The dataset I kept trying to finish
My collection work included attempts at uninterrupted 24-hour and seven-day acquisition. Those targets exposed useful infrastructure problems, but they also risked making every research question wait for the next reliability campaign. I needed a way to study the verified data already available without declaring the whole recording complete.
Three things had to stay separate. Process uptime says the collector was running. Durable records say observations reached storage. Certified coverage says a particular data product and interval passed its native integrity and continuity checks. None of those statements automatically establishes the other two.
Use verified hours without inventing the gaps
The research reader selects native PASS product-hours: hour-sized units for a specific exchange, instrument and data product that satisfy the existing certification rules. It verifies the original certificates and sealed inputs on read-only replicas, including file identities before use. It does not relax certification to admit an old failed capture.
For a study requiring several products together, only their shared eligible coverage can be used. Independent instruments can be studied separately. Uncertified intervals and unavailable dependencies remain exclusions, with reasons. A missing stretch of trades cannot be interpolated into a supposedly factual market history.
Recovery epochs mark new continuity segments after an interruption. Feature windows and later outcomes must not cross gaps, recovery boundaries or chronological sample splits. The study leaves space around those boundaries and warms up features again afterward. Information unavailable at the decision cutoff stays unavailable.
A concrete exploratory study
The real-data order-flow imbalance study used 459 native PASS Trade product-hours and processed 19,214,835 factual trades. Product-hours are summed across selected trade streams; they do not mean 459 uninterrupted hours on one instrument. These are research-input counts, not a whole-campaign PASS claim.
Here, order-flow imbalance compares buyer- and seller-initiated trade volume over a short window. The study asked how that measurement related to subsequent price changes. It used trade tapes, so unavailable order-book spread stayed missing. Stale prices and unknown inputs also remained missing rather than receiving convenient substitutes.
Only after defining that question does a statistic help. Rank correlation, often called an information coefficient or IC, describes whether larger feature values tend to accompany larger later returns. It does not establish a causal mechanism or show that a strategy would survive trading costs.
Reproducibility is not a fresh test
The recorded checks compared selected feature and outcome calculations with independent native replay and checked every evaluation against gap, split and information-time boundaries. Those checks support the calculation and selection behavior. They do not make the sample representative of every market condition.
The experiment log also retained prior exposure. Before the initial run, 416 source certificates had already been viewed; all 459 were exposed afterward. A later plumbing rerun explicitly recorded holdout reuse. Reproducing the output was useful for engineering verification, but it did not create new evidence about unseen markets.
Limitations and the next useful question
The study remains EXPLORATORY. A holdout is data reserved for evaluation; once examined or reused, it cannot be treated as a clean independent test. This work has no clean forward out-of-sample (OOS) validation and establishes no proven trading alpha or profitability.
I can use this smaller verified sample to examine whether a method deserves a genuinely fresh test while infrastructure work continues. Missingness, source recovery and prior exposure all constrain that decision. Waiting longer is useful only if it improves the evidence needed for the next question.
Evidence basis: the private Factory continuous-acquisition and gap-aware research reviews, native PASS reader and recorded real-data OFI reports. Financial result tables and raw observations are not reproduced.