A strategy should be represented as a testable economic policy, not as a backtest.

We treat a systematic strategy as a structured causal object:

Market state →Decision policy →Target positions →Realized positions →Conditional exposures →Outcomes

The chain is incomplete without the mechanism that explains why the return should exist, the data used to observe the state, the costs of implementation, the ways the strategy can fail, the evidence for each relationship, and its version history.

A strategy that moves beyond exploratory research should be able to fill in this record. Where a field is unknown, the record says so.

Identity

What strategy and version is this?

Mandate

What problem is it intended to solve?

Mechanism

Why should the return exist?

Universe

What can it trade?

Data

What information was available at each decision time?

State

Which conditions materially change the economics?

Policy

How does state map to action?

Sizing

How does action map to risk?

Exposure

What behavior is created across states?

Costs and capacity

What frictions apply, and at what scale does behavior change?

Failure

How can it lose or stop working?

Evidence

What supports the mechanism and the policy?

Falsification

What would make us revisit or reject it?

Production

Which artifacts implement the approved strategy?

Outcomes and provenance

What happened, and can the lineage be reconstructed?

A feature is evidence only to the extent that it measures a hypothesized mechanism or state. The sequence is mechanism, then the state it implies, then a measurable variable, then a feature, then a test. Not the reverse: available data, a feature search, an attractive backtest and a story afterward.

A mechanism names, as concretely as possible:

• who is paying for the opportunity, and for what risk or service;
• why it should persist once others know about it;
• which conditions should strengthen or weaken it; and
• which observations would contradict it.

Machine learning can combine information. It does not remove the obligation to define the economic problem. Models are implementation choices, not the thesis.

Each stage leaves a record, so that what was believed, what was fixed in advance and why a version was promoted or rejected can be reconstructed later.

Point-in-time correctness is a research property, not a data-engineering footnote.

For every material input we record the provider, field definitions, observation and effective timestamps, publication lag, revision behavior, transformation lineage and a version identifier, and we distinguish what was eligible in principle from what was available at the time.

Data observed on different bases, such as end-of-day values and intraday snapshots, is not silently combined. A strategy fails validation if its evidence cannot be reproduced from frozen source data.

Every complex model is compared with simple baselines. The question is whether it adds economically meaningful information, not whether it performs well in isolation.

Time-series research is evaluated walk-forward and reported by fold. Exploration can be broad; confirmation is narrow, frozen and auditable, and the same result is never described as both. The more choices exercised before a result appears, the higher the validation burden that follows.

If moving a threshold slightly destroys a strategy, the burden of proof rises. A robust mechanism usually produces a region of defensible behavior, a plateau or an ordering, rather than a single best parameter. Robustness tests attack the assumptions most capable of manufacturing the result: neighboring parameters, other windows, delayed execution, worse costs, removed features, simpler models and regime subsamples.

Many systematic strategies are nonlinear, and an unconditional statistic can hide the states in which the strategy matters most. The object of interest is often not a constant exposure but a function of the state: β = β(state).

The questions an allocator cares about are conditional. How does equity sensitivity change in drawdowns? What happens when implied volatility jumps, or the volatility curve inverts? Does the strategy become correlated with the rest of a portfolio precisely when liquidity falls?

Exposure is described in layers that are kept separate: mechanical (implied by positions), statistical (estimated from returns), conditional (within defined states), scenario (under specified shocks), policy (how intended positioning changes with the state) and realized (what occurred in production). The estimation method is always named.

What does beta mean when the policy changes with the state?

Kinds of evidence are kept distinct rather than compressed into one confidence score: mechanism evidence, historical structural evidence, out-of-sample evidence, robustness evidence, implementation evidence and live evidence. Hypothetical, simulated and live results are labeled as such, and live evidence accumulates alongside the original claim without rewriting it.

Not treated as evidence:

• an attractive backtest on its own;
• a result found after many unrecorded variations;
• a holdout that has already been looked at;
• feature importance, which is not causal evidence;
• agreement among correlated inputs, which is one signal counted several times; or
• a simulated result presented as a live one.

A strategy description is incomplete without an account of how it can fail: the mechanism stops persisting, the regime changes faster than the policy responds, data changes meaning, a hidden exposure dominates in stress, execution becomes impossible at modeled prices, or capacity is exceeded.

Every strategy approved for production carries explicit review conditions. A falsification rule does not have to mean automatic shutdown. It means the evidence triggers a defined review rather than being rationalized away.

P&L is a lagging, noisy signal, so monitoring watches assumptions: data integrity, calibration, positions, costs, target versus realized positions and exposure drift. It separates three situations that call for different responses: the strategy experienced a risk it was designed to carry; production deviated from the approved policy; or the thesis is no longer supported.

Every material result should say which code, data and configuration produced it, which decision record authorized the test, which periods were training, validation and holdout, and what was known before the result was observed.

Definitions are fixed before outcomes are read. A definition whose outcomes have been seen is not edited; a change becomes a new version. Decision records are append-only, and a change to production behavior is a new version.

The aim is reconstructability: a future researcher should be able to reproduce not only the result, but the decisions that led to it.

This standard is meant to expose uncertainty, not to remove it rhetorically. It does not claim:

• that a strategy approved under it will be profitable;
• that economic mechanisms are known with certainty, or that historical robustness guarantees future robustness;
• that machine learning is superior to rules, or that every strategy reduces to a single beta;
• that Volatility Street independently verifies or rates strategies it has helped develop; or
• that the same evidence threshold suits every level of capital and consequence.

Questions and disagreements are welcome. Contact us.