Every signal below is graded through the same gauntlet — non-overlapping t, Newey-West HAC, Deflated Sharpe, FDR. We publish what survives and label what doesn’t. Snapshot 2026-08-21.
Signals · research program · honest tiers
4 signals. Each plotted on its instrument, with the windows we’d have called.
We run a signal-research loop over our data. What survives sorts into honest tiers: one directional signal clears the defensible bar (livestock seasonal), one is on the watchlist (the feeder/live cattle spread), two are conditional-volatilityrisk flags (BTC & Monero on-chain), and one is a macro-liquidity regimeconditioner (Fed balance sheet & reverse-repo vs. SPY/BTC). Below, each is drawn on its instrument — the shaded windows are exactly when the signal fires. The volatility and regime signals forecast move size and backdrop, not direction, so they carry no equity curve; we label each as what it is, not dressed up.
Project scorecard · updated 2026-08-22 05:37 UTC · the loop works the weakest first
Building the business — every dimension, one loop
75.4%overall
Next up: Signals proven — livestock_le_dec(DSR 0.840, net Sh 1.04); gfle_spread(DSR 0.841, net Sh 0.99); hele_spread(DSR 0.930, net Sh 0.83) [* = clears net Sh>=1.0 AND DSR>=0.95]
Signals proven0 / 3 defensible
livestock_le_dec(DSR 0.840, net Sh 1.04); gfle_spread(DSR 0.841, net Sh 0.99); hele_spread(DSR 0.930, net Sh 0.83) [* = clears net Sh>=1.0 AND DSR>=0.95]
CNBC sector ETFs + top congress/insider stocks (no key); Tiingo would add full universe
Data product3 / 3 eval packages
hospital/RE/shipping sellable packages (G1)
Bulk delivery3 / 3 parquet verticals
S3/parquet drops for funds (G2)
The research loop · loop #1342 · updated 2026-08-22 05:29 UTC
Every 20 minutes, we grade our data through the gauntlet
5.8%feeds coverage
961,745hypotheses tested
707series graded
1candidates
95forward-track obs
Metrics we’re chasing
Defensible signals1 / 3
boc_canada_yields.value:mean [z20] -> LQD (watch) ic_is=0.2055 ic_oos=0.0782 forward_ic=-0.2487 n=25 WARNING: forward IC sign OPPOSITE to in-sample -- disconfirmed live
Best candidate OOS Sharpe0.0782 / 1
961745 ic_tests -> E[max|t|] under the null = 4.87 (honest bar ~5.9); best observed |haircut_t| = 4.13; dsr>=0.95 passers = 0 -- best result is BELOW what pure search produces
Feeds graded707 / 1,718
every real feed table, at least once
Defensible = net Sharpe>=1.0 @25bps OOS + DSR>=0.95 + forward-confirmed. We publish only what survives — the number that matters is candidates, and it moves only when the data earns it.
WATCHLIST
Live Cattle — December seasonal
CME Live Cattle (LE)
1.45×strategy equity
The signal: Go LONG live cattle for the month of December, flat the rest of the year.
● short ● long — each entry we'd have signaled.
Compounded return of following the signal — starts at 1.00×.
1.052sharpe
2.74t hac
0.8402dsr
1.044net sharpe
0.2413pbo
10years
FAILS the defensible bar: DSR 0.8402 < 0.95verdict
Live from benchmarks.py: Sharpe 1.052, Newey-West HAC t=2.74, DSR=0.8402, net-of-cost Sharpe 1.044, PBO 0.2413. Bar is net Sharpe>=1.0 and DSR>=0.95 -- NOT met.
WATCHLIST
Feeder/Live Cattle Spread — mean reversion
GF/LE ratio (feeder vs live cattle)
35.10×strategy equity
The signal: When the feeder/live cattle ratio dislocates >1 sigma from its 252-day mean, trade the spread back toward the mean (~5-day hold).
● short ● long — each entry we'd have signaled.
Compounded return of following the signal — starts at 1.00×.
1.0-1.4oos sharpe
-2.77hac t
yes (HAC-t -4.57)corn controlled
all 3regimes
SURVIVED adversarial verificationverdict
GF/LE ratio (feeder cattle / live cattle). OOS Sharpe ~1.0-1.4 net of cost, NOT corn-in-disguise (survives controlling for corn), mean-reverting in all 3 regimes, HAC-robust. Current z ~+0.9 (in-band). Watchlist - forward-confirming live before any claim of defensible.
The signal: When BTC on-chain activity (active-addresses 60d z) is elevated (z>1), next-day BTC moves are ~1.5x larger. A RISK/sizing signal: widen stops, cut leverage, hedge tails — NOT a directional call.
Shaded = dislocations the signal flagged (mean-reversion setups).
On-chain activity forecasts next-day BTC realized vol, INCREMENTAL to BTC own trailing vol (HAC-t 7.48, corr -0.11). Elevated-activity days: next-day return spread ~+58% wider. A volatility/risk signal for position-sizing + tail-hedge, not directional alpha. Live in the monitor.
The signal: When XMR transaction count (60d z) is elevated or equities are in risk-off, next-day XMR moves are larger — cut size, widen stops, hedge. NOT a directional call.
Shaded = dislocations the signal flagged (mean-reversion setups).
3.13hac t
yesincremental to trailing vol
+22-25% vol in SPY selloffs / VIX spikesrisk off amplifier
P(-8% day) doubles when z>2 (4.5%->9.1%)crash prob
VALIDATED conditional-volverdict
XMR direction is BTC/ETH beta (not tradeable); its VOLATILITY is its own and forecastable from tx-count, amplified by equity risk-off. Tail-risk quantified: crash probability doubles at extreme activity. Live in the monitor.
How to read the tiers
Defensible = clears Newey-West HAC t and the Deflated-Sharpe bar against the full multiple-testing ledger. This is the signal we stand behind. Watchlist= a real effect with positive out-of-sample return, but below the defensible bar (parameter-sensitive or thin) — we forward-track it live and promote only if it holds up. We would rather show you one honest signal than ten we can’t defend.