Research Under Controls.

From idea to verdict in five steps, each on published research

The Nexus overview shows research under controls as five steps. Here is each one: what Nexus does, the research it rests on, and whether it runs today.

1. 1

   Trial counted

   Before any data is read

   Every trial is booked in an append-only register before the first market-data read. The research pipeline's stages, batch calibration, strategy-builder long tests and promotion each check the booking again, so a run that was never booked, or is larger than its booking, is refused. Bookings made after the fact still count. The number of trials in a family sets the bar in step 5.

   One exception is open. Re-pricing a live book at newly measured costs creates a re-costed copy without a booking. Whether that copy counts as a new trial is awaiting a ruling.

   ResearchTry enough variants and the best one looks good by luck alone: the expected best Sharpe ratio of N strategies with no edge rises with N. Bailey, Borwein, López de Prado and Zhu (2014); Bailey and López de Prado (2014).

   Runs

   At those four entry points. Re-pricing a book at new costs does not book a trial yet.

2. 2

   Run authorised

   Spec, code, data, operator

   A counted run needs a one-shot authorisation that binds the specification's hash, the code commit, the data IDs, a review document and the operator's own sentence. It is root-owned, so no agent can write it.

   WhyThis is governance rather than statistics. It makes the trial counted in step 1 the one that actually ran.

   Runs

   For every counted run except exploratory sweeps.

3. 3

   Backtest runs

   Out of sample, with leakage removed

   The out-of-sample record a calibration reports comes from walk-forward analysis. Parameters are chosen on a training block and tested on the block that follows. The window then rolls forward, and the test blocks are chained into one record. Walk-forward efficiency, the out-of-sample score over the in-sample score, is reported for every run.

   One limit: the parameters a calibration stores for later use are not chosen by this train-then-test loop. They come from a separate rolling ranking that scores fixed configurations on its test blocks, with a correction for the number tried, and the walk-forward search starts from that winner.

   Training rows whose outcome overlaps the test block are purged, and one trading session is embargoed at the boundary. Both are counted in sessions, so weekends and holidays are handled. A fold left too thin is refused, not relaxed.

   ResearchWalk-forward analysis and its efficiency ratio: Pardo (1992, 2008). Purging and embargo: López de Prado (2018), chapter 7.

   Runs

   On every batch calibration, as the reported record; the stored parameters come from the separate ranking. The efficiency gate fails only a negative ratio; Pardo's 0.5 rule of thumb is not a gate.

4. 4

   Hold-out sealed

   Spent once, never refunded

   On the newer research path, a fixed three-year slice, September 2023 to August 2026, is sealed. A read through the governed route needs a break-seal token. The token spends one of two looks shared across the path's campaigns before the first sealed read, with no refund if the run then fails. A read without a token is logged and warned about, not blocked. A sealed result must clear a one-trial bar: a probabilistic Sharpe ratio above 95 % and a minimum Sharpe ratio.

   ResearchProbabilistic Sharpe ratio: Bailey and López de Prado (2012). A hold-out stops being out of sample once a result has been tuned against it, so each look is spent, not borrowed.

   Built

   No look has been spent yet. A read without a token warns rather than refuses. On the trend path the latest stretch of history is held back as a reference score, but it is not sealed.

5. 5

   Verdict by the rule

   Written before the result

   Each campaign writes its accept rule before any result exists, and closes on that rule. The survivor rule for a selected candidate is the exact familywise 5 % test, assuming independent strategies with normally distributed Sharpe ratios. Its out-of-sample Sharpe ratio must clear the 95th percentile of the best of N strategies with no edge, where N is the counted family from step 1, so more trials raise the bar. Three reversion campaigns closed negative on their rules in September 2026, with the sealed years left unread.

   ResearchFamilywise error control: Bonferroni, in Šidák's (1967) exact form. Why a new result needs a t-statistic nearer 3 than 2: Harvey, Liu and Zhu (2016).

   Rule in code

   At the research pipeline's survivor gate. No pipeline run has used it yet.

Guards that run on every change

Some failures do not show up as a bad result. They show up as a good one. These checks run before code reaches production, not after a result looks suspicious.

• Random-walk canary Runs

  Statistically powered tests feed the full search random-walk prices, which have no edge to find, and assert that the best result scores about zero. They run in CI and in the pre-deploy gate, and a non-zero score breaks the build. Look-ahead adds edge that a correction for the number of trials cannot remove, so it needs a guard of its own.

• Look-ahead tests Runs

  Dedicated tests check that no future data reaches a signal, and that delaying a fill manufactures no edge on random prices.

• Bad data stops the run Runs

  A dataset that fails its quality checks or has gone stale is refused at the start of a test and again at promotion. A nightly replay that hits a failed upstream fetch publishes nothing, and the next night retries.

• Sharpe never alone Runs

  In the strategy report every Sharpe ratio sits beside its standard error and the bars and years behind it, so a short record cannot pass as a strong one (Lo 2002).

Costs measured, not assumed

Cost is where most backtests flatter themselves. Nexus charges what the broker charges, refuses to guess, and says where it still approximates.

• Measured per leg

  Books that trade pay each leg's measured broker spread, half on each side, and overnight funding by asset class. A leg with no usable measurement is refused by name and never priced at a default.

• Broker minimums

  The broker's minimum trade size for each market is checked when a book is frozen and again before promotion, so a book that cannot be placed at its design capital does not reach the trader.

• Cost by year and hour

  On the newer research path, cost is resolved by year and by time of day. Older years are estimates: a per-year spread census from a market-data vendor is converted to the broker's terms by a ratio measured today, and a market with no broker measurement of its own is marked as estimated. A year with no census is never given a modern default. The cost tables are pinned by content hash, re-checked on every build, and every test also runs at 1.5 times cost. This path is separate from how trading books and the calibration search are priced.

• Costs at admission

  On the target-position engine that runs panel strategies, a position is opened or added to only if its forecast clears its own uncertainty plus a round-trip cost reserve. Trend books and the calibration search do not apply this hurdle.

• No-trade band

  Small changes in the target position are held inside a band, so the book does not pay to chase noise. Under proportional costs the optimal policy has exactly such a no-trade region (Constantinides 1986; Davis and Norman 1990). Nexus sets the band's width; it does not derive it.

• Break-even

  Each trading book reports, nightly, the multiple of its cost that would erase its edge.

Where it still approximatesTrading books charge today's measured spread across their whole history, and their report flags it. The calibration search charges a flat cost per trade; measured per-leg costs apply to the books that trade. On the newer research path, older years' costs are estimates scaled by today's broker ratio.

Robustness, before and after capital

A result that survives selection is then watched. Every active book is replayed each night and judged by rules written before it started.

• Forward replay. Every active book is replayed nightly at its frozen parameters. Its pre-registered review rules return PASS, KILL, INCONCLUSIVE or PENDING, and a KILL blocks promotion. PENDING, for a book with too few days, is advisory.
• Rolling performance. Rolling 21, 63 and 126-bar Sortino ratio and drawdown, with average drawdown and time under water.
• Bootstrap bands. Block-bootstrap bands at the 5th, 50th and 95th percentiles for Sortino ratio and drawdown, nightly. A diagnostic, not a gate (Künsch 1989).
• Paired comparison. In counted single-change trials, a candidate is scored against its control on the same data with a paired block bootstrap, and accepted only if it is no worse within a margin stated in advance. It is not a general acceptance gate: the broader Sharpe-uplift gate is built but not in use.
• Leave one out. Each sleeve is removed from the book in turn, to show how much of the result depends on it.
• Turnover. Annualised turnover is measured on every run and reported per configuration, per out-of-sample record and per leg.

How a quant plugs in

A strategy is a document of named slots, and your model fills the forecast slot. Bringing a model in is integration work through the strategy builder's extension points. It runs on an engine that already exists, under every control on this page.

1. 1

   Preview

   Score the strategy on the one declared window for its clock. Warm-up history is loaded but never scored, and every look is logged before any data is read.

2. 2

   Compose and save

   Change a stored strategy's settings within versioned, bounded menus, and save it under a name that is never reused.

3. 3

   Long test

   Submit the document. The trial is booked in the register first. The job then queues with an at-most runtime and memory quote, and runs to a report.

For a quantKeep the model in one component and the framework handles the rest: clean data, sizing, costs, counted trials and honest statistics. The two ways in, a forecast for a trend book or for a cross-sectional panel, are in the case study.

What runs, stated plainly

Each control is graded against the code, and the grade is shown here. Naming what is not built yet is part of the method.

• This page describes the framework. It shows no strategy returns and makes no performance claim.
• Grades come from an audit of the code and move as the code does. A grade is raised only when the code behind it is re-read.

Sources, read 2026-10-15: the research service's implemented-methods checklist, refreshed that day by an audit of its code as of 14 October with an adversarial re-check of every change of grade; its method catalogue and verified references; and an audit of the strategy builder for the interface section.

The reading list

Every work cited on this page.