AlphaLuck v2.1.2 — authorship and citation update

Adds Miruna Mazurencu-Marinescu-Pele as the fourth author and synchronizes citation and QuantLet metadata.

Data, calculation code and numerical results are unchanged from v2.1.1. The replication supplement retains the arm-deletion, model-block sign-flip, cash-closing-cost and cash formation-time checks in robustness_checks/.

The previous release and asset are preserved.

AlphaLuck v2.1.1

This release adds cash formation-time sensitivity, omitted-arm comparisons, model-block sign-flip reference calculations and closing-cost sensitivities to the existing Alpha Arena and Prediction Arena deposit.

Cash is ahead in 124 of 169 overlapping seven-day windows, with a median gap of 13.04 percentage points. At 10 and 100 basis points of closing costs, the counts are 123 and 116. The zero-cost cash gap is the negative of realised post-formation P&L. All windows share one market path; the sign-flip references require joint sign symmetry and do not establish significance across independent markets.

The robustness_checks folder includes derived account inputs and standard-library replay commands. Rebuilding the inputs from reproduction/ also passed, with five outputs byte-identical. Existing Alpha Arena and Prediction Arena inputs and calculations are preserved. The figures and citation metadata are updated.

The standalone ZIP contains the data and code needed for offline replay. It excludes the manuscript, submission documents and raw organiser archives.

AlphaLuck 2.1.0: Prediction Arena and revised Alpha Arena calculations

This research release adds the Prediction Arena trading-versus-holding comparison and the revised Alpha Arena calculations used in Trading versus Holding in a Live Language-Model Competition.

The attached AlphaLuck_replication_supplement_v2.1.0.zip is a standalone saved-data package. It contains Prediction Arena transaction reconciliation, external cash flows, wallet-paid network costs, settlement evidence, valuation bounds and formation checks at 12, 24 and 48 hours. It also contains corrected Alpha Arena funding alignment, all 336 formation hours, both joint-stress directions and prior-only threshold checks. Start with README.txt after extracting the archive.

The repository's reproduction/ tree includes the additional cost, cash/index, instrument-deletion and overlapping-window checks. Earlier standalone Quantlets retain their October inputs and are labelled accordingly; use replication_20261102/ for the revised margin calculations.

Saved-data calculations and unit checks passed. Prediction Arena results agree with the manuscript at the reported precision; sorted accumulation also makes the saved outputs reproducible across Python hash seeds. The revised Alpha Arena sweeps agree with the expected results within numerical precision.

The comparison remains descriptive. GPT is reported separately because of an unmatched order; one attributable account starts in cash. Formation time, valuation and margin assumptions matter. Raw organiser snapshots, complete operator decision logs, private feedback and submission documents are not distributed in this release. Third-party records retain their source attribution and licence notices.

Cite: Pele, D. T., Micu, D. I., and Curescu, V. R. (2026). AlphaLuck: Trading versus Holding in a Live Language-Model Competition. Research data and code, version 2.1.0. GitHub, QuantLet/AlphaLuck. https://github.com/QuantLet/AlphaLuck/releases/tag/v2.1.0

This is a research-data and software release, not journal submission or acceptance. No DOI has been assigned by this release process.