Platform
NovaQ takes market data in, turns it into decisions and carries those decisions through to execution, with a record of each step.
NOVAQ Ecosystem
What the platform does
| Capability | What it gives you | Result |
|---|---|---|
| Live market intelligence | Bias, regime, levels and session state, calculated as the market moves | Session statistics available during trading, always with sample size and confidence interval |
| Controlled execution | A signal passes safety checks and becomes a managed position | Signals pass a safety gate and budget checks before an order goes out |
| Reproducible research | Backtest and live logic share one definition | 87,260 trades reproduced with zero value mismatches on pinned data |
| Full reconstruction | Any decision can be followed from signal to fill | Correlation identifiers across the chain, a tamper-evident audit log and input replay |
| Fault tolerance | One failing service does not stop the rest | Supervised restarts, crash-safe journalling and an independent emergency record |
How it is built
Nine supervised service groups share one database, and each group has one job.
| Service group | Role |
|---|---|
| API | Web interface and gateway |
| Core | Scheduled analytics, data quality and maintenance |
| Data | Market data intake, gap repair, contract rolls and calendars |
| Features | Live bias, regime, levels and state |
| Exec | Signal arming, safety gate, position management and reconciliation |
| Trader | Hosts the AI trading agent |
| Neural | Hosts neural model inference |
| Shadow projection and observer | Record what happened and what was not executed, with no trading rights |
The operator works through 42 web pages covering markets, strategies, risk, research and system health. Behind them sit about 72,000 lines of platform Python, about 20,000 lines of C# for the trading platform and about 4,900 automated test functions.
The first generation ran as a single process, which was fast for research but limiting for live work. Each live incident went into a lessons register, and that register became the specification for NovaQ.
Data infrastructure
What it enables. Results are only as good as the bars beneath them. NOVAQ's data layer makes contract identity a checked fact.
Results.
- A rebuilt continuous corpus cut false jumps above 150 points from 6,541 to 108 while keeping 99.9% of about 5.5 million NQ bars.
- An audit of 341,992 files across 41 data versions found zero mixed-contract bars among about 11 million live minute bars.
- Three separate implementations reproduced the audit, and a synthetic mixing test caught 5 of 5 planted cases.
Coverage. NQ and ES minute data from 2010, with live minute and second data through NinjaTrader 8. The registry also covers GC, CL and micro contracts.
Principle. Prices are not back-adjusted or smoothed. Gaps between contracts are real market events and stay in the data.
Clean continuous corpus: false jumps and affected features
DR-LENS
What it does. DR-LENS is NOVAQ's session analytics engine. It translates the published defining-range method of TheMas7er into computable form and annotates about 1.1 million five-minute bars per instrument, about 4,100 sessions each, with more than 400 columns. We have extended the original DR-LENS with dozens of new analytical capabilities and parameters.
What it enables. It answers questions such as how often a given session state has historically led to a given outcome, with a confidence interval. Research can test a structural idea across 16 years without new data work, and the live system can show session statistics during trading.
A concrete result. A widely quoted session frequency reproduced at 80.7% on ES across 2,148 sessions, within one point of the published figure. The findings it produced are on the Research page.
From bars to counted session structure
Market state and regimes
- Regime classification places the market into states such as trend, rotation and expansion, with a stated level of confidence. It is rule-based, not a language model.
- Conditional probabilities always carry a confidence interval and sample size.
- Calibration compares what the system predicted with what happened.
- Conviction tracking updates the evidence for each strategy as results arrive. Hard safety gates are never softened by conviction.
Internal cycle speed
In October 2026 the median internal execution cycle fell from 16.5 to 2.77 seconds outside the windows that carry heavy provider work. One frequent query dropped from about 2.9 seconds to a few milliseconds, and session analytics from about 115 to 23 seconds per instrument. Each step was checked for identical outputs. This measures the platform's own processing, not order execution at the broker.
Median internal cycle, October 2026 (seconds)
From research to execution
- A hypothesis is recorded.
- It is tested with realistic costs and point-in-time data.
- A survivor receives a stable identifier in the registry.
- It runs in shadow or simulation with the same logic as the backtest.
- Parity checks compare live behaviour with research.
- Each deployment to an account is a typed, audited command.