Every strategy this desk has run, including the ones that died.
Odin keeps two ravens: Hugin for thought, Munin for memory. The sagas have him fear more for Munin. This is the memory — one line per book, opened and closed, with the number that closed it. Four of the five are dead.
What the desk believed at open, against what happened.
- Fusing social, Telegram and on-chain signals finds real setups
- A 1–10 conviction score separates good entries from bad
- Tight ROI targets bank winners early and protect the book
- The losses are fees and slippage; net them out and the edge appears
- One cohort printed +1.2%, so the approach works and needs tuning
- No positive slice at n≥5 — by side, source, ticker or hour of entry
- Conviction measured anti-predictive in all three cohorts
- The ladder cut winners at +0.7–1% while stops released losers at 3–6.4%
- A zero-fee, zero-slippage replay still lost about $150
- That +1.2% was two positions held 131.6 hours through an outage, into an SOL rally
The book stays closed absent new structural evidence. A parameter set that backtests better does not count as evidence — that is the mistake the +1.2% headline came from.
One book survived, and it needs no directional view.
Short the Hyperliquid perpetual, hold the spot leg, collect the funding difference. Delta-neutral, so the part that failed — predicting direction — is not involved.
Cost basis: taker fees plus 5 basis points of assumed slippage on every leg, every side, amortized over a 30-day hold. The one order-book walk in our data measured 0.5–2.9bps at this clip size on a calm day, so 5bps sits above what we have seen — and the Binance spot legs have never been measured at all. These figures were 6.4–7.7% before slippage entered the model; that earlier version was an upper bound and is what this page showed until 2026-08-13. LDO at 3.9% now sits below the 4.5% net APR the desk requires before a position is considered for capital, so on its own expectation it no longer qualifies. Paper positions on a single operator's own account, never real capital.
Build it, measure it, close it when it does not pay.
The pipeline runs and the ingestion works. What it produced is what closed the book.
Ingest
Public Crypto Twitter, Telegram channels and on-chain flows land in one stream. This part still runs.
Cluster
Related mentions are grouped and de-duplicated, sources weighted, each cluster scored with its sources kept. The score measured anti-predictive, so it gates nothing now.
Paper-test
Everything trades on paper against live prices, with taker fees and slippage modelled. No real capital, ever.
Judge
A cohort clears its breakeven win rate or it closes. Three did not. A fourth result that looked good turned out to be an outage, and was thrown out too.
There is nothing to buy.
One person's research project. No plans, no accounts, no checkout — the pricing page that used to sit here described a purchase path that was never built.
The dashboard
The desk interface, published as a demo build. Every figure in it is synthetic sample data, behind a permanent banner saying so. No account, no live market data, no positions behind it.
Open the demoThe write-ups
How the directional book was measured and why it closed, and how the funding-carry forward-test is run. Including the results that did not work.
Read the write-upsNo capital is managed for anyone and no real capital has ever been deployed. Every strategy described here was traded on paper. Nothing on this site is financial advice.