Dataurora
← Dataurora

How it works

The engine behind every strategy.

One pipeline turns the US and European equity markets into a short list of vetted picks. It runs in five layers: data, math, live-web research, a vote, and a signal that decides entry and exit. Every strategy on this site runs on the same engine.

Five layers, one pick.

A candidate enters at the top as a row in the market and leaves at the bottom as a position in the book, or it is dropped along the way. Each layer adds a different kind of scrutiny.

  1. Layer 1 · Data

    The whole market, as raw material

    Six kinds of data on every covered US and European company, pulled daily and stored exactly as it was known on the day. We own the substrate every signal is built on, not a feed of someone else's picks.

    • US + European equities
    • Prices · fundamentals · insider · analyst · news · earnings
    • Daily history since 2022
    • Point-in-time

  2. Layer 2 · Quant engine

    Math ranks the candidates

    Momentum and quality screens score the universe and rank it. Where a strategy has a walk-forward, out-of-sample backtest, it is validated before live capital; newer strategies publish theirs as the history accumulates. The warehouse of clean history is the part a competitor can't copy overnight.

    • 12-1 momentum
    • Quality screens
    • z-scored
    • Walk-forward · out-of-sample
    • No look-ahead

  3. Layer 3 · AI research

    Live-web research on every name

    Each surviving candidate gets a fresh research brief built from live web search by Perplexity, not stale model memory. It catches what the math can't see: an SEC probe, an acquisition, a broken supply chain.

    • Perplexity · grounded
    • Five-section brief
    • Per candidate
    • Live web, not training data

  4. Layer 4 · Verdict council

    Three AI models vote

    Gemini, Claude, and OpenAI each read the same brief and vote PROCEED, CAUTION, or SKIP. Two or more SKIPs veto the name. For live strategies the gate is fail-closed: no clear verdict, no trade. Every vote and reason is logged.

    • Gemini · Claude · OpenAI
    • PROCEED / CAUTION / SKIP
    • Two SKIPs veto
    • Live = fail-closed
    • Full audit trail

  5. Layer 5 · Signal & hold

    A signal decides entry — and exit

    Each pick carries one of five signals, Strong Buy to Strong Sell, with a strength coefficient from 1 to 10. The signal composes the quant rank with the AI verdict asymmetrically: the AI can make a signal more cautious, never stronger, and Strong Buy requires both the rank and the AI's confirmation. A bought name is then held until its signal says exit, with a dead band against flip-flopping — which is why the book does not mirror each day's fresh ranking.

    • Strong Buy → Strong Sell
    • Coefficient 1–10
    • AI demotes, never promotes
    • Hold-until-exit
    • Dead band

  6. A held position

    Bought on a real broker paper account, marked daily, held until its signal says exit, and shown to you in full: the signal, the rank, the factor values, the research brief, and all three verdicts.

One engine. Many strategies.

Each strategy is a product on top of the same pipeline. Adding one is a new set of frozen rules, not a new platform, so the cost of the next strategy is small and the rigor is identical.

A new strategy joins as a forward paper track with its rules frozen up front. Its walk-forward backtest is published once enough point-in-time history has accumulated — around September 2026 for the current strategies — and any rule change is pre-registered before it ships.

Why you can trust the numbers.

A track record is only worth as much as its bookkeeping. Six design choices make this one auditable instead of marketing.

See the track record →
IsolationPaper and live never touch
Paper and live run on separate broker accounts in separate environments. A live order can't reach the paper book: the broker confirms the connected account before any order is placed.
ReconciliationEvery pick reconciles end to end
Pick, decision, order, fill, position: one audited chain, reconciled every day against the broker's actual positions. The headline return and the equity curve are computed from the same executed fills.
BoundaryThe site is read-only
The website reads a small set of public database views through a restricted login. What crosses the wire is percentage returns, book composition as percentages and share counts, per-pick signals and research verdicts, plus public market reference data (prices, market caps) — never dollar balances, notionals, or NAV. The broker connection stays off the site.
Single sourceOne curve, one drawdown
The max drawdown you see is derived from the same published equity curve you are looking at — there is no separate marketing figure. That curve is built with the exact NAV methodology of the internal risk monitor — dividends credited on pay date, commissions included — verified to reconcile to 0.00 basis points.
Pre-registrationRules are frozen, changes are pre-registered
Factor formulas and weights are frozen. Changes to them are pre-registered in a dated, version-stamped document and judged on out-of-sample walk-forward results — never tuned quietly after the fact. Product-mechanics screens (like the investability ceiling) ship on documented priors instead — in the open, never quietly. Every strategy carries a pre-declared kill criterion; one strategy (dividend capture) was retired in June 2026 when it failed walk-forward validation.
AuditAdversarially audited from the inside
A July 2026 internal adversarial audit found zero critical money-safety issues and confirmed the paper/live isolation in both code and production. It is not a clean bill of health: the audit also logged non-critical findings, which are being worked down.

Deterministic brakes.

Around the signal engine sits a set of hard risk rules with fixed thresholds. They describe what the code does, not what we hope happens.

Drawdown ladder
Measured from the book's high-water mark: at −8% every new buy is halved; at −12% the book enters a defensive hold — positions are kept, not liquidated, and buying resumes automatically once the drawdown recovers; at −20% trading halts and only a manual restart resumes it. Beneath the ladder, −25% of committed capital is the catastrophic floor — a full stop that cannot be manually resumed while the book remains below it; in practice the tranche is retired.
Per-name stops
Each position carries a hard stop — −25% on the aggressive track, −15% on the conservative — or a 200-day-moving-average trend stop, whichever triggers first.
Sector cap
No new entry into a sector already at 30% or more of the book. Held names are not force-trimmed; the concentration decays through normal rotation.
Regime filter
When a market's benchmark index trades below its own 200-day average, new entries are cut to half size.
Earnings blackout
No entry into a name right before it reports earnings.
Churn & concentration
A daily turnover cap bounds how much new buying can enter the book in one day — protective exits are never blocked by it. An aggregate cap blocks new buying of any name already at 12% or more of a currency bloc's combined book value, summed across that bloc's books — held names are not force-trimmed, so a winner can drift above it.

Every brake above is deterministic code. The AI layer can veto a pick; it cannot override a brake.

See the engine at work, pick by pick.

Every strategy publishes its rules, its live paper track, and the reasoning behind each name — with walk-forward backtests published as they clear out-of-sample validation. Read it, then decide.