1.4.6 Advanced Extensions

Basic Prerequisites: Econometrics & Time Series (conditional moments, EWMA vol) and Probability & Measure Theory.

---

1. Intuition & Practical Objective

Momentum’s problem is not its mean - it is its risk: volatility that clusters and crashes concentrated in panic states. The advanced playbook attacks this directly by scaling positions with ex-ante risk. There are two canonical approaches:

1. Constant-volatility (volatility-managed) momentum - Barroso & Santa-Clara (2015). Scale the whole WML strategy by its own forecastable volatility: ${\text{WML}}_t^{\ast }=\frac{{\sigma }_{\text{tgt}}}{{\hat{\sigma }}_t}{\text{WML}}_t$. Because momentum volatility clusters, the scale is already tiny entering a crash, so the crash is mostly avoided. Verified result: Sharpe 0.53 → 0.97 (roughly doubled), monthly skewness $-2.47\to -0.42$, excess kurtosis $18.2\to 2.7$, worst-month return $-78.96\%$ effectively eliminated.
2. Dynamic (optimal) momentum - Daniel & Moskowitz (2016). Go further and scale by the forecastable mean as well as the variance, because momentum’s conditional Sharpe is itself forecastable (it collapses in panic states). The optimal dynamic weight keeps the strategy’s conditional volatility proportional to its conditional Sharpe ratio. Verified result: more than doubles the static Sharpe and delivers an annualized Sharpe of 1.18 across all markets/asset classes.

The intellectual core: risk-managing momentum is not just cutting volatility - it is re-allocating capital toward the states where momentum’s return per unit of risk is high (calm) and away from the states where it is low (panic rebounds).

---

2. Mathematical Ground Truth & Derivations

Barroso–Santa-Clara vol-managed momentum. Forecast WML’s monthly variance from the trailing 126 daily returns (scaled by 21 to monthly units),

$${\hat{\sigma }}_t^2=21\frac{1}{126}\sum _{j=0}^{125}{r}_{\text{WML},d(t-1-j)}^2,$$

then scale

$${\text{WML}}_t^{\ast }=\frac{{\sigma }_{\text{tgt}}}{{\hat{\sigma }}_t}{\text{WML}}_t,{\sigma }_{\text{tgt}}=12\%\text{annualized}.$$

Since WML is zero-investment and self-financing, scaling is unconstrained and the strategy stays self-financing (weights on the long and short legs vary in tandem). Because volatility is persistent (${\hat{\sigma }}_t$ is high entering a crash), the scale suppresses exactly the bad months. The mean of WML is also higher per unit risk after scaling: risk-managed momentum earns +2.04 pp/year more with 10.58 pp/year less volatility.

Daniel–Moskowitz optimal dynamic scaling. To maximize the unconditional Sharpe ratio, at each date scale so that the strategy’s conditional volatility is proportional to its conditional Sharpe ratio. Writing conditional mean ${\mu }_t$ and vol ${\sigma }_t$, the optimal position weight is

$$w_t\propto \frac{{\mu }_t}{{\sigma }_t^2},$$

i.e. capital goes where the forecasted return-per-risk is high. In panic states ${\mu }_t$ collapses (the written-call option-like payoff of losers makes momentum’s expected return negative in a rebound) and ${\sigma }_t$ spikes, so $w_t\to 0$ - exactly when the static strategy bleeds. This exploits the forecastability of the mean, which pure constant-vol scaling ignores, and is why the dynamic version beats the constant-vol version in spanning tests.

Why the mean is forecastable. In the panic state the short-loser leg is an out-of-the-money option-like position (loser beta $>3$), so its expected contribution flips sign and its convexity is large - measurable ex-ante from trailing market returns and VIX-type vol. Hence ${\mu }_t$ is not white noise; it is a function of the observable state.

---

3. Computational Implementation - risk-managing a crash-prone strategy

Standard library only. Builds a synthetic WML series with two embedded crashes (1932-style and 2009-style) preceded by elevated vol, then applies the Barroso scaling. Verifies: Sharpe improves, worst month and max drawdown shrink dramatically, realized vol approaches target.

Vol-scaling lifts Sharpe from $+0.72$ to $+1.18$, cuts the worst month from $-45.0\%$ to $-12.5\%$, and shrinks max drawdown from $-70.5\%$ to $-19.1\%$ - while holding realized vol near the 12% target. This is the same qualitative result Barroso–Santa-Clara find on real data (Sharpe 0.53→0.97, crashes nearly eliminated).

---

4. Failure Modes & First-Principles Breakdowns

1. Vol-scaling lags the crisis. The scale at month $t$ uses ${\hat{\sigma }}_{t-1}$; if a crash arrives with no preceding vol elevation (a pure gap/event shock), scaling cannot dodge it. It works because volatility clusters - it will not protect against a jump that does not cluster.
2. Volatility ≠ expected return. Constant-vol scaling equalizes risk but says nothing about where momentum’s mean is favorable. The Daniel–Moskowitz dynamic version adds a conditional-mean forecast; pure vol-scaling leaves the mean-timing on the table (their dynamic strategy beats the constant-vol strategy in spanning tests).
3. Forecast fragility. Conditional-mean forecasts in panic states are estimated from very few historical events (a handful of crashes in a century). Small-sample estimation error can flip the sign of ${\mu }_t$ exactly when it matters most - the parameter-estimation tail risk.
4. Leverage creep. When vol is low, ${\sigma }_{\text{tgt}}/{\hat{\sigma }}_t>1$, the scale levers up; the Barroso weights range roughly 0.13–2.00 with an average near 0.90. Excessive leverage in calm periods re-introduces the crash it was meant to avoid if the calm-to-panic transition is faster than the vol forecast.
5. Self-financing constraints are not free. Scaling WML assumes you can change the long/short dollar balance without cost; in practice the loser (short) leg is exactly the expensive, capacity-constrained one.

---

5. References

• Barroso & Santa-Clara (2015), Momentum Has Its Moments, J. Financial Economics 116(1)
• Daniel & Moskowitz (2016), Momentum Crashes, J. Financial Economics 122(2)
• Moskowitz, Ooi & Pedersen (2012), Time Series Momentum
• Baltas & Kosowski (2013), Demystifying Time-Series Momentum

---

6. Connected Graph Bridges

• Base: Econometrics & Time Series (conditional moments, EWMA) · Probability & Measure Theory
• Foundation: 05 · Failure Modes (the crashes this page tames) · 02 · Cross-Sectional · 03 · Time-Series
• Portfolio context: Risk Parity (vol targeting) · ES) · Backtesting Hygiene
• Home: Index Hub