Long-only buys asymmetric exposure, not just lower Sharpe
what you'll learn · Why long-only momentum has higher per-seed Sharpe variance than long-short momentum on the same data, and what that means for operators with no-short mandates.
Added a long-only momentum arm to the 14-arm harness. Mean Sharpe was −0.196 vs the long-short baseline. The interesting number wasn't the mean — it was the stdev: 1.487 across 5 seeds, more than double baseline's 0.581. Long-only doesn't just give up the short leg's contribution. It gives up the dollar-neutral diversification that flattens per-seed dispersion.
XsLongOnlyMomentumStrategy shipped as the catalog’s first
long-only variant — top quantile only, no short leg, default
gross_leverage = 1.0. The 14-arm harness ran it against the
same 5 seeds and same synthetic shocks as the long-short
baseline.
Result:
baseline (long-short): mean +1.001, stdev 0.581, min +0.173, max +1.656
long_only: mean +0.805, stdev 1.487, min −0.930, max +2.446Mean is down (Δ −0.196), as expected — the short leg of XsMomentum captures ~half the dollar exposure on each tick, and removing it gives up roughly that half of the signal. A common reading is “long-only loses against long-short by the short leg’s Sharpe.”
The more useful reading is in the stdev column: 1.487 versus 0.581. Long_only’s per-seed Sharpe dispersion is more than 2.5× baseline’s. The min went negative (−0.930). The max nearly hit +2.5 (the highest max of any arm in the 14-arm harness).
Long-only doesn’t just reduce expected return. It changes the shape of the return distribution across seeds.
Why the dispersion widens
The long-short variant trades roughly dollar-neutral: long the top quantile, short the bottom, weights sum to ~0. Half of the gross is long, half is short. When the panel signal is broadly right, both legs contribute; when broadly wrong, both legs lose — but the long leg’s loss is partly offset by the short leg’s gain (the cross-sectional ranking is still meaningful even when the panel is wrong about direction).
Long-only puts the same gross exposure on the top quantile alone. There’s no short leg to offset. The strategy concentrates all risk on a single bet: “the top-ranked symbols will outperform on absolute terms.” When right, full exposure → big gains. When wrong, full exposure → big losses.
The variance of “single concentrated bet” >> variance of “diversified spread bet” when the bets are correlated to the same underlying signal. The math is the same as a single-stock vs. a 2-stock portfolio’s variance at fixed gross exposure.
What this means for operators
Three concrete consequences:
-
The Sharpe loss vs long-short isn’t the whole cost. A long-only mandate trades a known mean penalty for an unknown variance penalty. The variance penalty is regime-dependent — the same long-only strategy will look much better in directional regimes (where the short leg would have lost) and much worse in chop (where the long leg can’t be hedged).
-
Risk budgets are different. A long-short strategy with
gross_leverage = 2.0has roughly the same per-period volatility as a long-only strategy withgross_leverage = 1.0on the same universe — the long-short’s dollar exposure is double but its dollar-neutrality halves the variance. The harness configures both at the same lookback / quantile / leverage for direct comparability, which means long-only’s risk budget is actually tighter than the long-short’s. An operator wanting equal risk budget would run long-only with gross_leverage closer to 0.5 or 0.7. -
The drawdown character is different. Long-only drawdowns are correlated to broad-market drawdowns (the top quantile typically includes high-beta names, which crash hardest). Long-short drawdowns are correlated to factor reversals (the long and short legs both moving against the ranking). The risk events aren’t the same; an operator who can’t tolerate broad-market drawdowns shouldn’t pick long-only even if their mandate allows it.
What this rules out
-
Not “long-only is worse than long-short.” Different shape, different risk profile. The harness’s mean Δ is one number; the operator’s constraint, regime, and risk tolerance pick between them.
-
Not “always use the higher-Sharpe variant.” Sharpe ratios ignore higher moments. The long-only’s wider min-Sharpe means the worst seed was meaningfully bad; an operator who can’t survive a −0.930 Sharpe seed shouldn’t choose the higher- expected variant on Sharpe alone.
-
Not “the dispersion difference is a synthetic artifact.” The dollar-neutrality argument is a property of the strategy shape, not of the data — long-only WILL have higher per-period variance than dollar-neutral long-short on any data where the rankings are correlated to a common factor (which is most data).
The discipline rule
When ranking strategies for an operator with a no-short mandate, do not compare to long-short baselines on Sharpe alone. Add a column for stdev across seeds (or stdev across rolling windows on real data). The dispersion gap is the hidden cost of giving up dollar-neutrality.