Mean Period Wise Spread (bps) leverage is 2 in Returns Analysis table

[Peropero0]

Problem Description

In Return Analysis section, the mean period wise return by factor quantile is computed as the period wise return of a portfolio made with assets belonging to the quantile. This information is useful to understand what forward returns one could expect when going long 100% a portfolio made of instruments belonging to a specific quantile.

The mean period wise spread though, is computed as the difference of mean period wise returns of specific quantiles, i.e. return of quantile 2 minus return of quantile 1. This should be useful to understand what forward returns to expect when building a long short portfolio.

Alphalens computes the spread as the difference of the two quantiles portfolios returns. In our example, this is like going 100% long the quantile 2 portfolio and 100% short the quantile 1 portfolio. In this way, the resulting portfolio is dollar neutral, but its leverage is 200%.

To account for a 100% leverage, Alphalens should rescale long portfolios to a total weight of 50% and short portfolios to -50%.

Please provide a minimal, self-contained, and reproducible example:

# imports
import pandas as pd
import alphalens

# build a dataframe with some signals for assets a and b
dates = pd.date_range('2020-01-02', '2020-01-07')
assets = ['asset_a', 'asset_b']

signals = pd.DataFrame({'signal': [1, 2, 2, 1, 1, 2, 1, 2, 2, 1, 1, 2]}, index = pd.MultiIndex.from_product([dates, assets]))
signals.index = signals.index.rename(['date', 'asset'])

# build a dataframe with some prices for the instruments
prices = pd.DataFrame(
    {
    'asset_a': [10, 11, 10, 12, 10, 11, 13],
    'asset_b': [10, 9, 10, 11, 12, 11, 12],
    'date': pd.date_range('2020-01-02', '2020-01-08')
    })

prices = prices.set_index('date')

# get clean factor and forward returns, we want to go long the asset with signal = 2 and short the asset with signal = 1
quantiles = alphalens.utils.get_clean_factor_and_forward_returns(
        signals['signal'],
        prices,
        periods=(1,),
        quantiles=2
    )

# alphalens mean return spread is -253 bps
alphalens.tears.create_summary_tear_sheet(quantiles, long_short=True)

# now I compute the mean return spread without alphalens.
# I compute the mean daily return by quantile as the average of daily return by quantile (this is trivial since we have just one asset per quantile)
spread_df = quantiles.groupby(['date', 'factor_quantile'])['1D'].mean().unstack()

# then I compute the daily mean return spread using the following formula:
# daily mean return spread = daily mean fwd return quantile 2 - daily mean fwd return quantile 1
long_weight = 1
short_weight = -1
spread_df['mean_return_spread'] = long_weight * spread_df[2] + short_weight * spread_df[1]

# this results in -253 bps as computed using alphalens
# notice that we are dollar neutral but with a leverage of 2
print(f"mean return spread = {spread_df['mean_return_spread'].mean() * 10000} bps")
print("")
print(f"Long weight = {long_weight},  Short weight = {short_weight}")
print(f"Long leg + short leg weights = {long_weight + short_weight}")
print(f"total absolute weights (leverage) = {abs(long_weight) + abs(short_weight)}")

# now do the same, but set long weights and short weights to +0.5 and -0.5 respectively
# notice that this results in a halved mean return spread because we are now dollar neutral and with leverage 1

long_weight = 0.5
short_weight = -0.5

spread_df = quantiles.groupby(['date', 'factor_quantile'])['1D'].mean().unstack()
spread_df['mean_return_spread'] = long_weight * spread_df[2] + short_weight * spread_df[1]

print(f"mean return spread = {spread_df['mean_return_spread'].mean() * 10000} bps")
print("")
print(f"Long weight = {long_weight},  Short weight = {short_weight}")
print(f"Long leg + short leg weights = {long_weight + short_weight}")
print(f"total absolute weights (leverage) = {abs(long_weight) + abs(short_weight)}")

Please provide any additional information below:
We would like to do a pull request to implement a solution.

Versions

• Alphalens version: 0.4.2
• Python version: 3.9.9
• Pandas version: 1.5.3

[stefan-jansen]

Fair point, a PR would be welcome.