Stock DossierLAB
Home / Data Methodology

Provenance and limitations

Data and Replay Methodology

How Stock Dossier Lab separates signal discovery, entry assumptions, forward outcomes, benchmarks, model versions, and the limits of historical evidence.

Point-in-time guardrailsUTF-8 dataReproducible fields

Dataset contract

What one replay row represents

One row represents a ticker that met one controlled experiment on one signal date. The assumed entry is the next trading session's open. Outcomes are measured at 1, 3, 5, 10, and 20 trading-day closes, alongside maximum favorable excursion, maximum adverse excursion, and QQQ/SPY-relative return.

Modelv4671_state_researchversioned rules
Signal dates120not calendar days
Entry basisNext openafter discovery
Sector modeneutralleakage guard

Pipeline

From raw bars to published evidence

1. Cached daily barsOHLCV data keyed by ticker and date, with missing symbols reported.
2. Point-in-time featuresMomentum, trend, relative volume, RSI, entry quality, and risk state calculated only from available bars.
3. Controlled experimentsStrict and one-rule relaxation profiles isolate what changed candidate density and outcomes.
4. Forward auditFixed entry assumption, multiple horizons, benchmarks, and complete-row checks.

Leakage controls

Defensive choices when perfect history is unavailable

  • Historical replay uses neutral sector scoring when point-in-time sector metadata cannot be proven.
  • The signal date ends feature calculation; forward prices are used only for outcome fields.
  • Entry begins at the next session open, not at a signal-date price discovered after the close.
  • A missing ticker or incomplete horizon remains missing instead of being silently imputed.
  • Every published table names the experiment, model version, window, and benchmark.

Interpretation

What the replay does not prove

Not independent trades

Daily observations for one ticker can overlap. Row count therefore overstates independent sample size.

Not live execution

The replay does not reproduce every order type, spread, partial fill, tax, or liquidity event.

Not untouched validation

Rules evolved during the research period. A future frozen out-of-sample test is still necessary.

Not personal advice

The model studies groups and states. It does not know a reader's finances, objectives, or risk capacity.

Research downloads

Machine-readable summaries

V2 publishes a compact JSON evidence file for reproducibility without redistributing the full underlying market-data cache.