Sixteen strategy ideas, one per lens, went through fourteen judged matches. The champion, the runner-up, and the best eliminated idea then went to testing. Every number on this page comes from live public sources pulled 10–12 July 2026; one result was revised after a source-data fault, and the correction sits in the Kaub section.
Every match ran the same way. Both sides attack, both defend, one rebuttal each, and an independent judge scores it: creativity 30%, credibility of edge 30%, implementability 20%, capacity & risk 20%. The final wasn't delegated — the desk head judged it personally. Scores are weighted totals out of 100.
← the bracket scrolls sideways →
Lost its quarter-final while being judged more original than the idea that beat it: creativity 82 against 68. Every objection was about sizing (thin cif-Rhine swap depth, capital idling between episodes), not about whether the effect exists. It went to testing as the dark horse; results below.
Champion, runner-up, dark horse. For each: find the strategy's observable footprint in free public data, test the cheapest falsifiable claim first, and write the kill criteria before looking at results. The verdict stamps reflect the data.
When the Rhine falls below loading thresholds at the Kaub gauge, barges carry less fuel upriver and inland German prices detach from the seaboard. The trade: buy the inland premium when the hydrology says a squeeze is coming.
Eight years of daily Rhine levels against the Germany-minus-Netherlands heating-oil spread from the EU Weekly Oil Bulletin, ex-tax. We used the Lobith gauge (2017–2024, zero gaps) because Kaub itself only publishes 31 days of history. Nine low-water episodes showed up below the 10th-percentile threshold.
The inland spread averaged +124.0 EUR/1000L in episode weeks against +78.2 in month-matched baseline weeks (Welch t = 3.45). The 2018 episode peaked at +265.6. 2022 peaked at +170.9. Water level leads the spread by about three weeks, and the effect only bites below the 10th percentile — which is exactly the episodic design the strategy claims.
The first pass reported a +91.9 uplift at t = 4.41. A check of the chart's tail found a fault: after the Netherlands stopped reporting in late 2022, the source repeated its last price for seven weeks, then printed a wrong flat 1902.0 for six more. Those thirteen weeks sat in the baseline. With them removed, the uplift is +45.8 at t = 3.45.
National consumer prices dilute the Rhine-corridor barge differential, so the uplift is a lower bound; the tradable version needs paid Argus/Platts cif-Rhine assessments. Serial correlation inflates weekly t-stats; the effective sample is two major episodes and six minor ones, once each. 2022 is confounded by the EU energy crisis. And measured as a ratio instead of a spread, 2022 weakens while 2018 strengthens to 1.42×. Both views are in the data.
No widening across at least 4 historical episodes, or the premium fully priced within 2 days of the gauge crossing. Either one kills it.
Kaub has been under 150cm for a month (minimum 63cm on 5 Jul 2026). An episode is underway now.
River level above, inland premium below
Weekly, 2017–Dec 2022 · red bands = low-water episodes (< p10 = 723cm) · tail cut where the NL series went stale
| Episode | Days | Min level |
|---|
Sources: Rijkswaterstaat DD-API 2.0 · EU Weekly Oil Bulletin (NL series ends Dec 2022 after stale-data cut)
Insurers hedging roughly $190B a year of buffered annuities must sell 1-year S&P wings at whatever price they find. The pitch says that flow has compressed the smile's convexity: buy the cheapened wings, sell the body, and collect.
The full live SPX chain: 29,364 options at the 10 Jul 2026 close, spot 7575. We reduced it to the smile term structure (ATM vol, 25-delta wings, skew and butterfly at six tenors) and checked the pitch's two cheapest falsifiable claims: 1y skew at 3–4 vols, and a visible pin at the 1y point.
The 1y 25-delta put skew printed 4.54 vols. That's above the pitched 3–4 range and below the pre-2020 norm of 5–6. The butterfly term structure runs smooth through 12M (1.16, 1.13, 1.10 across 9/12/18M) — no dip at the tenor the flow supposedly pins. Front convexity is modestly richer (fly/ATM 7.5% at 6M against 6.8% at 12M), which weakly supports the roll-down story but is generic to equity smiles. Today's entry discount is thinner than the pitch claimed.
One delayed-quote day can't test the historical compression claim; either the smile partially re-steepened after the 2023–24 sell-side notes, or the pitch overstated it. Annuity FLEX hedges cluster at month-start anniversary strikes, so a FLEX-specific discount could hide off the listed surface entirely. And the snapshot is from mid-July, outside the quarter-start issuance window.
Four quarters of daily snapshots with no widening of the 12M−6M fly gap in issuance windows. Or 1y skew persistently at 5 vols or more — nothing left to harvest.
A daily snapshot job is live, building the history free data doesn't have. The full test needs CBOE DataShop (about $1–5k) or OptionMetrics ($25–50k/yr), plus OCC FLEX open interest and LIMRA sales for attribution.
SPX smile term structure — snapshot 10 Jul 2026
Vol points by tenor · shaded bands mark the pitch's claimed range and the pre-2020 norm
| Tenor | ATM | 25Δ put | Skew | Fly |
|---|
Source: CBOE delayed quotes, full chain · vendor IVs, delta-space interpolation, parity-checked forward
Cancelling a scheduled 10b5-1 insider sale is the cheapest legal way to act on good news. Since 2023 it must be disclosed, but only in 10-Q footnotes that Form 4-keyed quant pipelines can't see. The pitch: buy the cancellations, ride the drift.
Every 10-Q and 10-K since the rule took effect: 22,249 filings matching termination language. A classification pass read each one, and 94% turned out to be negative boilerplate ("no officer or director terminated..."). What's left is 981 real terminations — almost exactly the ~300 a year the pitch estimated. We priced 662 of them, market-adjusted against SPY from the filing date.
The unconditional signal is flat. Mean 21-day CAR: +0.34% (t = 0.50), hit rate 45.7%, median −0.89%. The 63-day mean looks alive at +1.6%, but the median is −4.65% and the trimmed mean goes negative; that tail was warrants and lottery names, not drift. So the tournament final's central attack — that nobody has shown drift after disclosure — now has pilot data behind it.
First pass showed +5.6% mean at 63 days. It was three SPAC warrants the CIK-to-ticker map had picked instead of common stock (+1219% on one). One mapping fix later, the effect vanished.
If the strict-filter version (CEO/CFO only, unsold shares remaining, no subsequent Form 4 sales) is also flat at 150+ clean events, the idea is dead, not just diluted.
The filters are the strategy's entire remaining case. Next fidelity layer: XBRL ECD-tag parsing, an officer-rank filter, and Form 4 ingestion for the no-subsequent-sale screen.
Distribution of 21-day abnormal returns — 659 events
Market-adjusted CAR vs SPY, first close after filing · 2.5%-wide bins, tails clipped at ±40%
| Horizon | N | Mean | Median | Hit rate | t |
|---|
Sources: SEC EDGAR full-text search + filings · Nasdaq daily closes · Apr 2023 – Jun 2026
The funnel
Everything above reruns from a local repo: pure Python, no dependencies, one SQLite database per strategy, every stage resumable, raw responses cached for audit. The tests update on a schedule.
| Pipeline | Cadence | What accumulates | Paid upgrade when justified |
|---|---|---|---|
| rila | daily, post-close | its own 1y-surface history; no vendor sells the backfill | CBOE DataShop $1–5k · OptionMetrics $25–50k/yr |
| dog | quarterly, after the 10-Q wave | fresh terminations and CARs; the next layer adds Form 4 screens | — (EDGAR is free; the filters are engineering) |
| kaub | daily levels; weekly fuel | live episode tracking (one's running now) | Argus/Platts cif-Rhine $15–30k/yr |