Validation Note · #361154 · Cross-asset futures
Network momentum replicates as a mechanism, not a magnitude
Xingyue (Stacy) Pu, Stephen Roberts, Xiaowen Dong & Stefan Zohren · Oxford-Man Institute / arXiv q-fin
overstated
1.51→0.62
claimed → measured Sharpe
| Claimed | Measured | Δ | |
|---|---|---|---|
| Sharpe | 1.51 | 0.62 | −0.89 |
| Annualised return | 22.2% | 9.0% | −13.2pp |
| Max drawdown | −19.9% | −19.5% | +0.4pp |
| Plain 12-mo trend rule | — | 0.80 | beats it |
▲Replicate▲Honesty✗Deflate✗Crisis—Generalise▲Frictions✗Decay✗Originality
What the paper claims
Learn a graph linking ~40 futures by momentum co-movement, trade each asset on its neighbours' momentum, long/short at 15% vol. Claimed: Sharpe 1.51 net of costs against 1.13 for trading each asset's own momentum, 64 futures, 2000–2022.
What we measured
Same method on the roll-adjusted 40-contract reference panel, 2020–2026: Sharpe 0.62, return 9.0%, drawdown −19.5% (claimed −19.9% — the risk number is honest). The lift over individual momentum is +0.58 but the paired t-statistic is 1.5. A plain 12-month trend rule scores 0.80 on identical data.
Why the verdict
C2 decided it. No configuration clears multiple testing (best t = 1.81, Deflated Sharpe 0.74 against a 0.95 bar, PBO 0.71), and the model trails the naive trend benchmark the paper never shows.
Threats to validity
- The paper: its headline claim, network beats individual momentum, is never significance-tested in the paper itself.
- Our validation: 40 of 64 contracts, US-only outside commodities, on a 2020–2026 window — the paper's 2000–2022 daily magnitude is out of public-data reach.
Your library