called price_to_book, combine it with projected_market_cap using zscores and winsorization, then creates another information tearsheet based on our new (and hopefully improved) alpha factor.
Notice how the IC figures are lower than they were in the first chart. That means the factor we added is making our predictions worse!
In [ ]:
def make_pipeline():
net_income_moving_average = SimpleMovingAverage( # 1 year moving average of year over year net income
inputs=[factset.Fundamentals.net_inc_af],
window_length=252
)
market_cap_moving_average = SimpleMovingAverage( # 1 year moving average of market cap
inputs=[factset.Fundamentals.mkt_val],
window_length=252
)
average_market_cap_per_net_income = (market_cap_moving_average / net_income_moving_average)
net_income = factset.Fundamentals.net_inc_qf.latest # the last quarter's net income
projected_market_cap = average_market_cap_per_net_income * net_income
price_to_book = factset.Fundamentals.pbk_qf.latest
factor_to_analyze = projected_market_cap.zscore() + price_to_book.zscore()
return Pipeline(
columns = {'factor_to_analyze': factor_to_analyze},
screen = QTradableStocksUS() & factor_to_analyze.notnull()
)
pipeline_output = run_pipeline(make_pipeline(), '2010-1-1', '2012-1-1')
pricing_data = get_pricing(pipeline_output.index.levels[1], '2010-1-1', '2012-2-1', fields='open_price')
new_factor_data = get_clean_factor_and_forward_returns(pipeline_output, pricing_data)
create_information_tear_sheet(new_factor_data)
We found that the first iteration of our alpha factor had more predictive value than the second one. Let's see if the original alpha factor might make any money.
create_returns_tear_sheet() splits your universe into quantiles, then shows the returns generated by each quantile over different time periods. Quantile 1 is the 20% of assets with the lowest alpha factor values, and quantile 5 is the highest 20%.
This function creates six types of charts, but the two most important ones are:
Run the following cell, and notice how quantile 5 doesn't have the highest returns. Ideally, you want quantile 1 to have the lowest returns, and quantile 5 to have the highest returns. This tear sheet is telling us we still have work to do!
In [ ]:
from alphalens.tears import create_returns_tear_sheet
create_returns_tear_sheet(factor_data)
In this lesson, you experienced a few cycles of the alpha discovery stage of the quant worfklow. Making good alpha factors isn't easy, but Alphalens allows you to iterate through them quickly to find out if you're on the right track! You can usually improve existing alpha factors in some way by getting creative with moving averages, looking for trend reversals, or any number of other stratgies.
Try looking around Quantopian's forums, or reading academic papers for inspiration. This is where you get to be creative! In the next lesson, we'll discuss advanced Alphalens concepts.
In [ ]: