Does volatility scaling create the TSMOM edge?
A common objection to time-series momentum (TSMOM): the famous backtest only looks good because of volatility scaling — i.e. leverage dressed up as an anomaly. Jo & Kim (2019) re-test that claim with futures data. This lesson walks through what they actually find, with numbers taken only from their paper.
- What TSMOM and volatility scaling mean in plain English
- Why Kim, Tse & Wald (2016) argued the edge was “just” scaling
- How Jo & Kim (2019) re-test: unscaled vs scaled, and with a passive-long control
- The paper’s monthly alpha ranges — and why “scaling creates the edge” is incomplete
- How this sits next to our vol-targeting teardown and the validation gauntlet
The claim under the microscope
Moskowitz, Ooi & Pedersen (2012) documented time-series momentum: for each asset, go long if its own past return was positive, short if negative — typically a 12-month look-back and a one-month hold. Their published strategy also used volatility scaling: size each futures position inversely to its own recent volatility (often toward a fixed risk target such as 40% annualized vol). That makes high-vol contracts smaller and low-vol contracts larger, so the portfolio mixes assets more evenly by risk.
Kim, Tse & Wald (2016) pushed back. In Jo & Kim’s summary of that critique: without volatility scaling, TSMOM alphas look similar to a simple passive long futures book; TSMOM only statistically beats passive long with scaling in the 1985–2009 window. The implication many readers took away: maybe there is no TSMOM anomaly — only leverage / risk-parity packaging.
That is a serious objection. If true, “time-series momentum” is a story about how you size, not about predictive signs from an asset’s own past returns.
What Jo & Kim re-test
Jo & Kim (2019), Revisiting the Time Series Momentum Anomaly (Annals of Economics and Finance 20-2, 767–782), ask a sharper question than “do the two alphas look different?” They ask whether a passive long futures factor can explain away TSMOM when you put both in the same regression — and whether that answer depends on volatility scaling.
- Universe: 60 futures (9 equity, 13 bond, 31 commodity, 7 currency), January 1984–August 2017.
- Main comparable window: January 1985–December 2009 (same span as MOP / Kim–Tse–Wald for the headline tables).
- Strategy: 12-month look-back, one-month hold; unscaled vs volatility-scaled versions of TSMOM and of a passive long futures book.
- Tests: Fama–French 3-factor + momentum, Fama–French 5-factor + momentum, and Asness–Moskowitz–Pedersen (2013) factors — with and without the passive-long control.
The numbers (Table 2, 1985–2009)
Jo & Kim report that TSMOM produces a significant alpha with and without volatility scaling. From their discussion of Table 2 (Panels A–C):
| Specification | Monthly alpha range | t-stats (paper) |
|---|---|---|
| Unscaled TSMOM (factor models, no passive-long control) | 0.18% – 0.38% / mo | 2.22 – 4.36 |
| Volatility-scaled TSMOM (same) | 0.80% – 1.20% / mo | 5.18 – 7.21 |
| Unscaled TSMOM with passive-long control | 0.16% – 0.36% / mo | 1.99 – 4.18 |
| Volatility-scaled TSMOM with passive-long control | 0.69% – 1.08% / mo | 4.50 – 6.48 |
Source: Jo & Kim (2019), Table 2 discussion (Jan 1985–Dec 2009). Ranges span their FF3+mom, FF5+mom, and Asness et al. (2013) specifications.
Two pedagogical takeaways fall out of those cells:
- Scaling changes the magnitude — scaled alphas are several times larger in the paper’s tables. Jo & Kim themselves note that scaled positions are leveraged, and that in futures practice margins of roughly 5%–20% of notional are common, so an unscaled book is not “more realistic” in a naïve sense.
- Scaling does not create the statistical anomaly in their tests. Unscaled TSMOM still shows significant intercepts across models; adding the passive-long control shrinks alphas a bit but does not wipe them out (0.16%–0.36%/mo unscaled; 0.69%–1.08%/mo scaled).
One-sentence summary of the paper’s punchline
In Jo & Kim’s sample and factor horse-race, TSMOM remains a significant residual after you strip out volatility scaling and after you control for a diversified passive long futures factor — so “the edge was just vol-scaling / leverage” is an incomplete reading of the evidence they report.
Does it survive past 2009?
They extend the same regressions through August 2017 (Table 3). Their summary: TSMOM still produces significant positive alphas whether positions are scaled or not, whether or not the passive-long control is included, and across the same family of factor models. Global Fama–French specifications (Tables 4–5) point the same way in their write-up.
That does not mean “trade this tomorrow.” It means that, within this paper’s construction, the post-publication extension did not erase the anomaly they measure.
How this fits Quant for Free
Our volatility-targeting teardown asks a related but different question: does targeting risk improve the shape of returns for a given signal? Useful, and still not a free lunch — leverage, estimator choice, and caps matter.
This Learn piece is the conceptual sibling: when someone dismisses TSMOM (or a trend book) with “it only worked because of vol-scaling,” Jo & Kim’s tables are the counter-evidence you should actually cite — unscaled alphas still show up, and a passive long control does not fully absorb them. Pair that honesty with the validation gauntlet: disclose how many variants you tried, report scaled and unscaled side by side, and use tools like the Deflated Sharpe Ratio so a pretty scaled curve does not get a free pass.
- TSMOM
- Time-series momentum: long/short each asset from its own past return sign (here, 12×1).
- Volatility scaling
- Sizing inversely to ex-ante volatility so risk contributions are more equal (risk parity–like).
- Passive long futures
- Always-long diversified futures book used as a control for “just being in futures.”
- Alpha (intercept)
- Residual monthly return after the chosen factor model — the paper’s measure of the anomaly.
- Unscaled vs scaled
- Same sign rule; scaled multiplies by a vol target / σt, unscaled does not.
Where to go next
AnalysisVolatility targeting teardown — when risk targeting helps the curve, and when it is just leverage Analysis
Dual momentum teardown — a related absolute/relative momentum case study Tool
Deflated Sharpe Ratio — discount a proud Sharpe for how many variants you tried Learn · Module 5
The validation gauntlet — overfitting, out-of-sample, multiple testing, costs