Skip to content

ASRQuant Research Discovery

ASRQuant 1.1.0 introduced, and 1.2.0 preserves, a conservative research-discovery layer for the ASR Research and Analyst teams. Its job is to transform evidence into falsifiable research candidates. It never treats an automatically generated idea as a proven novelty claim.

One workflow

import asrquant as asr

board = asr.discovery.weekly(
    data=curve_history,
    domain="fixed_income",
    n=10,
)

print(board.to_frame())
project = board.start(0)
cycle = asr.weekly_cycle(board, 0, launch_friday="2026-08-14")
cycle.publication_pack("weekly/WR-001")

The pipeline is:

evidence -> observation -> research question -> hypothesis -> falsification rule -> ResearchProject -> robustness -> claim audit -> publication pack.

Discovery sources

asr.discovery.weekly() can combine several independent sources:

  • market or curve data: mean/variance shifts, tails, serial dependence, relationship breaks, PCA residuals and curve-regime changes;
  • literature: source-linked hypotheses, limitations, assumptions and corpus-relative novelty evidence;
  • model disagreement: points where competing models give materially different predictions;
  • robustness grids: specifications for which a reported result is unstable;
  • curated research catalogues: testable starting points that remain explicitly labelled NOT_ESTABLISHED for novelty.

Friday-to-Friday contract

The ASR weekly cycle is eight dated checkpoints, including both Fridays:

Day Stage Required result
Friday N Launch question, hypothesis, owners, falsification rule, evidence contract
Saturday Prior art nearest literature, competing explanations, definitions
Sunday Data design frozen sample, provenance, lags, validation split
Monday Baseline simplest valid benchmark/reproduction
Tuesday Main experiment primary estimate/model and diagnostics
Wednesday Robustness alternatives, subperiods, costs, stress and falsification
Thursday Review independent reproduction, limitations, claim audit
Friday N+1 Publish research note, notebook, figure, evidence and next question

board.weekly_plan(...) returns this plan as a dataframe.

Evidence, not automatic novelty

Every ResearchCandidate contains:

  • research_question;
  • hypothesis;
  • falsification_rule;
  • methods;
  • data_requirements;
  • source_observations;
  • risks;
  • priority_score;
  • novelty_status.

The default novelty status is NOT_ESTABLISHED. A research team must perform and document the prior-art search before words such as new, novel, first or original are used in public communication.

Hand-off to the existing research engine

ResearchBoard.start() creates the normal ASRQuant ResearchProject; nothing is lost by starting from discovery. The project can then use literature ingestion, data planning/fetching, feature and signal construction, hypothesis tests, backtests, robustness, decisions, manifests and HTML reports.

Publication pack

WeeklyResearchCycle.publication_pack() creates a reproducibility-oriented directory containing:

  • research_brief.md;
  • RESEARCH_NOTE.md;
  • REPRODUCIBILITY_CHECKLIST.md;
  • CLAIM_AUDIT.md;
  • weekly_plan.csv;
  • cycle_status.csv;
  • project_manifest.json;
  • standard folders for data, notebooks, figures, tables, source code and evidence.

This is a starting contract, not evidence that a result is scientifically correct. The independent review step remains mandatory.