Treasury Futures Curve: Carry and Momentum

About this project

A systematic study of carry, momentum and curve strategies on U.S. Treasury futures from January 2010 to July 2026. Continuous, back-adjusted futures are rebuilt from the exchange roll map, curve trades are hedged to be DV01-neutral using the daily cheapest-to-deliver forward risk, and every strategy is evaluated net of tick costs with HAC Diebold-Mariano inference against a duration-momentum baseline.

Why it matters

Time-series momentum on outright duration is well known. The question a rates desk actually faces is whether the same signal works on the curve. It does not: the slope mean-reverts, and applying a trend signal to it is a systematic way to lose money even on a perfectly hedged book. The asymmetry is the finding; the honest part is that the curve trades do not statistically beat the outright baseline.

Methodology

• Six futures roots (TU, FV, TY, UXY, US, WN), first three generics; continuous series rebuilt from FUT_CUR_GEN_TICKER with difference back-adjustment so roll gaps are never booked as returns.
• DV01 from Bloomberg’s daily cheapest-to-deliver forward risk (CONVENTIONAL_CTD_FORWARD_FRSK), the real per-day sensitivity rather than a regression proxy.
• No look-ahead: signals use t−1 data, execution at the t settle, monthly rebalance, 126-day warm-up, leverage capped at 3×, volatility scaled to 10%.
• Costs of one tick per contract per side using each contract’s own tick size (the 10-year tick was halved during the sample).
• Diebold-Mariano with Newey-West long-run variance at a 21-day lag plus the Harvey-Leybourne-Newbold small-sample correction, cross-checked against HAC-robust OLS.

Strongest findings

• On the identical DV01-neutral 2s10s book, momentum loses (net Sharpe −0.79) and mean-reversion earns +0.71; the book’s P&L regression slope on the 10-year yield change is −0.001, so the gap is signal, not hedge error.
• Net of one tick per contract per side, 2010–2026: duration momentum 0.68 (max drawdown −23%), 2s10s mean-reversion 0.71 (−20%), 5s30s mean-reversion 0.66 (−27%), 2s5s10s butterfly 0.07 (−38%).
• Against the baseline, HAC Diebold-Mariano gives DM = 0.83 (p = 0.41) for 2s10s and 0.68 (p = 0.49) for 5s30s: neither curve strategy statistically dominates outright momentum.
• Every number was re-executed from the canonical notebook against the original exports in August 2026; that re-execution exposed a transaction-cost loader bug, corrected on 2026-08-23, and the figures here are from the corrected run.

Robustness and caveats

• The mean-reversion direction for the curve is economically motivated but was confirmed on the full sample; a walk-forward out-of-sample test is still required before treating the curve results as a hard performance claim.
• Single country, single currency; costs are a flat tick-per-side approximation without market impact or financing.
• Inputs are Bloomberg Terminal exports and are not redistributed; the repository ships the exporters, the checklist and a documented free-data route.
• Correction (2026-08-23): the original notebook silently loaded no contract statics, so the table first published as "net" was gross of tick costs (0.71 / 0.77 / 0.73 / 0.12; same-book −0.73 / +0.77; DM p 0.32 / 0.37). The costed numbers are shown here and in the re-typeset paper; no conclusion changed — the ordering holds, both DM p-values move further from significance, and the asymmetry widens slightly.