Signal research on the data catalog, published in full.

A nightly grader sweeps every feed in the fleet: Spearman IC against a mapped asset basket at 1–5 day leads, chronological 70/30 split, winsorized, with every test logged so the multiple-testing exposure is fully disclosed. The first sweep (7,210 tests) produced 61 candidates that passed the in/out-of-sample gates; the test infrastructure was then rebuilt to grade every signal at its real publication date and the full search rerun — 15,035 tests across 296 feeds, none above threshold (detail under “Null results” below). No signal carries “offerable” status: that requires ≥60 live forward observations with confirming sign. Candidates that pass receive multiple-testing adjustments (Harvey–Liu, deflated Sharpe) and conservative per-trade costs before entering the paper book. The live paper track record (/track-record, public, no key) opened 2026-07-14 at zero and accrues only forward. The re-verification that eliminated 26 of 27 lead-lag pairings is documented in full at /leadlag.

Four free-tier feeds, each paired with the instrument it historically relates to and the measured IC_oos / t / n — modest effects, sign-consistent in and out of sample. See the four free signals →

Daily NRC reactor exhaust reconstructs the official EIA nuclear generation print to 1.5% average error, ~14.9weeks before it publishes — and conditioning on unplanned reactor tripsfinds UNG drops an excess −322bps over the following 5 days (t = −3.42 on 64 events, sign opposite the naive story). Published with full caveats (event clustering, roll drift, seasonality) and now paper-trading forward. Chart includes the ±1.96σ wave-bound envelope. Full signal page →

Twenty years of USDA boxed-beef cutout exhaust against live-cattle futures (LE=F): 5-day cutout momentum mean-reverts into forward futures returns — a weak but sign-consistent effect across train and test, reported as graded, with decile analysis and rejected candidate features included. Full signal page →

As of 2026-07-16, every signal is graded at the moment its data was actually published, not the date it covers — the most common source of spurious alt-data results. Under that standard, nothing in the library clears the bar: results that previously appeared significant reflected publication lookahead, with the model implicitly using data before it was available. The test infrastructure was rebuilt around a publication-lag clock and the full search rerun — 15,035 tests across 296 feeds, zero above threshold. Best t-statistic in the ledger is 4.1 against a 5.3 requirement; best false-discovery q-value is 0.49 against the 0.10 requirement. The product is the point-in-time data and the testing infrastructure that produced these results.

Two feeds remain on the board as sub-threshold candidates — emerging-market bond spreads into SPY and Canadian government yields into bond funds — both forward-tracked, with no claims attached. The detail below predates the publication-lag rebuild and is consistent with the same zero-survivors result.

Crypto on-chain × BTC/ETH (2026-07-16): 11 deep feeds (hashrate, MVRV, miner flows, DEX volume, DeFi TVL, fear/greed…), up to a decade of history, 330 tests through the full gate stack (BH FDR q≤0.1 against the whole 11,440-test ledger family, deflated-Sharpe floor). Survivors: 0. Widely cited on-chain indicators post large in-sample t-stats (hashrate t=−5.5) that collapse out of sample (IC_oos≈−0.03); in/out-of-sample sign agreement was 237/540 ≈ 44% — a coin flip.

Rates & FX feeds × TLT/IEF/SHY/LQD/HYG/UUP/FXE/FXY/EMB (2026-07-16): 19 deep feeds (gilt/bund/Canada curves, breakevens, reserves, repo…), 1,230 tests. Survivors: 0.