Opening the research
Opening the research
Public record
Every strategy version, every result and every product change, dated. Failures included.
v009 hedged altcoin-perp positions against Bitcoin with a Kalman-filtered hedge ratio. Its hedge tracked so tightly that the spread it trades never deviated enough to enter: zero trades across 168,000 model fits, and the walk-forward failed its pre-registered gate. The follow-up changed the one parameter the v009 verdict named: with a slower hedge, v010 did trade, and lost −0.51% per trade across 118 unseen trades (33% win rate), with only 1 of 14 pairs profitable on its own. Both count as failed under the record's rules and stay public. Nothing was retuned to look better, and no holdout period was spent.
A floating assistant on every page, and a page of its own at /ask, lets anyone question the live record in plain language. It reads the same files the dashboard and the signal log are built from, as they are at the moment you ask: collector health, every live strategy state, published signals, the market snapshot, the rule hashes frozen before testing, and what we have posted publicly. It has no write path at all, so it cannot trade, edit a file or change a strategy. When a question is about our data it answers from what it read and lists it; it never invents a Crosswake figure, date or position. Anything else it answers from general knowledge, without pretending the record shows it. It is rate limited, and it reports paper results as paper.
Half the book follows each coin’s own trend over 20, 60 and 120 days, long or short. The other half buys the strongest third of 37 futures over the last month and shorts the weakest third, rebalanced weekly. These rules were already studied on all past data elsewhere, so no clean backtest is left: v008 is judged on its live paper record only, from 18 October 2026. The bar is set now: after 90 days, a Sharpe above 0.5, both halves positive, a drawdown under 25%, and better risk-adjusted returns than holding the same coins. It sends no signals yet.
Between 8 and 10 October the channel received 10 short alerts from v004, a strategy this record already lists as failed. That should not have happened. Its live paper result agrees with the test: 8 closed trades, 2 winners, −0.84% average per trade after costs. v003 and v004 now run on paper for observation only and send nothing, the same as v007.
v005 and v006 sent their first live paper signal, an ATOM entry at 2.066 after the 10 October daily close. It was about 6.7% under water later that day. That is ordinary for this strategy: in testing roughly two trades in three lost, and the winners paid for them. The trade stays open until the frozen exit rule closes it, and the result will be published either way.
Research now reads Hyperliquid’s public order book, so depth and market impact are measured in-house instead of listed as facts we lack. That is a cross-venue proxy for spot liquidity, and it is recorded as one. The buying side of the same pipeline, which decides whether a missing fact is worth paying for, is documented and runs in simulation only: no real money has moved and no on-chain payment has been made.
Published v005’s portfolio view on unseen data: equity curve, drawdown, capital deployed, benchmarks against holding BTC and the altcoins, and every trade as a CSV. Every trade was also re-priced at fills 1 and 4 hours late; the result barely moved.
Signals now arrive as short cards with entry, exit discipline and reason. Market updates on ten major coins post after every four-hour close, with facts computed by software and checked before posting.
Intraday momentum on BTC, ETH and SOL futures was negative on data it had not seen. It is kept on paper for observation only and sends no signals.
The v005 rules applied to a wider universe of 36 coins, including newer listings. Judged on its live record only, alongside v005.
243 trades from January 2025 to September 2026: +1.5% average per trade after costs, but a confidence interval that includes zero. Recorded as inconclusive and moved to live paper tracking, with its record counted from 1 October 2026.
Shorting altcoins that held up while Bitcoin fell lost money after costs. Failed.
Hour-scale catch-up trades on lagging altcoins lost money or traded too rarely. Failed.
Delayed, cost-aware entries produced no tradable signals once real costs were counted.
Altcoins react to Bitcoin within seconds, far too fast for a person to act on. The research moved to slower timeframes.