1. Separate description, prediction, and action
A regime label compresses observations into a category. A forecast makes a statement about an outcome not yet observed. A decision rule maps information into an action under stated constraints. These are three different objects, even when a research notebook places them on adjacent lines. A correct classification need not predict returns, and a useful forecast need not justify a particular action after costs or constraints. Start by stating which object is being evaluated. This is a framework for examining research claims, not investment advice or a recommended allocation.
2. Write a label dictionary before an action dictionary
A label dictionary should define inputs, units, comparisons, data eligibility, and boundaries. Avoid attaching favorable or unfavorable meanings merely through names such as safe, defensive, or opportunity. A higher-volatility category is a statement about the selected volatility measure, not proof that every position is unsuitable. NBER's recession methodology illustrates why a carefully defined economic category also cannot be reduced to a single convenient observation. A separate action dictionary, when a study requires one, must explain objectives, timing, constraints, and the consequence of unresolved inputs.
3. Treat information timing as part of the claim
A label reconstructed from revised data answers a different question from a label recorded at the time. Likewise, a filtered state probability calculated using parameters estimated on future observations is not fully decision-time evidence. Retain both the data cutoff and the estimation cutoff. Scikit-learn's guidance on data leakage explains the general problem: information unavailable at prediction time must not influence model construction or evaluation. In a regime study, apply that principle to scaling, threshold selection, feature selection, state naming, and any proposed action mapping, not just the final estimator.
4. Specify the additional prediction experiment
To investigate whether a label contains information about a later outcome, define that outcome and horizon before inspecting results. Keep contemporaneous returns separate from returns beginning after the label was available. State how overlapping observations, repeated episodes, and unequal state occupancy will be handled. A category populated by many adjacent dates is not necessarily supported by many independent episodes. Compare against an appropriate baseline and report uncertainty. These design choices cannot be inferred from an attractive state chart, and a small illustrative table cannot establish empirical predictability.
5. Evaluate the action mapping separately
A proposed mapping needs its own evaluation because it introduces choices absent from the classifier: implementation delay, turnover, transaction assumptions, exposure limits, and fallback behavior. Two researchers can agree on every historical label yet obtain different strategy results from different mappings. BIS work on market conditions distinguishes dimensions such as volatility and liquidity, reinforcing the need to specify what an execution assumption actually measures. Do not attribute all differences in a strategy comparison to regime insight when leverage, asset universe, or baseline exposure also changed.
6. Use the fictional table as a claim audit
The table presents invented research cases, not recommended trades. Each row begins with something the classification process might establish and asks what remains untested. A clear label can justify examining the corresponding archive entry; it does not supply a missing return forecast. An ambiguous label can justify documenting uncertainty; it does not prove that doing nothing is optimal. A missing observation can justify marking a result unavailable; it does not establish a neutral economic state. This distinction prevents operational exceptions from becoming unearned economic conclusions.
7. Report claims at their actual evidence level
A concise report can state that a rule is reproducible, a label is provisional, a relationship was explored, or a decision rule was tested on a held-out period. Those statements are not interchangeable. Preserve unsuccessful comparisons and distinguish exploratory choices from frozen specifications. Report classification stability alongside any downstream assessment so readers can see what changed. The strongest conclusion may be that the label organizes information usefully while its investment relevance remains unestablished. That is a meaningful research result, provided it is not promoted into a performance claim.
Inspect the comparison
| Fictional research case | What is established | What still needs evidence |
|---|---|---|
| Valid inputs exceed a declared threshold | A reproducible category assignment | Any forecast or action benefit |
| Inputs sit close to a boundary | Sensitivity to a small measurement change | An appropriate decision response |
| A required release is unavailable | An incomplete information set | Whether any fallback rule is useful |
Primary references
References support the stated methods and data conventions. Original fictional examples are not observations from the cited sources.