Hypothesis Discovery — ASRQuant 1.2.0

ASRQuant 1.2.0 promotes hypothesis discovery to a first-class quantitative research API.

The design separates two questions that must never be conflated:

1. Is there evidence worth testing? — data/literature/model/robustness evidence.
2. Is the idea novel? — a separate prior-art question that cannot be established automatically from one corpus.

Public API

import asrquant as asr

asr.hypotheses.from_data(...)
asr.hypotheses.from_literature(...)
asr.hypotheses.from_model_disagreement(...)
asr.hypotheses.from_robustness(...)
asr.hypotheses.discover(...)
asr.hypotheses.search(...)
asr.hypotheses.audit(...)

Data-driven discovery

ideas = asr.hypotheses.from_data(
    data={
        "rates": rates,
        "macro": macro,
        "equities": equities,
    },
    domain="fixed_income",
    targets="equities::VALUE_MINUS_GROWTH",
    horizons=(1, 5, 20),
    lags=(0, 1, 5),
)

print(ideas.to_frame())

The default screening layer can generate candidates from:

• chronology-safe lagged relationships;
• high/low state or regime differences;
• cointegration candidates;
• structural mean/variance changes;
• serial dependence and tail behaviour;
• correlation breaks;
• yield-curve factor and regime diagnostics through the existing discovery engine.

Data-mining controls

Every statistical screening run records:

• total number of tests performed;
• chronological discovery and holdout sample sizes;
• discovery effect size;
• raw discovery p-value;
• Benjamini-Hochberg FDR q-value;
• holdout effect and p-value;
• transformation applied to each variable;
• data status and falsification rule.

Data statuses are deliberately explicit:

• EXPLORATORY
• DATA_SUPPORTED
• OUT_OF_SAMPLE_SUPPORTED
• INCONCLUSIVE
• FALSIFIED
• LITERATURE_DERIVED

A data status is not a novelty status.

Literature-driven discovery

The existing source-linked literature engine is preserved:

corpus = asr.LiteratureCorpus.from_pdfs("papers/", topic="interest rates")
ideas = asr.hypotheses.from_literature(corpus, topic="interest rates")

Page-level source excerpts, corpus fingerprints and conservative corpus-relative labels remain part of the workflow.

Combined discovery

ideas = asr.hypotheses.discover(
    data=research_panel,
    papers="papers/",
    domain="fixed_income",
)

When both data and literature are supplied, data-generated candidates are audited against the supplied corpus. The audit may return:

• PRIOR_ART_FOUND
• CORPUS_RELATED
• POTENTIAL_GAP
• CONTRADICTORY_LITERATURE
• NOVELTY_NOT_ESTABLISHED

None of these labels means global novelty has been proven.

Search

matches = asr.hypotheses.search(
    "forward curve instability before rate regime transitions",
    hypotheses=ideas,
    papers=corpus,
)

print(matches.hypotheses)
print(matches.excerpts)

The search surface returns generated hypothesis matches and source-linked literature excerpts separately.

Novelty audit

audit = ideas.audit(0, corpus=corpus)
print(audit.summary)
print(audit.closest_matches)

The novelty audit is intentionally fail-safe. Without a supplied corpus it returns NOVELTY_NOT_ESTABLISHED. Even with a corpus, the result is explicitly corpus-relative and must be followed by a documented prior-art review before public novelty language is used.

ResearchProject hand-off

A selected hypothesis enters the existing ASRQuant research pipeline directly:

project = ideas.start(0)

The hand-off preserves:

• hypothesis statement;
• research question;
• predictor and target;
• expected sign and horizon;
• data evidence;
• suggested methods;
• falsification criteria;
• prior-art references;
• data and novelty statuses.

From there the normal ASRQuant workflow continues through data planning, feature construction, econometrics/ML, portfolio/backtest experiments, robustness, decision governance and publication artifacts.

Research rule

ASRQuant may automate candidate generation, screening, ranking and evidence organization. It does not automate scientific truth. A candidate remains a research object until the evidence, assumptions, prior art and falsification tests have been reviewed.