Four independent sleeves, each running the same disciplined pipeline once a day — from raw market data to a position you can audit. This is the route every signal takes before it ever moves a dollar.
Everything that's been built, at the system level — four trading sleeves, a market-regime gauge, and the Smart Money tracker, all running on one shared platform. Outside data flows in at the top; positions, alerts, and the website fan out at the bottom.
Full-history replay → risk gate → position state, once a day per sleeve.
Four-pillar market-weather score. Context only — never wired to a trade.
SEC filings ingested, scored, and surfaced across six views.
Breadth, sector indices, trend-flip radar, MA detector. Research only.
A weekly column drafted by AI from the sleeves’ own telemetry. It reports on the system, never feeds it.
Once per scheduled time each day, every sleeve runs the same six-stage pipeline end to end. No human pushes a button; nothing is decided by feel. The strategy logic re-runs against the full price history, the risk gate has to clear, and only then is a position allowed to change.
Each sleeve walks this path on its own — its own data, its own gate, its own saved state. They never compare notes.
Fetch the latest daily closes — and, for some sleeves, an auxiliary non-price input — from multiple sources with automatic fallback when one is slow or wrong.
Prices are cached and cross-checked, so a single disagreeing or down feed gets caught instead of trusted, and the cache rebuilds itself if the host wiped it.
The full strategy re-runs against the entire price history, not just today’s bar, so the decision carries years of context rather than a snapshot.
Before any position is permitted, an explicit risk filter — trend, volatility, rates, or market structure, depending on the sleeve — has to pass, and if the environment isn’t safe the sleeve stays in cash.
The target position is compared to the saved state; if it changed, a phone alert goes out and the new state is written to a version-controlled file, and if nothing changed, the system stays quiet.
Every run pings a health monitor, so a silent failure — a cron that didn’t fire, a host that went down — gets noticed instead of going dark.
A missing or untrustworthy feed never produces a guess. It pushes the sleeve toward cash or skips the trade — the system fails toward safety, not toward invention.
Alongside the four sleeves sit read-only tools built on the same free public data. Not one of them touches a trading decision — they exist for context and research, which is exactly why there is nothing in them to overfit.
A daily 0–10 “market weather” score from four capped pillars — trend, volatility, credit spreads, and macro/breadth — built from free public market and FRED data. It is read-only context: no sleeve trades off it, so there is nothing to overfit. State is persisted and an alert fires only when the regime changes.
A separate service that ingests free SEC filings — 13F holdings, N-PORT marks and Form 4 insider transactions — for a roster of concentrated managers, scores their highest-conviction moves, and surfaces them across six views. Filings lag ~45 days and cover longs only, so it is a research lens, not a live signal.
A family of scheduled scanners kept apart from the trading logic: market breadth, US sector indices, a monthly trend-flip radar across sectors, commodities and countries, and a moving-average detector. They describe what the market is doing. No sleeve reads any of them.
A weekly column drafted by a language model from the sleeves’ own telemetry — what each one did that week and why — then reviewed and edited before it goes out. It reports on the system; it has no input into it. Being straight about the workflow: a draft left unreviewed by the deadline publishes anyway, so treat it as commentary, not as a vetted research note.
A systematic strategy is only as honest as its worst data day. The harder the system tries to keep trading through bad inputs, the more likely it is to act on a number it shouldn't trust — so quantsleeve is built to do the opposite, and lean toward inaction when something looks wrong.
Prices come from more than one feed, cross-checked against each other, so no single provider can quietly steer a decision.
When data is missing or untrustworthy, the sleeve moves toward cash or skips the trade rather than fabricating a signal.
Each sleeve has its own feed and pipeline, so one source failing on one sleeve never blocks or corrupts the other three.
A strategy publishes signals in the open long before any capital goes behind it — the track record has to hold up on live, unedited signals first.
Automation without supervision is just a faster way to be wrong unnoticed. The system runs itself daily, but it's instrumented to flag its own silence — and reviewed on a regular cadence by critics who don't take its results on faith.
Each sleeve computes its signal on its own scheduled cron, every trading day, with no manual trigger.
Every run reports in to a health-monitoring service, so a missed run or a downed host surfaces immediately. A run reporting success is not proof of a healthy system, though — a job can complete and still advance nothing. So a separate daily watchdog checks the data itself: every tracked source must have moved on, judged in trading sessions rather than elapsed hours, and it alerts when one has gone stale.
Each strategy is audited monthly by independent AI critics acting as CMTs, hedge-fund PMs, and risk managers, with the reports kept in the project repo.
Everything on this page is the architecture — the shape of each sleeve, the categories of filter it applies, and the order the pipeline runs them in. What stays private is the calibration: the exact windows, thresholds, ceilings, triggers, cooldowns, and position sizes that make each strategy work. The methodology is open for audit and the backtests are fully interactive, but the numbers that turn the shape into an edge are not published.
Want to see the engine run? Open the interactive backtests and change the window, slippage, and position size yourself.