Register of hypotheses
Every hypothesis this programme has tested, the bar it had to clear, and the verdict it received. Each bar was fixed in a dated commit before the scoring run — a threshold chosen after seeing the result is not a threshold, it is a description.
Dead hypotheses are published with their measurements — they are dead, so the numbers cost nothing. The one surviving strategy carries its verdict only: it is being traded forward, so its construction and its figures stay closed. Counts are computed from the entries below, not maintained separately.
Fade extension from session VWAP, measured in ATR units
The original signal. Gross alpha inside the live entry band measured roughly 30× below the cheapest cost hurdle, and the sign inverted beyond five minutes. A synthetic control showed the same entry losing money even on a perfectly mean-reverting tape — the anchor enters an excursion too early. Retiring it is what prompted a pre-registered slate instead of one metric at a time.
Ride a large extension rather than fading it
The inverse of vwap_ext, and the direction that study had accidentally suggested was stronger. It did not survive its own registration: nothing in the round-1 slate approached the bar.
Deep order-flow skew continues in the direction of the skew
Flow level as a continuation signal. Sub-hurdle. Together with H7 this closes the question of whether order-flow level leads price at these horizons: it does not.
Extension and order flow pointing the same way
The best of the round-1 slate at +2.23 bps at thirty minutes — and positive on only 4 of 10 days, far short of both the level bar and the consistency requirement.
Fade an extension when order flow opposes it
Standout of the five-day preview at 1.5–1.9× the hurdle, short of the pre-committed 2× bar. Five further days were banked without being scored, and a rehearsal on them was explicitly declined. They inverted the result: the highest-conviction bucket went from best to worst. Small frequent wins, catastrophic losses exactly where the signal was most confident.
Opening-range breakout continues
Sub-hurdle on the full preview and not re-run in round 2. It also taught the de-clustering lesson: a persistent state re-sampled every thirty seconds produces six figures of "events" that are one event, so event counts were treated as autocorrelated from then on.
Trend continuation on a relative-volume spike
Scored zero events — because the relative-volume input was an unwired stub frozen at a constant, not because the market said no. Recorded as VOID rather than FAIL, and the distinction became a standing rule: a miscalibrated instrument is not a market verdict, and it earns exactly one threshold re-registration.
The same idea, once relative volume was actually computed
Instrument fixed, threshold deliberately unchanged so the re-run tested the hypothesis rather than a new one. Sub-hurdle.
Resting order-book pressure as a directional signal
The most consistent effect measured anywhere in the programme — positive 10 days of 10 at the short horizons, rising conviction buckets, 64,000 events — and roughly a fifth the size of a retail round trip. A real microstructure fact, observed for nothing, and untradeable at these costs.
Acceleration in order flow, rather than its level
Twelve events in ten days. The registered threshold turned out to sit near the 99.99th percentile of the real distribution, so the registration voided itself under the event-count floor rather than producing a verdict on noise.
The same idea, re-registered at a calibrated threshold
The one permitted re-registration, spent on the most selective threshold that still cleared the event floor: 1,860 events across 11 days. Indistinguishable from noise, with a negative top-conviction bucket. Final for the flow-acceleration family.
The H3 mechanism against a rolling anchor instead of session VWAP
The closest intraday miss. Changing the anchor turned the mechanism from negative to positive with 8/10-day consistency at every horizon — vindicating the criticism of the laggy anchor — and reached 0.97× of the single-cost hurdle, about half the promotion gate.
Twelve-month momentum, skipping the most recent month
Fired on roughly 45% of in-universe days against a registered 30% degeneracy cap — a metric that fires on half the market is not selecting anything. Voided by its own registration before a verdict could be claimed.
Top-conviction momentum at a measured threshold
The largest event-alpha ever measured here — roughly 3× the promotion bar over 66,574 events and ten years, with monotone conviction buckets — and 52.8% of quarters positive against a pre-committed 60%. The registration had already written the verdict: a metric that made all its money in one regime is a regime bet.
Proximity to the 52-week extreme
Cleared the level bar on the clean instrument and ran 18 quarters of 36 — a coin flip. Failed on consistency. Its first reading also showed inverted conviction buckets; that inversion was contamination, and was retracted.
Buying a sharp weekly move against a longer trend
Far below the hurdle. Notable for what was withdrawn: the first run showed a catastrophic negative, which turned out to be unadjusted corporate actions making ordinary splits look like one-week crashes. The metric was buying fake dips that never reverted. On a clean instrument it is merely flat.
Expansion in realised volatility as an entry trigger
Reached 0.99× of the bar with 19 of 39 quarters positive and a negative top-conviction bucket. Close, in the way that does not count when the bar was fixed in advance.
The momentum signal as a monthly top-K book, not a per-event metric
Failed all four legs at both capitals over 118 rebalances, and lost to its own costless equal-weight benchmark. Pooled per-event alpha is not harvestable return: a long-only book forfeits the signed metric’s short side, top-K selection reaches past the best bucket into winner’s-curse territory, and month-long holds are not 66,000 overlapping windows.
Low realised volatility as a standalone factor
Did exactly what defensive factors do — about half the benchmark’s drawdown — and lagged a roaring equal-weight benchmark badly on every consistency leg. Calm, and beaten by buy-everything.
Low market beta as a standalone factor
Same shape as F1, same outcome. Both were later retested as hedge sleeves rather than standalone strategies, and failed that too.
Momentum with a long-term trend filter for regime management
Failed every leg at both capitals. The filter did what the textbooks promise on the one crash it was designed for, then whipsawed away two other years: it trimmed the tail and the return together. Momentum is now closed at three levels of construction — per event, as a portfolio, and regime-managed.
A defensive sleeve blended with the one surviving strategy
The combined book was genuinely good and still failed, which is the interesting part: its risk-adjusted return came in below the incumbent alone. The incumbent’s drawdown was already lower than its would-be hedge’s, so the sleeve could only dilute. A hedge earns its place only when drawdown is the binding constraint.
Long-term reversal (construction withheld — live strategy)
Cleared all four legs of the registered gate at both capitals, and later cleared a second, stricter risk-adjusted gate built after the fact. Qualified, not celebrated: its traded sample is dominated by a single unusually favourable regime for buying beaten-down names, and it is the construction most exposed to delisting assumptions. Now in forward paper against a protocol registered before the first observation; first scheduled verdict around February 2027.