Factor Investing

Factor investing and smart beta strategies — systematic exposure to return drivers.

When to Activate

• User is building factor-based portfolios (value, momentum, quality, low-vol, size)
• Designing smart beta strategies or evaluating factor ETFs
• Analyzing factor crowding, timing, or cycle dynamics
• Constructing multi-factor portfolios and managing factor interactions
• Evaluating implementation costs and capacity constraints of factor strategies

Core Concepts

What Are Factors?

Factors are persistent, systematic drivers of returns across asset classes. The academic canon:

Factor Construction

Long-short factors:

1. Rank stocks on the characteristic (e.g., B/P for value)
2. Form quintile or decile portfolios (cross-sectionally within a universe)
3. Go long top quintile, short bottom quintile
4. Weight by equal-weight or value-weight within each leg

Long-only smart beta:

1. Start from a broad benchmark (e.g., Russell 1000)
2. Tilt weights toward desired factor exposure
3. Apply constraints (sector neutrality, turnover, tracking error)

Factor Definitions in Practice

Value: Book-to-Price, Earnings Yield (E/P), EBITDA/EV, CF/P, Sales/Price. Composite signals (multi-metric) are more robust than any single measure.

Momentum: 12-month return minus the most recent month (12-1 momentum). The last month is excluded to avoid short-term reversal contamination. Typical holding period: 3-12 months.

Quality: Gross profitability (GP/Assets), ROE, earnings stability, low leverage, low accruals. AQR's QMJ factor combines profitability, growth, safety, and payout.

Low Volatility: Realized volatility (trailing 1-3 years), beta, or idiosyncratic volatility. BAB (Betting Against Beta) is the levered/de-levered formulation from Frazzini-Pedersen.

Size: Market capitalization. Small-cap premium has been weak post-publication. Survives mainly in micro-caps and in interaction with other factors (small-cap value).

Methodology

Step 1: Factor Signal Construction

import pandas as pd
import numpy as np

def compute_value_signal(fundamentals):
    """Composite value signal from multiple metrics."""
    signals = pd.DataFrame()
    signals['bp'] = fundamentals['book_value'] / fundamentals['market_cap']
    signals['ep'] = fundamentals['earnings'] / fundamentals['market_cap']
    signals['cfp'] = fundamentals['cash_flow'] / fundamentals['market_cap']

    # Cross-sectional z-score each metric
    for col in signals.columns:
        signals[col] = (signals[col] - signals[col].mean()) / signals[col].std()

    # Composite: equal-weight average of z-scores
    return signals.mean(axis=1)

def compute_momentum_signal(prices, lookback=252, skip=21):
    """12-1 month momentum."""
    total_return = prices.pct_change(lookback)
    recent_return = prices.pct_change(skip)
    return total_return - recent_return

def compute_quality_signal(fundamentals):
    """Profitability-based quality."""
    gp_a = fundamentals['gross_profit'] / fundamentals['total_assets']
    roe = fundamentals['net_income'] / fundamentals['book_value']
    accruals = (fundamentals['net_income'] - fundamentals['cash_flow']) / fundamentals['total_assets']
    # Z-score and combine
    signals = pd.DataFrame({'gpa': zscore(gp_a), 'roe': zscore(roe), 'accruals': -zscore(accruals)})
    return signals.mean(axis=1)

Step 2: Portfolio Formation

def form_factor_portfolio(signal, returns, n_quantiles=5, weighting='equal'):
    """
    Sort stocks into quantiles based on signal, compute long-short returns.
    """
    quantiles = signal.groupby(signal.index.get_level_values('date')).transform(
        lambda x: pd.qcut(x, n_quantiles, labels=False, duplicates='drop')
    )

    long_mask = quantiles == (n_quantiles - 1)
    short_mask = quantiles == 0

    if weighting == 'equal':
        long_ret = returns[long_mask].groupby('date').mean()
        short_ret = returns[short_mask].groupby('date').mean()
    elif weighting == 'value':
        # Value-weight within each leg
        pass

    factor_return = long_ret - short_ret
    return factor_return

Step 3: Multi-Factor Portfolio Construction

def multi_factor_portfolio(signals_dict, sigma, risk_aversion=1.0,
                           sector_neutral=True, max_active_weight=0.02):
    """
    Combine multiple factor signals into a single portfolio.

    Approaches:
    1. Signal blending: average z-scores, then optimize
    2. Portfolio blending: optimize each factor separately, average weights
    3. Integrated optimization: optimize with all factor tilts simultaneously
    """
    import cvxpy as cp

    # Approach 1: Signal blending (preferred — avoids cancellation)
    composite_signal = sum(signals_dict.values()) / len(signals_dict)

    n = len(composite_signal)
    w = cp.Variable(n)
    w_bench = np.ones(n) / n  # equal-weight benchmark

    # Maximize factor exposure subject to risk constraints
    objective = cp.Maximize(composite_signal.values @ w)
    constraints = [
        cp.sum(w) == 1,
        w >= 0,
        cp.norm(w - w_bench, 'inf') <= max_active_weight,  # position limit
        cp.quad_form(w - w_bench, sigma) <= te_budget**2,   # TE constraint
    ]

    if sector_neutral:
        for sector_mask in sector_masks:
            constraints.append(
                cp.sum(w[sector_mask]) == cp.sum(w_bench[sector_mask])
            )

    prob = cp.Problem(objective, constraints)
    prob.solve()
    return w.value

Step 4: Factor Timing (Use with Caution)

Factor timing attempts to vary factor exposures based on macro or valuation signals. Evidence is mixed:

def factor_valuation_signal(factor_returns, lookback=60):
    """
    Factor valuation spread: compare current value spread
    to historical average. Wide spreads -> overweight value.
    """
    # Value spread: median P/E of cheap quintile vs expensive quintile
    # When spread is wide (cheap stocks are very cheap), value factor is attractive
    pass

def macro_factor_timing(macro_data):
    """
    Macro-based factor rotation:
    - Early cycle (recovery): small cap, value, high beta
    - Mid cycle (expansion): momentum, quality
    - Late cycle (slowdown): quality, low vol, defensive
    - Recession: low vol, quality, bonds
    """
    pass

Arnott et al. (2019) show factor timing based on valuation spreads has modest efficacy. Most practitioners use fixed strategic factor weights with limited tactical adjustment.

Step 5: Factor Crowding Analysis

def detect_crowding(factor_signal, short_interest, fund_flows):
    """
    Crowding indicators:
    1. Factor valuation spread compression (cheaper than usual = less crowded)
    2. Pairwise correlation of factor stocks (high = crowded)
    3. Short interest concentration in short-leg stocks
    4. AUM growth in factor ETFs/funds
    5. Factor return autocorrelation turning negative (reversal = crowding)
    """
    pass

Crowding matters because:

• Crowded factors are vulnerable to rapid unwinds (quant crisis of August 2007)
• Transaction costs increase as many managers trade the same signals
• Alpha decays as the factor becomes well-known (publication effect)

Examples

Evaluating Factor ETF Exposure

# Regress ETF returns on Fama-French 5 factors + momentum
import statsmodels.api as sm

ff_factors = pd.read_csv('F-F_Research_Data_5_Factors_2x3_daily.csv')
etf_returns = pd.read_csv('MTUM_returns.csv')  # iShares Momentum ETF

merged = etf_returns.merge(ff_factors, on='date')
X = merged[['Mkt-RF', 'SMB', 'HML', 'RMW', 'CMA', 'Mom']]
y = merged['ETF_excess']

model = sm.OLS(y, sm.add_constant(X)).fit()
# Check: MTUM should have high Mom loading, modest market beta

Implementation Cost Analysis

def factor_turnover_analysis(signal, rebalance_freq='monthly'):
    """
    Key cost metrics for factor strategies:
    - One-way turnover per rebalance
    - Estimated market impact (higher for small-cap, momentum)
    - Net-of-cost factor premium
    """
    # Momentum: 80-100% annual one-way turnover (expensive)
    # Value: 20-40% annual turnover (cheap)
    # Quality: 15-30% annual turnover (cheapest)
    # Low-vol: 25-40% annual turnover (moderate)
    pass

Novy-Marx and Velikov (2016) show that many published anomalies disappear after realistic transaction costs. Momentum is especially vulnerable.

Quality Gate

• Factor signals use point-in-time data — no look-ahead bias in fundamentals
• Cross-sectional signals are sector/industry neutralized where appropriate
• Momentum signal excludes the most recent month (reversal contamination)
• Multi-metric composites used rather than single-metric signals
• Transaction costs and turnover estimated realistically (especially for momentum)
• Factor crowding risk assessed — check valuation spreads and strategy AUM
• Out-of-sample and out-of-region tests performed (US factor ≠ global factor)
• Factor interactions checked — value and momentum are negatively correlated (beneficial)
• Capacity constraints estimated for small-cap factor strategies
• Reporting includes both gross and net-of-cost factor returns
• Compared against simple market-cap benchmark over the full cycle (including drawdowns)