Approach / Research standards
We separate an economic mechanism, an observable signal, a statistical forecast and a funded decision. Each transition requires different evidence.
HYPOTHESIS → EVIDENCE → DECISIONA RESEARCH STANDARD
Each stage asks for stronger evidence. A result advances only when the next economic question has an answer.
Define the instrument universe, venue, event time, receipt time and eligible observations before measuring a pattern. Preserve missing, stale and rejected records in the accounting.
The pattern disappears under valid timestamp, instrument-mapping or data-quality checks.
Reproducible measurement; no claim of prediction yet.
This is a methodological framework, not a report of five completed Side Street studies.
A candidate can stop at any gate. Negative findings and insufficient measurement support remain useful outcomes when they resolve a specific question.
Features must use information received before the decision. Source age, reconstruction gaps and unavailable history are part of the measurement—not convenient exclusions after an outcome.
Compare on common eligible populations. Activity, recent price and direction-neutral controls challenge whether a richer representation adds the information it claims to add.
Overlapping forecasts can contain few distinct episodes. Report censoring, temporal concentration and dependence alongside probability scores; reserve later chronology for confirmation.
Bind results to versioned code and input identities. Preserve failures and interruptions. Selected calculations receive separate numerical or artifact checks; those checks are not fresh market replication.
Visible-depth replay does not establish actual fills, maker queue position, endogenous impact or deployable capacity. Completed empirical work remains exploratory; later replication and economic admission are separate decisions.
Language models already assist with source synthesis, experiment design, implementation and criticism. Their output is accepted through inspectable artifacts, deterministic checks and responsible scientific judgment.
We are developing supervised workflows with bounded tasks and explicit data access. The intended measure of progress is accepted, reproducible work—not generated prose, agent count or a claimed stream of discoveries.
A controlled throughput evaluation and Claude integration are planned. Neither is an established platform advantage.
Selection risk and the research literature