Research
NOVAQ's research asks what the market actually does, measures it against strict controls and keeps both what holds and what does not.
Main findings
| Finding | Result | Evidence |
|---|---|---|
| Diversified portfolio | 53 rule-based books: Sharpe about 2.15, return to drawdown about 23, 7 of 7 calendar years positive (2026 partial) | Historical test, 2020–2026 |
| Portfolio diversification | About 12 effective exposures in the book, with mean daily correlation near 0.13 and close to zero on the worst days | Historical test, two methods |
| Session size is forecastable | AUC near 0.7 (0.5 is chance) and positive skill in all 7 years (2026 partial), while direction stayed near a coin flip | Walk-forward, 2020–2026 |
| Early volume reveals the rest of the day | The slowest fifth of days had about half the remaining range of the fastest fifth, in every year tested | Descriptive, 2020–2026 |
| Volatility carries across sessions | A wide earlier session led to a wide opening hour about half the time, against one third by chance | Descriptive, both halves stable |
| Position dominates | Earlier sessions, composite bias and regime labels added no measurable direction once price position was known | Leave-one-year-out |
| Famous statistics reproduce | A well-known session frequency matched within one point, and 15 of 17 public NQ statistics matched within about four | Independent reproduction |
| Forward-to-backtest check | No family with a robust gap on 2,272 paired trades, with a wide interval. A technical check, not evidence of profit | Mostly simulation |
What can be forecast before the open: AUC by target
How research is done
| Stage | Purpose |
|---|---|
| Specification | Rules become computable, and ambiguity is written down |
| Pre-registration | Key studies fix their rules and success criteria before the first run |
| Point-in-time inputs | Each input must be known at the moment of the decision |
| Costs | Commission and slippage are included, with sensitivity tests |
| Multiple-testing control | False discovery correction across all variants |
| Out-of-sample and walk-forward | Parameters are chosen on earlier data and carried forward |
| Robustness | Monte Carlo, bootstrap, placebo and random-path controls |
| Independent review | A separate reviewer looks for reasons the result could be false |
| Forward | Candidates run under a stop rule written in advance |
Research in numbers
An inventory on 3 November 2026 catalogued 698 studies and separate tests across all NOVAQ repositories. Together they tested over 3 million variants and configurations. That total includes reruns and overlapping grids, and one study's count of about 1.55 million is approximate. 106 studies were pre-registered.
The filter does its job: of 75 large studies with a known outcome, 39 left no survivor. That is how a small set of results earns its place.
Autonomous multi-agent research
AI agents let NOVAQ run larger studies with more checks than one person could manage. In one pre-registered study, eight parallel groups logged 3,254 tests across 76 items, with 33 skeptic verdicts and 1,296 additional skeptic tests, all under a global false discovery correction.
The result is breadth with a paper trail: each study keeps its hypotheses, code, results and verdicts together.
Portfolio research
Strategies that look different can be the same bet. In June 2026, correlation analysis of daily results reduced 36 atomic strategies to 18 micro-streams and 12 macro bets. In October 2026, a separate measure on 55 books gave about 12 effective exposures. Twenty scalping variants turned out to be a single momentum exposure, which is why risk budgets are set per bet.
From strategies to independent bets (June 2026)
Results across the range
Research is useful when it can also close a question. Negative results stay in the record because they save future work.
| Study | Question | Result |
|---|---|---|
| Claim validation | How many library claims survive a full gauntlet? | For example, 2 of 217 |
| Auction statistics (joint) | Do they beat a random path? | For example, 2 of 21 measurements differ |
| DR/IDR register | Do high hit rates become profit? | Core frequencies replicate, none of 44 strategies passed |
| Options positioning (joint) | Do levels predict direction? | They help with size, not direction |
| Order-flow indicator | Does a strong-looking signal survive? | None of over 34,000 uses, and one fragile candidate from a wider search |
| Open types (joint) | Does the first hour give a trade? | It describes structure, none of 13 direct setups passed |
| AI Silver Bullet review | Can a public 25,000-configuration study be reproduced? | Reproduced on another engine, net profit not confirmed |
Joint research
Part of the NOVAQ programme is carried out jointly with Vašek (BTE). Joint studies are published in generalised form: the definitions, parameters and rules needed to reproduce them are not disclosed.