Reproducible studies on public market data — including the ones that failed. Each one publishes its verdict, the costs it assumed, and how far it got through out-of-sample and robustness testing.
Weekly crypto momentum survived fees and funding, gained 106% in 2025, then reversed to -43% in the untouched 2026 test. Volatility scaling did not save it.
A 16-altcoin BTC-led portfolio achieved gross Sharpe 10.73—but turnover costs were still 26.6 times larger than its alpha.
BTC-crash → ETH/SOL cross-asset signal: Sharpe 0.96 on 96 events, WFE 0.47, costs eat 62% of edge. WATCH verdict. 69-day corpus, no sealed OOS.
BTC statistically led SOL at one minute and the gross Sharpe reached 5.75. The catch: execution costs were 327 times larger than the edge.
One of 276 crypto hour-pairs looked statistically compelling. Then realistic costs and chronological validation made it vanish across BTC, ETH and SOL.
Four honest negative results from selective ML, order-book, macro and on-chain research—and the reusable boundaries each failure added to the knowledge base.
We preregistered the trapped-long sequence (OI expansion, aggressive buying, failed breakout, OI contraction) and tried to kill it. It died. Hyperliquid has no OI history; this is a Binance 5m negative result.
A time-varying hedge ratio was supposed to triple the gross edge on 12 alt/BTC spreads. It reduced spread variance by -24.6% (i.e. raised it) on 11 of 12 pairs, cut gross edge to +2.40%/yr against a 9.36%/yr bar, and broke even at 8.14 bps against the 8.81 bps it pays.
PCA on 12 alt/BTC spreads shows the portfolio already had 9.5 effective bets daily, not 2.3. Raising it to 11.5 moved returns from +0.70%/yr to +0.59%/yr, because cost eats 76-79% of gross edge.
The break-even fill rate for passive execution on 12 alt/BTC spreads is 88%, against 14% if fills were a random draw from the signal set. Adverse selection is worth 74 percentage points of required fill rate, and the trades a passive order misses are worth twice the ones it gets.