Approach
The research programme
Six connected areas of technical work, one shared standard of evidence, and an explicit account of what would count as being wrong.
Programme
Areas of technical work
Each area exists because a question in another one could not be answered without it.
Market Microstructure
How price forms at short horizons. Order flow, liquidity provision, the sequencing between quote and trade, and the mechanics of how information reaches price rather than the question of what information exists.
- Intraday price discovery
- Liquidity and depth dynamics
- Structure formation and breakdown
Derivatives Analytics
Options positioning as measurable data rather than folklore. Open-interest dynamics, strike-band aggregation, the relationship between positioning and the underlying, and the hedging behaviour that connects them.
- Open-interest and positioning reconstruction
- Put/call positioning measures
- Positioning around scheduled and unscheduled events
Event-Based Research
Mechanically defined events, aligned in event time. The discipline here is definitional: an event rule that cannot be written down cannot be tested, and one that can be tuned after the fact has already failed.
- Structural break detection
- Event-time alignment and windowing
- Confirmation lag and its consequences
Statistical Modelling
Estimation that respects the data-generating process. Clustered and bootstrapped inference for intraday observations that are not independent, explicit confound control, and correction across the whole family of tests run.
- Association versus prediction
- Multiple-comparison correction
- Day-clustered and block-bootstrap inference
Applied Machine Learning
Applied sceptically, and applied to markets. A flexible model fitted to financial data will always find something; the entire value of the exercise lies in whether the something survives a period the model has never seen. The methods are standard — the discipline around them is the work.
- Time-series and event-aligned feature construction, without look-ahead
- Classification and forecasting on financial datasets
- Held-out evaluation, evaluated once
- Discarding specifications that decay out of sample
Real-Time Market Systems
The engineering half. Reconstructing the past is a different problem from observing the present, and the gap between them — latency, partial snapshots, revisions — is itself a research subject.
- Streaming ingestion and state management
- Online detection with explicit confirmation lag
- Instrumentation and reproducible logging
Architecture
How work moves from data to result
The same seven layers for every study. Select a layer to see what happens there and why it is separated from the next.
Layer 01
Data
Index, options-chain and derived reference data, ingested with provenance recorded. Vendor-licensed data is never redistributed; what can be shared is documented in each paper's disclosure statement.
Philosophy
Research philosophy
Seven commitments. They are listed here so that any published result can be checked against them.
- 01
Define the question first
A question written after seeing the data is not a question, it is a description. Specification comes before estimation, and what was pre-specified is disclosed separately from what was exploratory.
- 02
Separate association from prediction
That two things move together is a fact about the past. That one anticipates the other is a claim about the future. They require different tests, and conflating them is the most common failure in applied market research.
- 03
Control obvious confounds
If a signal is measured around a price move, recent price movement is the first thing that must be controlled for. Not doing so manufactures results that look strong and mean nothing.
- 04
Prefer out-of-sample validation
A held-out period, held back honestly and evaluated once. Re-tuning after a failed out-of-sample test converts the holdout into training data and destroys the only real evidence available.
- 05
Report negative results
A signal that fails is information — often more useful than one that appears to work, because far fewer people publish it. Failures are written up in full, with the same method and the same detail.
- 06
Make methodology reproducible
Every published figure should trace back to code that runs end to end from raw data. A result nobody can re-derive cannot be checked, and a result that cannot be checked is not a result.
- 07
Avoid overfitting narratives to data
The story is fitted last, if at all. A mechanism that is 'consistent with' the evidence is described as consistent with it, not asserted as the explanation.
These are commitments, not achievements. Where a published paper falls short of one, the shortfall is stated in that paper's limitations section.