AlphaLuck

Data and code for Trading versus Holding in a Live Language-Model Competition, by Daniel Traian Pele, Diana Ioana Micu, Victor Rareș Curescu and Miruna Mazurencu-Marinescu-Pele. Repository: https://github.com/QuantLet/AlphaLuck.

Current replication

replication_20261102/ contains the saved inputs, calculations and results for:

• Prediction Arena's trading-versus-holding comparison, with independently reconciled transactions, external deposits, network fees and settlement evidence;
• formation at 12, 24 and 48 hours, with the same closing date;
• corrected Alpha Arena funding alignment, all 336 formation hours, both joint-stress directions, and the checks using only preceding margin observations.

At first-day formation, holding preserves more terminal wealth for three of the four attributable Prediction Arena accounts. One starts in cash. GLM's sign changes at earlier formation, and settlement-only bounds leave Gemini's sign uncertain. GPT is reported separately because of an unmatched order. These are descriptive comparisons and do not establish a general advantage of holding or persistent model skill.

Saved-data replay needs no credentials or new downloads. With Python 3.11+ and the packages in replication_20261102/requirements.txt installed:

cd replication_20261102/alpha_arena
python src/reproduce_threshold_results.py
python -m unittest discover -s tests
cd ../prediction_arena
python calculate_replication.py
python calculate_twr.py
python formation_sensitivity.py
python -m unittest test_replication.py test_formation_sensitivity.py
python independent_inputs/test_history.py

Read the replication instructions for valuation assumptions, source attribution and the limits of the reconstruction. This folder supplements the earlier reconstruction; it does not contain raw organiser snapshots.

Reconstruction and earlier standalone analyses

reproduction/ retains the public Hyperliquid records, AI-Trader inputs and the numerical tables, figures and Alpha Arena code. Its README explains the offline checks and which parts of a complete reconstruction still depend on unavailable organiser archives.

The following six compact Quantlets retain their October inputs and results. For the margin calculations use replication_20261102, not the older isolated-margin values in AlphaLuckFormationHour.

Quantlet	Reproduces
AlphaLuckRiskProfiles	Initial-capital volatility and gross-exposure profiles across the observed arms
AlphaLuckRiskReturn	Model-level and within-arm risk-return associations, recomputed from account data
AlphaLuckPnlAttribution	Signed accounting contributions, with timing and execution combined
AlphaLuckDenominatorSensitivity	Paired capital-base comparisons, shared-date reweighting and deletion sensitivities
AlphaLuckFrozenBenchmark	Aggregation of hourly-funded endpoints and conditional closing-cost and common-price stresses
AlphaLuckFormationHour	The frozen-book comparison for every formation hour, to the close and over fixed 1-, 3- and 7-day horizons, under account-wide and passive-isolated margin policies

Each compact folder remains independently executable. See DATA.md for the boundaries between archived observations and derived numerical inputs.

Citation and licences

Cite Pele, Micu, Curescu and Mazurencu-Marinescu-Pele (2026), AlphaLuck: Trading versus Holding in a Live Language-Model Competition, research data and code. https://github.com/QuantLet/AlphaLuck. Machine-readable metadata are in CITATION.cff.

Code and authored documentation use the MIT licence. The Envio / moose-code Polymarket archive retains CC BY 4.0 attribution; other third-party records retain their source notices. This does not relicense third-party material. Raw organiser snapshots, private feedback and licensed external-price datasets are not added. GitHub publication does not establish indexing in the QuantLet catalogue.

The repository includes replication_20261102/robustness_checks/: arm deletions, model-block sign-flip reference calculations, cash closing costs, and cash comparisons across 169 overlapping seven-day windows and 336 formations to the close. These are descriptive sensitivities on one market path. Run python replay_checks.py and python cash_formation_sensitivity.py from that directory; both use only the Python standard library.