Research
Research programme
NeoZorK Research studies time series and complex systems in markets and well beyond them, and builds the AI, verification and data infrastructure that such research needs. Its method is simple to state and hard to practise: register the hypothesis and the test first, evaluate out of sample, report failures with the same care as successes, and ship each experiment so that others can rerun it. The work builds on the econophysics and superstatistics literature and on the practice of pre-registered research.
How a claim earns its status
Method
- RegisterState the hypothesis and the test before looking at results.
- FreezeFix protocol, data and hashes (SHA-256) so the experiment cannot drift.
- TestOut of sample, walk-forward and Monte Carlo stress, net of realistic costs.
- ReportPublish the verdict with a status, including failures.
- ReproduceShip the experiment and its exports so a stranger can recompute the number.
Honest applications
What this could be good for
Monitoring across heterogeneous sensors
A threshold that survives a change of source, so that alarms need less per-source calibration. Realistic today for roughly a quarter to a third of source pairs, and with known failures.
Honest model validation
Purged validation, multiple-testing control and a register of refuted claims are reusable on their own, whatever the model.
Risk models with heavy tails
A documented baseline for what a heavy-tailed, clock-driven model does and does not add over standard historical simulation.
A template for evidence
The pre-registration, hash-freeze and verdict discipline, released as open tools, is the part others can use immediately.
None of this is an investment recommendation or a claim of trading profit. Directional prediction of prices was tested repeatedly and not found. The record is single-operator; independent replication is what would change that.
What is not yet known
Open questions
- Does a one-shot heavy-tail estimate beat rolling extreme-value estimation and adaptive conformal baselines?
- Does the regime labelling beat HMM, Markov-switching GARCH and change-point baselines?
- Can short histories of 50 to 250 points be calibrated against a placebo donor, not just against themselves?
- What carries the threshold-transfer effect once more families are sealed and tested blind?
Pillars
Research areas
Time-series science
Regimes, non-stationarity, forecasting limits and honest baselines.
Econophysics
Market dynamics studied with the tools and the scepticism of physics.
Superstatistics
Fluctuating parameters and heavy tails as a structural feature, not noise.
Quantum-inspired modelling
Cross-disciplinary frameworks tested like any other hypothesis.
Evidence & reproducibility
Preregistration, frozen protocols, hashed inputs, negative results published.
AI architecture
Verification-first alternatives to slow LLM-only generation.
Projects by domain
Where the work lives
Counts come from the project catalogue.