Trend structure
Fast and slow moving averages must agree, price must break a twenty-bar channel, and the four-hour picture must point the same way as the fifteen-minute one.
FINDEUS
Plutus of Finance
FINDEUS trades currencies, crypto and equities from three separate processes against three separate brokers, sizes every position against the account it is actually risking, and declines far more trades than it takes. It also tells you which of its own numbers it does not believe. The restraint is the product.
Invite only. Ask for a code and say what you would use it for.
Read live from the running system through /api/public and /api/health, both of which are allow-lists: configuration and liveness only. No balance, no position and no profit-and-loss is published anywhere on this site.
The book
Read live from each engine as you load this page, not typed here and stamped with the date somebody last edited it. None of it is hidden behind the login, because a page that only reports its machinery when the machinery is busy is a brochure.
Reading the book…
The ratio above is the argument of this whole page expressed as a number. Most of what is refused is refused by one gate: an entry is skipped outright when the quoted spread is more than 15% of ATR, a setting two points from a cliff that turns this strategy from −15% to −96%. That is why it sits where it sits, and why the honest description of this book is one that is refusing rather than one that is winning.
The governor has separately cut the whole book to half risk. It judges the most recent sixty closed trades and offers no verdict at all below forty, so this is the first window in which it could say anything; what it said was that the book's expectancy is distinguishable from zero on the wrong side. It cannot raise risk, only reduce it.
Reading the book…
Long holds were designed in rather than suffered: daily bars, a 6% minimum stop and a minimum dwell mean positions measured in weeks, because a venue charging 1.88% a round trip cannot support anything shorter. Whatever the open positions are currently worth is unrealised, which is to say it is not a measurement of anything.
Reading the book…
The youngest of the three, and the most constrained. The +131R backtest figure quoted further down this page has produced nothing live, and the one thing worth saying about it here is that it is a nineteen-position number and this account is capped at five.
Every figure above is a COUNT read from the running engines through /api/book at —, alongside /api/public and /api/health. All three are allow-lists: counts, configuration and liveness. This site publishes no balance, no position value and no profit-and-loss, so none of these numbers says what any trade earned — and a book that has closed nothing shows a dash rather than a zero, because an empty sample is not a result.
The system
Most of a trading system is the part that says no. FINDEUS runs every candidate through the same chain, records which link stopped it, and shows you the answer.
Refused at any gate → written to the journal with the reason, and the bar is over.
Fast and slow moving averages must agree, price must break a twenty-bar channel, and the four-hour picture must point the same way as the fifteen-minute one.
A directional-strength reading below thirty means the market is ranging. Breakouts in a range are where trend systems bleed, so those are skipped outright.
No new position within two hours before or four hours after a high-impact release. This is the one filter in the whole project that survived a placebo test.
A trade that hands a tenth of what it is trying to win to the spread before being right about anything is refused outright. This is the one part of the Virtus-Fin score that is allowed to say no, and it is allowed because it is a subtraction rather than a statistical claim.
One percent of net asset value per currency trade, a cap on open positions, a daily loss stop, a total drawdown halt, and a currency-exposure cap so several trades cannot quietly become one bet. Each venue carries its own envelope rather than a shared one — seven concurrent positions on currencies, eight on crypto, five on equities — because these are three accounts of different sizes and a cap sized against one of them funds nothing on another. The live currency values are in the panel at the top of this page, read from the running system rather than typed here.
Orders go in fill-or-kill with a slippage bound, and a stop and a target ship with every entry. The trailing stop was switched off in August after it was measured: a trend book's expectancy lives in its right tail, and a stop that ratchets in on ordinary noise cuts exactly the trades that pay for the losers. Turning it off was better in both halves of the sample and is the single largest improvement made to the currency book.
Every signal, block, fill and close is written to a durable journal — a mounted volume on the currency book, versioned remote storage on the other two — so the record survives a restart, and so post-mortems use data rather than memory.
Discipline
Zeus blinded Plutus so that wealth would fall on the deserving and the undeserving alike — impartially, rather than to whoever the gods happened to favour. A trading system should be blind in the same way: to the story, to the conviction, to the trade you badly want to take.
A trading product is easy to make look good. These are the shortcuts FINDEUS does not take, and they are the reason to trust the numbers it does show you.
Every strategy here was tested against a matched null — random filters dropping the same number of trades — and most of them failed. The dashboard reports the t-statistic next to the return, so a good-looking result that is indistinguishable from luck looks exactly like what it is.
FINDEUS reads the macro tape through Claude with live web search, on a daily interval and only when a key and a budget are configured. That view is advisory and stays advisory: it is scored against realised forward returns for at least 120 observations, and only a rank correlation that clears a threshold promotes it to a trading input. It has never been promoted, and the morning briefing email was switched off in August.
Spread is measured as a fraction of the risk on each trade, not assumed away. On the currency book the modelled figure is of the order of two hundredths of one R — about 1.6 pips against a two-ATR stop — and when the deployed configuration was measured over four years of real bid/ask rather than modelled, it was paying a quarter of its target to the spread, which is why that configuration was retired. On Robinhood's crypto book it was measured at 1.88% of position value — the same on Bitcoin as on a meme coin, which is how you know it is a fee wearing a spread's clothing. When the measured cost exceeds the measured edge, the panel says so instead of rounding it into the equity curve.
Nothing that decides whether trading may continue lives on a filesystem that disappears. The currency book keeps its halt flag, journal and risk state on a mounted volume; the other two have no volume, so theirs is written to durable storage with a generation counter that rejects a stale write rather than overwriting it. A stop you set from your phone stays set through a restart, a redeploy and a crash.
Signing in gets you the book. Placing, closing or halting anything asks for a second password every single time, and the server checks your role before the order is queued.
The equities backtest returns +131R. That figure trades nineteen names at once, and a twenty-dollar account holds one — so the same trades were replayed through the cap the account actually has, two hundred random orderings at each cap, because which of four simultaneous breakouts you get is luck. At one concurrent position the median is −2.9R and 157 of the 200 orderings lose. At two it is +1.4R and 87 of them do. At three it is +13.5R and none of them do. That sits on the dashboard above the equity curve, in the same size, because a headline number computed under conditions you do not have is the most expensive kind of true. The live cap is five, raised only after the same walk-forward showed the account could actually fund the extra slots — the third position is paid for by the first two having already worked, which a static read of today's share prices says is impossible.
Markets
The strategies never reference a currency pair — they take a price series and return a signal. Adding a market is mostly plumbing. Deciding whether that market is worth trading is the hard part, and it is the part most systems skip. Each venue runs as its own process against its own broker, so a fault in one cannot reach the others, and a fourth view adds them up — because three books each stopping themselves at their own daily limit can still lose the sum of those limits between them, and no single bot can see that number.
The majors and the crosses between them on fifteen-minute bars over the REST API, with cross-currency P&L conversion so a yen trade is sized in the account's own currency. Nothing streams: a decision is made on a bar that has finished, so a tick-by-tick feed would buy latency this strategy has no use for.
LiveSigned requests, fractional-coin sizing to eight decimals, a resting stop and a 24/7 clock. A round trip here costs 1.88% of position value — the same on Bitcoin as on a meme coin, which is how you know it is a fee wearing a spread's clothing — and that single number decides the whole strategy shape: daily bars, a 6% minimum stop, weeks-long holds. Concurrent positions were measured at +0.70 average correlation, which makes eight of them roughly 1.4 independent bets rather than eight. The dashboard says so rather than counting to eight. The watchlist was widened from fourteen names for a reason that is arithmetic rather than ambition: fourteen left the book empty on a quarter of all days, and with holding time already stretched as far as the evidence supports, the number of names is the only lever left on how often anything happens at all.
LiveThe constraint here is not the one everybody names. This is a cash account, so the Pattern Day Trader rule does not apply — that governs margin. What binds is that Alpaca will not accept a fractional stop order, so a live position must be a whole share, and one share of the cheapest large cap risked 31% of the account. Hence a watchlist of liquid names under twenty dollars. The other thing FX and crypto never face: the exchange is shut 17.5 hours a day, and a stop is an instruction to sell once a price trades, not at it.
LiveVirtus-Fin
Every strategy here answers a yes/no question: did the close clear the channel? A market that ground sideways for forty bars and then jumped three ATR on five times its usual volume, and one that ticked a cent over the line in the middle of chop, are the same event to that question. Virtus-Fin replaces the yes/no with a number — five factors, each observable at the close of the signal bar, minus what the round trip costs in units of the trade's own risk.
Ranking a venue's own trades by the score gave a rank correlation of +0.019 on crypto and −0.052 on equities across 995 trades. Using it to break ties when more signals fire than slots exist beat a random ordering in both halves of the sample in 28 of 54 crypto configurations, and 0 of 27 equity ones. So on crypto and equities it ships switched off: computed on every instrument, recorded against every entry, deciding nothing. On forex it was switched on anyway, in August, on an instruction rather than on evidence — it sizes there and still refuses nothing, and the page that explains it says so in the same words.
Each book fits the five weights on its own closed trades. Before those weights may build the live score, they must order trades they were never fitted on — in both directions across the middle of that book's history. One hold-out is a coin flip: a rank correlation over sixty trades has a standard error near 0.13, so half the time an ordering carrying no information clears a one-sided test.
A trade that pays more than a tenth of its target to the spread is refused. That is a subtraction, not a statistical claim, and it is the only part of the score allowed to say no. Measured over four years of real bid/ask, the configuration this retired was losing 0.19R per trade across twelve thousand of them.
Each venue shows its stage — collecting, fitted, tested-and-rejected, or scoring with its own weights — the distance to the next one, and the two out-of-sample numbers. When it says the weights are not being used, that is the state of the evidence rather than a placeholder.
That last one is a game, and it is the shortest way to see what the measurement above actually means: twelve breakouts, take or skip, with the same five readings the engine sees. It takes about two minutes, and the point of it is that you will not be able to pick the winners either.
Control
Until August this was true of one venue out of three. Stopping the equities book meant opening a shell on the machine it runs on, and clearing a halted crypto book meant the same. That is a poor place to keep the control that decides whether a live account may trade — you reach for it at exactly the moment a terminal is the last thing you want.
Not "is the process up" — whether the next signal would be taken. Each blocking condition is listed in the order the engine checks it, with the control that clears it beside it. A book stopped by the clock is shown differently from one a guard has latched, because an alarm that fires every night is one nobody reads by morning.
Halting blocks new entries and closes nothing — it says so on the button, because an operator who taps it expecting safety is still fully in the market. Flattening sells everything and then halts, since closing the book and re-entering on the next bar was never what the button meant.
This site holds no broker credentials of any kind. A command is written to a queue and collected by the engine in about twenty seconds, so a compromise of the dashboard cannot place an order — it can at worst show you old numbers.
Every button carries the book it acts on, and so does the confirmation. This is not decoration: the keyboard palette used to send its Halt and Flatten to the currency account from every screen in the app, so reaching for it on the equities tab would have flattened forex.
A control surface that silently omits things teaches you the missing capability does not exist. Each venue lists what still lives on the host — the verified flatten that re-reads the book from the broker, the risk-state surgery, the forward-test reset — with the command, and why it is not a button.
Security
There is no open registration. Every account comes from a single-use code the owner mints, valid for fourteen days; you can ask for one, and a person reads the request. Behind the login is a funded brokerage account, and a public signup form on it would publish the balance to anyone who found the URL.
Viewer sees the book and nothing else. Trader can work positions. Owner can change risk, halt the bot and manage the team. The server enforces this on every command — hiding a button is courtesy, not a boundary.
Removing someone takes effect on their next request, not whenever their cookie happens to expire. Passwords are stored as scrypt hashes, and changing yours signs out every other device.
Nothing can push to the trading process. It polls for queued commands, which expire in minutes — so a stale instruction from an hour ago is never acted on, and a compromised dashboard can at worst show you old data.
Ask
There is a robot in the corner of every page on this site. Ask it how the score works, what stops a trade, or what happens when the book starts losing — and tell it how much detail you want. It answers the same question three ways.
“Five grades, each one worth a different amount, all added together. Then take away the cost.” One idea per sentence, and an analogy instead of a term.
What the thing is, why it is built that way, and what it refuses to do. No arithmetic, no assumed vocabulary.
The weights, the cost multiplier per venue, the sample a figure came from. The same numbers the Virtus-Fin page publishes.
It is not a language model. It is a written list of answers with a face: every one of them matches a claim made elsewhere on this site, and the measured lines under them are read from /api/book when you ask. Ask it something nobody wrote down and it says so rather than inventing an answer — which is the same rule the trading engine follows about trades.
Access is handed out one person at a time, by hand. Tell us what you would use it for and you go on the waiting list; there is no automatic approval and nothing here signs you up to anything.