Live managed futures: what the funds actually paid, and what the vendor index overstated

Question. Every trend figure in this repository comes from AQR’s published TSMOM series, which states no fee, transaction-cost, slippage or financing basis anywhere. What did real managed-futures funds pay their investors over the same months, and how big is the gap?

Decision it informs. Whether Experiment 011’s unresolved verdict can be lifted by changing the instrument, and whether the 7.7 pp/yr survivorship-plus-backfill-plus-cost haircut this repository applies to the vendor series is the right size.

Out of scope. Whether to hold a trend sleeve — that is decision 0004, and nothing here changes it. Which fund to buy — that is the recommendation.

as of 2026-08-16. Experiment 012, exploratory, result unresolved.

---

Conclusion

1. The vendor series did not overstate what fund investors earned. It understated it. Over the 78 months both series exist, AQR’s TSMOM index returned +1.95%/yr excess of cash at a Sharpe of 0.141; an equal-weight index of 46 live managed-futures funds, net of their own fees, returned +2.84%/yr at a Sharpe of 0.329. Regressing the vendor series on the live one gives an alpha of −1.36 pp/yr (HAC t = −0.39).
2. The 7.7 pp/yr haircut is outside the interval on the only window that can measure it. At matched volatility the vendor series earned −2.62 pp/yr relative to the funds, 95% interval [−10.91, +5.68]. A +7.7 gap sits above that interval. This does not refute the bound where it matters — the bound is applied to 1985–2011, where no fund return exists — but on the only live evidence available it points the other way.
3. The overlay result is still unresolved, and the reason has moved. The matched-volatility gap for a 50% live-trend overlay against the leverage-matched control is +1.27 pp/yr [−2.62, +4.95] against an MDE₈₀ of 4.76 pp/yr. Experiment 011 was unresolved because of its series; Experiment 012 is unresolved because of its window. Replacing the instrument did not settle it and cannot, at 78 months.
4. The break-even haircut collapses from 9.57 pp/yr to 2.54. That is a fact about 2019–2025, in which trend earned a quarter of its long-run mean, not about the leg’s construction.
5. Attrition is severe and every figure is a lower bound. Of 25 funds filing at the window’s open, 13 (52%) stopped filing inside 6.5 years, including a $1.78bn fund.

---

1. The census: what is in it, and what can never be

Screened over the union of the 2019Q4 and 2025Q4 SEC N-PORT structured data sets — 14,742 fund series — on series names only, with no return read while screening. The mandate and exclusion patterns are Experiment 008’s, reused verbatim because they were frozen before any fund return was downloaded.

Unlike Experiment 008 there is no exchange-listing test, no asset floor, no expense-ratio cap and no inception cutoff. Experiment 008 audited a shelf you could buy; this measures what the mandate delivered, so screening down to the survivors of a size test would reintroduce the selection the exercise exists to remove. Mutual funds, tiny funds and dead funds are all in.

Two survivorship holes, and the second is worse

• Public N-PORT filings begin in 2019. A managed-futures fund that closed before 2019Q4 is invisible to both censuses.
• A fund that both launched after 2019Q4 and closed before 2025Q4 is in neither census, so it is missing from the index entirely. Unlike the first hole this one deletes funds from inside the window. Closing it needs the intermediate quarterly censuses, which this experiment does not read; that is the cheapest single improvement available.

Both holes remove funds that failed, so both flatter every return below. A live index that is too high makes the vendor-less-live difference too low — the direction that would hide the gap this experiment went looking for.

Two further limits on the count itself. A diversified futures programme whose name carries none of the mandate tokens is not in the census at all. And the frozen exclusion pattern has one known defect, recorded rather than fixed: it removes the Credit Suisse Managed Futures Strategy Fund ($343m) because the sponsor’s name contains “credit”. Editing the pattern after seeing what it caught would destroy the provenance that makes reusing it worthwhile.

Attrition, 2019-07 to 2025-12 — a lower bound, twice over

The five largest to stop: ASG Managed Futures Strategy Fund ($1.78bn, last filed 2023-03), Equinox IPM Systematic Macro ($316m, 2019-12), CTIVP–AQR Managed Futures ($266m, 2020-03), JNL/AQR Managed Futures ($134m, 2021-03), 361 Managed Futures ($131m, 2021-01). Only three of the thirteen carry an isFinalFiling flag, so “stopped filing” is the observable and “liquidated” is an inference.

---

2. The live index

Equal weight across the funds that filed a return for that month. A fund that died contributes until it died; a fund that launched contributes from launch; nothing is backfilled. That is the property a CTA peer-group index does not have. Between 18 and 33 funds report in any month of the window.

Annualised, excess of the Ken French one-month bill, 78 months 2019-07…2025-12:

The index is not a diversification artefact. The median individual fund with at least 36 months returned +2.82%/yr at a Sharpe of 0.280 — essentially the index’s mean, at a higher volatility. The index’s advantage is lower volatility, not a higher return.

The equity correlation of −0.109 is the mechanism the sleeve is supposed to supply, and it held: over the 2022 equity drawdown, 2022-01…2022-09, US equity compounded −25.3% while the live net index compounded +26.4% (the vendor series, +34.2%).

Robustness, on subsets defined by the census and never by a return

The second arm matters most: AQR authors the comparator series and also runs five of the funds in the index, so the headline comparison is not fully independent. Removing them cuts the live Sharpe from 0.329 to 0.230 — still above the vendor’s 0.141 — and shrinks the matched-volatility difference from −2.62 to −1.23 pp/yr. It does not change the sign.

The last arm is a caution rather than a result: two funds (Virtus Rampart Multi-Asset Trend, Virtus AlphaSimplex Managed Futures Strategy) show cross-share-class return dispersion of 69 and 41 pp/yr, against a median of 1.10 across the 28 multi-class funds. That is almost certainly a filing artefact rather than 40 percentage points of fee. They are kept in, because dropping them makes the live index better and the conservative choice is the one that does not.

---

3. The decisive comparison: AQR TSMOM regressed on the live funds

AQR_t = alpha + beta · live_t + e_t, 78 monthly observations, Newey–West with 3 lags.

Read the interval, not the point estimate. The instrument cannot distinguish −2.6 from zero. What it can do is exclude +7.7: the repository’s standing CTA bias bound sits above the upper end of both intervals. On these 78 months, a 7.7 pp/yr haircut to the vendor series is not a conservatism, it is a misstatement.

Three things this does not license.

• It is not a measurement of the pre-2019 period, which is where the haircut does its work. The vendor series earned +16.09%/yr at a Sharpe of 1.343 over 1985–2011, +4.17% at 0.315 over 2012–2025, and +1.95% at 0.141 over this window. The 7.7 pp/yr bound is applied to the long-window mean, and no fund return exists to test it there. This experiment measures the gap only where the gap is smallest.
• The 7.7 pp/yr figure was never a claim about AQR’s index against funds. It bounds survivorship and backfill in hedge-fund CTA databases, whose fee load and reporting regime are different. Transferring it to a registered fund is the error Experiment 008 was written to correct, and this result is consistent with that correction rather than a surprise against it.
• It does not clear the vendor series. AQR still states no cost basis, still reconstructs its full history on every update, and still shows a 12.11 pp/yr pre- to post-publication decay. What has been ruled out is one specific explanation — that the series flatters relative to what funds delivered — on one specific window.

---

4. The overlay, re-run with the live leg

Same base portfolios, same weights, same leverage-matched control and same matched-volatility statistic as Experiment 011. The trend leg carries a zero modelled fee, and that is arithmetic rather than optimism: Item B.5 has already deducted the fund’s fee, its trading costs, its slippage and its roll. Experiment 011 charged 1.45%/yr to a series that had never paid one. The borrow spread of 0.59% is charged to the levered controls on exactly the same terms as to the overlay.

78 months, 2019-07…2025-12, net:

The two benchmarks answer different questions and are never combined.

This is unresolved, and it could not have been anything else. The gap is a quarter of the smallest effect 78 months can detect at 80% power. Experiment 011 had 485 months and an MDE₈₀ of 2.82 pp/yr; six years of N-PORT gives 4.76. Trading a contaminated instrument for a clean but blunt one is the whole trade this experiment made, and the trade did not produce an answer.

The break-even haircut on the live leg is 2.54 pp/yr against the levered control and 2.14 against the unlevered, against Experiment 011’s 9.57 and 9.16. That is not a statement about the leg: the sweep is linear in the sleeve’s mean, and the live window’s mean is 2.84% where the 1985–2025 vendor mean is 12.07%.

---

Verified, assumed, open

Verified. The census counts, the attrition of the opening cohort, the index’s moments, the regression and both differences, all from Experiment 012 against hash-pinned French and AQR files and the two committed N-PORT censuses. The overlay arithmetic is Experiment 011’s code, imported rather than reimplemented, so the two results are produced by the same functions.

Assumed. That Item B.5 means the same thing across 46 filers. It is unaudited, and Form N-PORT General Instruction G lets each filer use its own internal methodology. The repository’s cross-source check returned an HTTP error for all 44 US, all 25 ex-US and all 109 corrected-frame tickers, so Item B.5 is the sole measurement of every fund return here and no independent corroboration exists (evidence base). Also assumed: that the equal-weight mean across a fund’s share classes is the right representative, and that EDGAR’s series-filtered filing feed reaches back to a fund’s first N-PORT filing — it does not always, and DBMF is the case in hand: it launched 2019-05 and its feed begins 2021-07.

Open.

1. The pre-2019 gap is unmeasured and unmeasurable from this source. It is where the 7.7 pp/yr haircut is actually applied, and nothing here reaches it.
2. The within-window survivorship hole is closable by reading the intermediate quarterly N-PORT censuses. That is the cheapest improvement available and it would tighten every attrition figure on this page.
3. The index is equal-weight across funds, not across dollars. It is what the average managed-futures fund delivered, not what the average managed-futures dollar earned. N-PORT gives net assets at two dates only.
4. Pretax everywhere, and managed futures are the worst case for that omission.

---

Consequence for this repository

1. The 7.7 pp/yr CTA bias haircut is a scenario, not a bias estimate for this series. It is measured on hedge-fund CTA databases, and wherever it is quoted against the AQR TSMOM series it carries this measurement beside it: on the only window where the comparison is possible, the vendor-less-live difference is −2.62 pp/yr with a 95% interval of [−10.91, +5.68].
2. Experiment 011’s unresolved verdict stands, but its stated reason is superseded. The verdict was attributed to the vendor series being unmeasurable. It is measurable, and over the measurable window it is not the problem. The binding constraint is the length of any window on which live fund data exists.
3. A trend sleeve remains a risk-reduction claim, not a return claim. The live evidence supports the correlation (−0.109, and +26.4% through a −25.3% equity drawdown) and cannot resolve the mean.
4. No sleeve is promoted and decision 0004 stands. Nothing here is confirmatory: the specification was written after Experiment 011’s result was known, and that cannot be undone by re-running it.