Risk Model¶
FactorRiskModel is a fundamental-style multi-factor risk model. Each day it regresses the stocks' returns on their
factor exposures, which are qex expressions, plus a market factor and, optionally, industry dummies. The regression
coefficients are that day's factor returns. From them it estimates a factor covariance matrix and each stock's
specific risk, and uses those to forecast a portfolio's risk and attribute it to the factors.
It needs the risk extra: pip install "qex[risk]" (statsmodels).
from qex import FactorRiskModel, PRICE_ONLY_FACTORS
import pandas as pd
model = FactorRiskModel(style_factors=PRICE_ONLY_FACTORS, include_industries=True, industry_column='sector')
model.fit(data, lookback=252) # returns the model
model.factor_returns # dates x factors: each day's factor returns
model.factor_covariance # factors x factors, annualized
model.specific_risk # each stock's annualized specific (idiosyncratic) volatility
weights = pd.Series({'AAPL': 0.1, 'MSFT': -0.1, 'JPM': 0.05, 'BAC': -0.05})
total, factor, specific = model.portfolio_risk(weights) # annualized volatilities
model.factor_attribution(weights) # each factor's contribution to variance
How it is estimated¶
- Exposures are the style factors' expressions, evaluated on the data, plus 1 for every stock on the
marketfactor and a 0/1 dummy per industry wheninclude_industriesis set. - Point in time. Day t's returns are regressed on the exposures known at the close of day t − 1, so a factor built from day t's own return (short-term reversal, volatility) can't explain that return with hindsight.
- Returns follow qex's rule for gaps (Halts, Delistings and Gaps): a day with no price has no return and leaves that day's regression.
- The last
lookbackdays are fitted, one cross-sectional regression a day, over the stocks with every exposure and a return; a day with fewer thanmin_observationsstocks (default 10) is skipped. - Covariances are exponentially weighted, with a
half_life(default 63 days), and annualized (× 252). They are proper weighted covariances: with equal weights, the factor covariance is exactly that of the factor returns. - Forecasts use the exposures at the close of the last fitted day, so
portfolio_riskforecasts the risk of a book held from that day on.
Factors¶
A factor is a FactorDefinition(name, expression, description): any qex expression that gives each stock a
number each day. Two sets come with qex:
DEFAULT_STYLE_FACTORS (needs fundamental fields market_cap and book_to_price):
| Factor | Expression |
|---|---|
size |
rank(field('market_cap')) |
value |
rank(winsorize(field('book_to_price'), 0.05)) |
momentum |
rank(returns(252) - returns(21)) |
volatility |
rank(volatility(60)) |
liquidity |
rank(adv(20)): 20-day average dollar volume |
short_term_reversal |
rank(-returns(5)) |
PRICE_ONLY_FACTORS (prices and volume only):
| Factor | Expression |
|---|---|
momentum |
rank(returns(252) - returns(21)) |
volatility |
rank(volatility(60)) |
liquidity |
rank(adv(20)) |
short_term_reversal |
rank(-returns(5)) |
medium_term_reversal |
rank(-returns(20)) |
Your own:
from qex import FactorRiskModel, FactorDefinition
factors = [
FactorDefinition('earnings_yield', "rank(field('earnings') / field('price'))", 'Earnings yield rank'),
FactorDefinition('beta', "rank(ts_beta_shrunk(returns(1), returns(1, 'market'), 252))", 'Market beta rank'),
]
model = FactorRiskModel(style_factors=factors).fit(data)
style_factors=None uses DEFAULT_STYLE_FACTORS. A factor that can't be computed on the data (a missing field) is
skipped with a warning.
For a statistical model instead, with factors found from the returns themselves, see pca_factor in the
Operator Reference.
API reference¶
FactorRiskModel ¶
Multi-factor risk model.
The model decomposes stock returns into
r_i = sum_k(X_ik * f_k) + epsilon_i
Where
r_i = return of stock i X_ik = exposure of stock i to factor k f_k = return of factor k epsilon_i = specific (idiosyncratic) return
Attributes:
| Name | Type | Description |
|---|---|---|
style_factors |
List of style factor definitions |
|
include_industries |
Whether to include industry dummy factors |
|
industry_column |
Name of industry column in data |
|
half_life |
Half-life for exponential weighting (days) |
fit ¶
Fit the risk model using cross-sectional regression.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
data
|
Data dict or LazyData with price and factor data |
required | |
lookback
|
int
|
Number of days to use for estimation |
252
|
min_observations
|
int
|
Minimum observations required per stock |
10
|
Returns:
| Type | Description |
|---|---|
FactorRiskModel
|
self (fitted model) |
get_exposures ¶
Get factor exposures for a specific date.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
data
|
Data dict or LazyData |
required | |
date
|
Optional[str]
|
Date string (default: most recent) |
None
|
Returns:
| Type | Description |
|---|---|
DataFrame
|
DataFrame with stocks as rows, factors as columns |
portfolio_risk ¶
portfolio_risk(weights: Series, exposures: Optional[DataFrame] = None) -> Tuple[float, float, float]
Compute portfolio risk decomposition.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
weights
|
Series
|
Portfolio weights indexed by ticker |
required |
exposures
|
Optional[DataFrame]
|
Factor exposures (default: use fitted exposures) |
None
|
Returns:
| Type | Description |
|---|---|
Tuple[float, float, float]
|
Tuple of (total_risk, factor_risk, specific_risk) as annualized std |
factor_attribution ¶
Compute risk attribution to each factor.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
weights
|
Series
|
Portfolio weights |
required |
exposures
|
Optional[DataFrame]
|
Factor exposures (default: use fitted) |
None
|
Returns:
| Type | Description |
|---|---|
Series
|
Series with variance contribution of each factor |
FactorDefinition
dataclass
¶
Definition of a style factor.