It counts; it doesn't forecast. Every close, Abacus strips the market out of each name, scores what's left from 0 to 100, reads which of four tapes we're in, and lists only the trade shapes that have paid in that tape before — each with its replayed number, its tested exit, and a size. Then it marks its own book.
Abacus evolves in named generations: each one keeps everything the last one proved and adds only what beat it in replay. Archimedes is the first, live since the launch on June 10, 2026, and the one behind every number on the desk today. The name is for the man who measured a crown by the water it pushed aside — which is the engine's whole move: take the market out of a stock and weigh what is left.
Named for the Bernoullis, who first wrote down that the value of a gain depends on who holds it and that a long enough run of trials tells the truth. Bernoulli builds on Archimedes, not over it — the same rules, plus whatever answers two questions the Archimedes replay raised. Today's pair is built on distance — the leader furthest ahead of the laggard — and the replay says the laggard leg is what fails in an up-tape; in study is a pair built on timing instead, which name is early in its move and which is late. And a run is listed today on its recent strength alone; in study is whether asking for the longer trend behind it keeps the losers off the list. Both replayed to 2008, both scored on the desk's own book after costs. What survives ships as Bernoulli; what fails goes on the timeline as a finding set aside.
Generations are evolutionary, never a restart: a new name means a rule was added or changed because the replay and the book said so — never a coat of paint. Each is named for someone who counted value before it was a discipline — all ages, women and men.
Every trade on the desk carries one tested number. This is how it is made, what the desk's own book has done with it, and how the engine changes — including what was tested and set aside.
Most of a stock's daily move is the market. The engine subtracts it and scores what is left, 0 to 100, as Alpha. A trade is a shape drawn from that score — a fresh strength run, a weakness run against the index, a leader-vs-laggard pair — with an entry, a hedge, an exit and a size already decided. The number beside it is the average outcome of that shape, in the current tape, at its exit, replayed walk-forward over 2019 → 2026 with the tape read at each entry and no knowledge of what came after. Beside it: how often it finished up, how long it was held, and the worst one-in-twenty. A shape whose number is negative in the current tape is not offered — it is here, in red, not on the desk. In a dislocation the desk lists nothing: the plain long's number there is the index's own rebound, not the name's.
| shape | expression | Trendingtoday | Flat | Grinding bear | Dislocation |
|---|---|---|---|---|---|
| Strength run | long · plain | +0.6%41% · 9d · worst -7.4% | +0.6%43% · 8d · worst -7.5% | +0.9%48% · 8d · worst -8.2% | +0.5%47% · 7d · worst -11.2% |
| Strength run | long · vs the index | +0.4%36% · 9d · worst -6.4% | +0.1%35% · 8d · worst -6.5% | +0.2%37% · 8d · worst -6.6% | -0.6%33% · 7d · worst -7.7% |
| Strength run | long · vs its sector | +0.3%38% · 9d · worst -5.9% | +0.1%37% · 8d · worst -5.9% | +0.2%39% · 8d · worst -5.8% | -0.6%32% · 7d · worst -7.1% |
| Weakness run | short · vs the index | +0.1%36% · 8d · worst -6.3% | +0.1%37% · 8d · worst -6.2% | -0.4%32% · 7d · worst -6.3% | -0.7%32% · 6d · worst -8.8% |
| Weakness run | short · plain | -0.0%34% · 8d · worst -7.2% | -0.5%33% · 8d · worst -7.8% | -0.9%29% · 7d · worst -9% | -1.3%30% · 6d · worst -12.6% |
| Leader vs laggard | pair · wide gap | -0.2%40% · 16d · worst -15.8% | +0.2%43% · 16d · worst -15.1% | -0.4%37% · 16d · worst -14.8% | +0.1%46% · 16d · worst -15.7% |
| Leader vs laggard | pair · mid gap | -0.2%41% · 15d · worst -13.5% | +0.2%42% · 16d · worst -13.3% | +0.5%45% · 16d · worst -12.6% | -0.0%39% · 15d · worst -13.6% |
| Leader vs laggard | pair · narrow gap | -0.5%40% · 15d · worst -14.4% | +0.3%43% · 16d · worst -10.9% | not enough history | not enough history |
Exits: Alpha closes under 60, or day 21 (longs) · Alpha closes over 40, or day 21 (shorts) · three Narrowing days in a row, or day 21 (pairs). Pair buckets: gap: wide ≥55 · mid 42–54 · narrow <42.
A run is a temporary state — most end within two weeks. Runs the engine's checks agree on last about twice as long, in every tape. The clock on each trade is drawn from these curves. It is a statement about time, not about price.
checks agree all runs most agree checks split
A rule must be positive over the whole window, not in a chosen slice.
Positive in a clear majority of years, and no single year outweighs the good years together.
The desk's own book, run with the change after costs, must be better for it.
Findings that fail are logged too — the failures are half the record.
Alpha · The tape · Exits by state · True peers · Tested numbers · The house book
The list and the house book hold at most three long runs from one sector; the rest fall to the next shape in rank. Replayed 2019 to today on the desk's rules with the book as the objective: Sharpe 0.46 → 0.60, worst drawdown −16.9% → −13.2%, 2022 −6.7% → −2.5%, return up. A rotation is usually a sector rotation.
Effect The book stops eating a sector rotation whole. Per-trade numbers are unchanged; this is portfolio construction, not scoring.
Abacus — Alpha · Breadth · Attributed · Covariance · Utility · Score — is introduced as the engine behind every number on the desk, with Archimedes as its first generation and Bernoulli named while in study. Nothing in the rules changed; the name records what was already running.
Effect Every future change to the rules lands under a generation name on this page, so the shading tells you which rules were in force.
The desk's front door is now a short list of fully specified trades the engine has qualified — entry, hedge, exit, size — each with one tested number: the average outcome of that shape, in the current tape, at its exit, replayed over 2019 to today. Shapes whose number is negative in the current tape are not listed. The house book starts: every qualified trade, taken by the rules, after costs.
Effect The desk's rules, run as one 20-slot book over 2019–2026 after costs: about 1.0 Sharpe with runs unhedged, about 0.6 market-neutral.
Nine measures of the market's condition at entry (dispersion, rotation, leadership breadth, a leaders-minus-laggards factor, sector dispersion, the index's own return) were tested against the outcome of every leader-vs-laggard pair since 2019. None separates the years the pair paid from the years it lost. A trailing-year number predicts the next month slightly backwards.
Decision Set aside: the pair's number is an era property. It is quoted pooled, never by cycle, and the leader-vs-laggard pair is not offered in an up-tape.
The same fresh-leader entry, hedged four ways: with nothing, with the index, with its sector, with a laggard stock. Hedging with the index or the sector keeps the run's number positive in every tape but a dislocation; hedging with a laggard turns it negative in an up-tape, because laggards get bought in rotations.
Decision Runs are quoted plain with the index or sector hedge as a switch; weak names are offered only against the index; the laggard pair returns in flat and grinding tapes, where selling the weak name pays.
Shorter caps, two Narrowing days instead of three, hard stops on the spread, trailing give-backs, profit takes, and an exit on the legs' own Alpha were all replayed on the same entries. Every one lands within a fifth of a percent of the shipped rule. Exits change the shape of the outcome — the leg exit halves the worst case at a third of the holding time — not the mean.
Decision Exits unchanged. The leg exit is a candidate for grinding tapes once a third such tape has been seen.
Runs the checks agree on last about twice as long as runs they split on — in every tape. Their return is better in up and flat tapes and worse in the one grinding year on record.
Decision Agreement conditions the run clock on the trade detail. It does not move the tested number, the rank, or the size.
A stretched bullish read in a grinding or dislocated tape has its conviction trimmed and the ticket says why. Dislocation now shows whether the index is still falling or rebounding off its low, because the two halves tested differently.
Effect Removes the one cohort that lost in both 2020 and 2022 without touching entries elsewhere.
Four market states read from the index alone — Trending, Flat, Grinding bear, Dislocation — settled once a day, with a posture line for each. Twenty years of replay chose the four and their persistence rules.
Effect Every tested number the desk quotes is now conditioned on the tape at entry.
A pair that wobbles a point past its threshold stays on the board inside a grace band instead of churning off and on. Retained pairs went on to outperform fresh pair-days in testing.
Effect Board churn fell from most of the board each day to a handful; exits became real events.
Pairs leave on three Narrowing days in a row; directional runs leave when Alpha closes back through 60. Both replaced fixed holding periods after a study of every exit rule the desk could express.
Effect Kept most of the payoff in less than half the time with about half the worst case. Still the exits in use.
Peers are found by how names move together after the market's influence is stripped out, not by sector label. Some cross sector lines.
Effect Tighter hedges — and, later, the finding that a tighter hedge is not a stronger thesis.
Relative-strength scoring across stocks, sectors and the market: every name's move with the market taken out, scored 0–100 as Alpha, with paired long/short ideas built from the gaps. The rules Archimedes still runs on.
Effect The first generation begins; everything above was added to it, nothing replaced.
Open the deskOperator's ManualThe WorkshopChangelogDaily Pulse