LAMINAR: Measuring an Order-Flow Coherence Label Across Twelve Futures Instruments

This working paper evaluates LAMINAR, a session-relative order-flow classifier, across 15 CME futures tapes representing 12 instruments, 194 exchange-day sessions, and 36,533,928 trades. The sample spans equity index, energy, metals, interest-rate, and currency futures. The study examines three distinct properties of the classifier: how frequently it selects observations, how its selections respond to controlled changes in the input, and whether previously selected segments retain elevated selection rates when placed in a different context.

The classifier is evaluated using a common numerical price convention, exchange-day session boundaries, and explicit accounting for its four regime labels. Observations record the classifier’s state at the last trade in each populated two-minute reporting bin. The LAMINAR label occurs in every included session. Pooled tape incidence ranges from 6.35% to 9.67%, while individual session rates range from 2.69% to 14.39%. These are frequencies of sampled classifier states, rather than estimates of continuous elapsed-time occupancy.

A control grid comprises 70 executions over seven source instrument-sessions in Japanese yen, 2-year Treasury, gold, and S&P futures. Controls include aggressor-sign permutation, price-increment permutation, block reordering, independent resampling of session marginals, synthetic episode insertion, and transplantation of previously LAMINAR-labelled real segments. Repeated random seeds characterize variation under the chosen transformations; they are not additional independent market sessions.

Independently resampled controls, which preserve the original timestamp sequence, produce mean incidence at least as high as the real inputs in three of the four tested instruments. In the tested S&P session, price-path permutation reduces incidence from 7.94% to 4.92–5.56% across two seeds. Positive controls reveal different patterns: synthetic insertions produce higher mean inside than outside incidence in gold, yen, and the 2-year future, but lower incidence in the S&P. Grafted real segments produce higher mean inside incidence in gold and the S&P, but lower incidence in yen and the 2-year future. These results distinguish aggregate selection frequency from sensitivity to particular perturbations and retention of a label under changed context.

The contribution is a measurement protocol and empirical account of what an order-flow coherence label establishes under specified observation and control conditions. The findings support conditional statements about the tested inputs and transformations. Predictive returns, executable trading value, and causal explanations based on market density are outside the study’s established results.

The accompanying evidence package contains derived session-level receipts, recorded control outputs, price-representation checks, generated tables and figures, manuscript source, figure-generation code, and audit results. Licensed raw market events are not redistributed. The classifier’s proprietary construction is outside the disclosure scope; the supplied materials document the reported measurements and their supporting evidence.