End-to-End Research Workflow¶

ASRQuant 1.2.0 is designed to connect research stages without forcing every project into one monolithic object.

1. Define the question¶

Start with a falsifiable statement, a target, an information set and a horizon. Use asr.hypotheses when the question is being generated from data, literature, model disagreement or robustness instability.

2. Acquire and validate data¶

panel = asr.data.load("research_panel.csv", date_column="Date")
quality = asr.data.validate(panel)

Document data source, timing convention, revisions and transformations before testing the hypothesis.

3. Discover or register hypotheses¶

ideas = asr.hypotheses.discover(
    data=panel,
    targets="target_return",
    horizons=(1, 5),
    lags=(0, 1),
    min_observations=120,
)

Candidate ranking is a screening mechanism, not proof of economic significance or scientific novelty.

4. Build signal diagnostics¶

For cross-sectional research, use asr.alpha for rank/z-score transforms, forward returns, information coefficients, quantile portfolios, long-short spreads and turnover.

5. Separate factor and risk explanations¶

Use asr.factors to estimate exposures and decompose factor/specific risk. Use asr.risk for VaR, Expected Shortfall, risk contributions and scenarios.

6. Construct the portfolio¶

construction = asr.portfolio.optimize(
    returns,
    method="hierarchical_risk_parity",
)

Portfolio construction is distinct from signal generation. Record the mapping from signal to weights and every constraint.

7. Backtest with chronology and costs¶

backtest = asr.backtesting.run(prices, target_weights)

The backtest should encode the information set, rebalance timing, execution delay and costs rather than infer them after the fact.

8. Run robustness checks¶

Change windows, costs, portfolio construction, data vintages and reasonable modelling assumptions. Do not interpret a single parameter setting as a research conclusion.

9. Produce a reviewable decision¶

A research output should preserve enough information for another reviewer to answer:

What was known at each time?
Which transformation created each feature?
Which hypotheses were screened?
Which model and parameters were used?
How were weights constructed?
Which costs and delays were assumed?
Which robustness checks passed or failed?