USDA publishes daily boxed-beef cutout prices and 5-area cash cattle prices — the packer’s realized margin between what beef sells for and what the cattle cost, before anyone else’s model has seen it. The tradeable instrument graded here is CME Live Cattle futures (LE=F). Three candidate features were built and graded on identical terms; the winner — cutout_mom5, the 5-day log-momentum of the choice cutout — is the only one with a sign-consistent train/test relationship: a mean-reversion signal (strong cutout momentum tends to precede negative forward LE=F returns). It is weak — near transaction cost — and does not clear a tradeable bar. Every number below, including the two rejected candidates, is reported as graded, and the fair test runs forward on /track-record.
Three features were built from the same USDA exhaust before any grading happened: packer_margin_d5 (5d change of choice cutout minus 5-area steer cash), chsel_d5 (5d change of the choice/select cutout spread), and cutout_mom5(5d log-change of choice cutout alone). Selection of the primary feature was by maximum mean |train IC| across k=1…5 — the test set was never touched for selection, only for confirmation after the fact.
Every feature is lagged 1 trading day (USDA daily reports finalize after futures close), winsorized at 5 prior-250-day standard deviations (strictly-prior window, no lookahead), and graded by Spearman IC on a 70/30 chronological split. cme_settle carries LE and HE settlement data on the box but only 40–48 rows spanning ~5 recent days — too short to grade, documented rather than used.
packer_margin_d5 is effectively noise (train IC flips between roughly +0.01 and −0.01 across lags; test IC drifts to −0.01 to −0.03 with no consistent direction). chsel_d5 flips sign outright — positive in-sample, negative out-of-sample by a wide margin at every lag — the textbook shape of an unstable, non-generalizing fit. Both are rejected and shown in full below, not omitted.
Solid black: daily packer margin ($/cwt), downsampled to 420points across the full 2016–2026 LE=F futures history for legibility. The grey envelope is the ±1.96σ accuracy wave-bound: the prior-250-trading-day mean margin ± 1.96× the prior-250-day standard deviation, computed strictly from data available before each day — no lookahead, which is why the band is absent for the earliest ~1 trading year. Colored ticks mark the five largest cutout 5-day-change spikes (|z|>2 vs. a strictly-prior 250-day distribution, 5-day cooldown between events) by magnitude: blue = up-spike, red = down-spike.
Solid black = CME Live Cattle futures close (LE=F), the tradeable instrument. Dashed blue = USDA choice boxed-beef cutout. Both indexed to 100 at the start of the sample so the two series — priced in different units ($/cwt live cattle vs. $/cwt wholesale beef) — can be compared on one axis. They move together over the long run, which is exactly why the cutout is a candidate leading indicator: the question graded below is whether its short-horizonmomentum leads LE=F, not whether the two are correlated in levels (they trivially are — cattle cost is the largest input to beef price).
cutout_mom5is the primary feature — the only one of the three that keeps the same sign in-sample and out-of-sample. It does not clear a tradeable bar on its own: |IC| stays under 0.07 even at k=5, and no cell here has been checked against a formal significance threshold with multiple-testing correction. It is reported because the sign held, not because the magnitude is exploitable yet.
| Lag (days) | n (full) | IC (full) | t (full) | IC (train) | IC (test) |
|---|---|---|---|---|---|
| 1 | 2,507 | -0.0098 | -0.49 | -0.0019 | -0.0353 |
| 2 | 2,506 | -0.0138 | -0.69 | -0.0046 | -0.0417 |
| 3 | 2,505 | -0.0292 | -1.46 | -0.0213 | -0.0526 |
| 4 | 2,504 | -0.0503 | -2.52 | -0.0432 | -0.0681 |
| 5 | 2,503 | -0.0511 | -2.56 | -0.0442 | -0.0687 |
t = IC · √((n−2)/(1−IC²)), computed on the full-sample IC and n. No parameter search was performed on the primary feature after selection — k=1…5 are the only lags ever graded.
| Feature | Lag | IC (full) | IC (train) | IC (test) |
|---|---|---|---|---|
| packer_margin_d5 (rejected — noise) | 1 | 0.0002 | 0.0098 | -0.0269 |
| 2 | -0.0042 | 0.0017 | -0.0197 | |
| 3 | -0.0131 | -0.0069 | -0.0320 | |
| 4 | -0.0161 | -0.0140 | -0.0256 | |
| 5 | -0.0058 | -0.0067 | -0.0091 | |
| chsel_d5 (rejected — sign flips train→test) | 1 | -0.0132 | 0.0079 | -0.0614 |
| 2 | -0.0094 | 0.0217 | -0.0785 | |
| 3 | -0.0183 | 0.0159 | -0.0979 | |
| 4 | -0.0249 | 0.0062 | -0.0974 | |
| 5 | -0.0235 | 0.0012 | -0.0807 |
Sort every day into deciles by that day’s cutout_mom5 reading, then average the forward LE=F return in each bucket. At k=1 the top−bottom spread is a flat, non-monotonic -4.4 bps. At k=5 the headline spread is -4.0 bps, but that number is not a clean decile-1-through-10 ramp — deciles 1–4 (weakest momentum) average solidly positive forward returns while the top-momentum deciles are mixed, not uniformly negative. Direction matches the negative IC above; the shape is not monotonic enough to trade the decile spread on its own.
Instead of the daily lead-lag IC, condition on the event: what does LE=F do in the days after an abnormally large 5-day move in the choice cutout fires, using a strictly-prior 250-day z-score and a 5-day cooldown so events don’t double-count the same move?
| Event type | n | +1d bps (hit%) | +2d bps (hit%) | +3d bps (hit%) | +4d bps (hit%) | +5d bps (hit%) |
|---|---|---|---|---|---|---|
| Up-spike | 30 | 42.7 (57%) | 32.4 (53%) | 29.8 (63%) | 23.9 (60%) | 16.3 (57%) |
| Down-spike | 20 | -28.9 (50%) | -58.4 (50%) | -74.8 (50%) | -39.7 (55%) | -3.0 (60%) |
Sample-size note: n = 30 up / 20 down over the full LE=F history — small samples, no multiple-testing correction, and hit rates near 50–60% are not, by themselves, statistically distinguishable from noise at these counts. Shown because the direction is consistent with the mean-reversion story (down-spikes precede negative forward LE=F moves at +2d/+3d; up-spikes precede positive ones through +4d), not because it is offerable.
The data lead is real; the measured edge is too small to trade.The cutout genuinely prints before models see it, and the sign of the relationship held from train to test at every lag — that is why this survives as research at all. But the magnitude is the problem: the best cell is a test IC of −0.069 at k=5, and the decile top-minus-bottom spread is roughly −4 bps per event — the same order as a tick or two of slippage in LE (one tick = $0.025/cwt = $10 per 40,000-lb contract). After friction, the edge as-measured rounds to zero. No cell here survives a formal significance threshold with multiple-testing correction, and two of the three candidate features were rejected outright. Nothing on this page is proven alpha.
The fair test is forward: this exact rule (fade 5-day cutout momentum in LE=F) is paper-trading live as PT-002-CUTOUT-LC on /track-record, marked against real CME settlement closes (@LC.1 continuous front-month via CNBC — 500 sessions on record, Jul 2024 through Jul 2026 and updating daily, never Yahoo). If the forward ledger surprises us, it graduates; the expected outcome, stated in advance, is that it retires.
USDA daily reports finalize after futures close -> every feature LAGGED 1 trading day; winsorize at 5 rolling prior-250d sd (shifted); rolling stats use strictly-prior data only
The cutout-momentum fade is paper-trading live as PT-002-CUTOUT-LCon the forward ledger — entries, marks, and retirement reasons all published, nothing deleted. That ledger, not this backtest, decides whether this becomes a signal. See /track-record for the live ledger, or /signal-001for the nuclear-fleet nowcast — the same treatment applied to a stronger data lead.