Datasets:

vuongtsc
/

alpha-research-methods

protocol_id stringlengths 4 4	stage stringlengths 11 25	objective stringlengths 53 95	fixed_variables listlengths 2 5	changed_variable stringlengths 4 52	promotion_gate stringlengths 56 96	stop_rule stringlengths 31 89	leakage_check stringlengths 55 96	provenance stringclasses 1 value
P001	Hypothesis registration	Turn an intuition into a falsifiable statement before candidate generation.	[ "economic mechanism", "expected direction", "target horizon", "failure condition" ]	none	The hypothesis names a causal channel and an observation that would disprove it.	Reject vague stories that can explain every possible outcome.	Confirm that all proposed inputs are observable at the decision timestamp.	original
P002	Source audit	Separate authoritative definitions from automated or approximate interpretations.	[ "source version", "publication date", "market context" ]	interpretation candidate	Every material concept has traceable provenance and uncertainty labels.	Stop when required definitions or redistribution rights cannot be established.	Verify that quoted or derived content is licensed for the intended use.	original
P003	Portability audit	Decide whether the economic mechanism survives a change of market.	[ "causal channel", "participant behavior", "measurement meaning" ]	target-market institution	All required institutional conditions exist or have an explicit target analogue.	Reject syntax-only portability.	Use only target-market data known at the decision time.	original
P004	Candidate trio	Separate source fidelity from target-market adaptation.	[ "universe", "evaluation window", "risk settings" ]	faithful, adapted, or motif-hybrid construction	Each variant has distinct provenance and changes only the declared construction layer.	Reject comparisons that change data, settings, and structure simultaneously.	Freeze the candidate definitions before reading evaluation results.	original
P005	Fixed-settings comparison	Attribute outcome differences to one research choice.	[ "universe", "delay", "decay", "neutralization", "truncation" ]	one candidate component	The comparison is reproducible and paired under identical settings.	Discard confounded comparisons rather than explaining them after the fact.	Settings are chosen from a predeclared policy, not optimized on the outcome.	original
P006	Sign canonicalization	Distinguish a useful inverse relationship from a direction mistake.	[ "candidate structure", "data", "settings" ]	direction only	The inverse direction is economically interpretable and pre-registered as a separate hypothesis.	Do not search repeated sign and parameter combinations after observing outcomes.	Record the original sign and the rule that allowed one inversion.	original
P007	Small pilot	Buy enough evidence to retire weak mechanisms cheaply.	[ "pilot budget", "promotion criteria", "candidate roster" ]	pre-registered candidate variants	At least one variant clears the predefined signal-quality gate and beats its baseline.	End the mechanism when the fixed pilot budget produces no qualifying evidence.	Do not add candidates after seeing partial pilot results.	original
P008	Sequential stopping	Prevent exploration budgets from becoming permanent taxes.	[ "maximum budget", "minimum effect", "decision checkpoints" ]	sample size at scheduled checkpoints	Evidence exceeds the predeclared continuation threshold.	Stop at the registered deadline; do not move the baseline after failure.	Document every exception and treat a redesigned experiment as a new version.	original
P009	Signal-quality promotion	Promote only candidates with sufficient native predictive quality.	[ "quality thresholds", "evaluation universe", "cost assumptions" ]	candidate	The candidate clears all native quality requirements without repair.	Do not optimize downstream uniqueness for candidates that lack native quality.	Thresholds are defined before evaluating the candidate pool.	original
P010	Downstream gate diagnosis	Name the exact layer blocking an otherwise strong candidate.	[ "qualified candidate", "native settings" ]	one downstream constraint	A paired intervention fixes the named gate without breaking prior gates.	Stop broad repair when paired canaries do not change the blocking gate.	Do not reuse final evaluation feedback as a generator training target without isolation.	original
P011	Robustness review	Check whether the mechanism survives reasonable changes without parameter mining.	[ "mechanism", "direction", "data provenance" ]	predeclared subperiod, universe, or nearby parameter	The effect remains directionally coherent across independent slices.	Reject isolated success that depends on one narrow slice or extreme parameter.	Hold out at least one slice that was not used in candidate selection.	original
P012	Public release review	Share methodology without releasing private signals, third-party content, or platform evidence.	[ "deny-set", "license policy", "public schema" ]	release artifact	Automated leakage scans pass and every row has original provenance.	Refuse publication on any exact overlap, secret, restricted metric, or uncertain license.	Compare all text against private expression stores and inspect the final manifest before upload.	original

Portable Alpha Research Methods

An original educational dataset for research agents and human researchers who need to turn quantitative intuitions into falsifiable, portable experiments. It focuses on economic mechanisms, semantic translation, and disciplined experiment design rather than distributing trading signals.

What is included

mechanisms — 24 records: hypothesis cards with economic intuition, portability conditions, failure modes, feature classes, and falsification tests.
translation_rules — 20 records: exact, approximate, rejected, and policy-level rules for preserving meaning across data systems.
experiment_protocols — 12 records: staged protocols for preregistration, paired tests, stopping rules, robustness, and safe publication.

from datasets import load_dataset

mechanisms = load_dataset("vuongtsc/alpha-research-methods", "mechanisms", split="train")
rules = load_dataset("vuongtsc/alpha-research-methods", "translation_rules", split="train")
protocols = load_dataset("vuongtsc/alpha-research-methods", "experiment_protocols", split="train")

Safety boundary

This repository contains no executable alpha expressions, private candidates, platform-derived backtests, portfolio outcomes, identifiers, credentials, data catalogs, or copied third-party formula collections. Every data row is an original synthesis released under CC BY 4.0.

The build compares every public text value against a private expression deny-set, rejects formula-like syntax and secrets, enforces original provenance, and uploads only a fixed file allowlist. See manifest.json for the build outcome.

Intended use

training research agents to write falsifiable hypothesis cards;
classifying faithful, approximate, and invalid research translations;
designing paired experiments and preregistered stopping rules;
teaching cross-market portability and leakage-aware research practice.

Limitations

The records are methodological guidance, not empirical claims.
Feature classes are intentionally generic and are not a data catalog.
No record should be interpreted as investment advice or evidence of profitability.
Users must validate market, data, legal, and operational assumptions independently.

License

CC BY 4.0. Attribution: vuongtsc, Portable Alpha Research Methods.

Downloads last month: 24

Number of rows:

Total file size:

36.8 kB