1339-GRP Persistent Dealer-Gamma Regimes Under Temporal Obfuscation: A Re-Run on Complete Inputs with a Controlled LLM Re-T

Presenter Information

Location

https://campus.kennesaw.edu/colleges-departments/ccse/events/computing-showcase/fa26-cday-program.php

Document Type

Event

Start Date

20-11-2026 2:00 PM

Description

Temporal obfuscation tests whether a large language model (LLM) reasons about numbers1 or recalls market history. Our earlier work reported that an LLM detects persistent dealer-gamma2 regimes in 30-day windows of obfuscated daily gamma exposure (GEX) of SPY, with a sharp rise3 after 2023. We re-run that test on the complete data. The published 2024–2025 inputs lacked the put4 leg of the options chain. We assembled six daily series (recovered, archived or rebuilt), replayed the5 deterministic rule, bridged o4-mini to gpt-6-luna on the published inputs, reran the 2024 and 20206 baselines with relative day labels, and repeated seven arms with an open-weights model. On the7 published inputs the rule detects all 501 windows of 2024–2025; on the complete series it detects 608 of 502 (12.0%), and gpt-6-luna 63 of 502 (12.5%). The two models agree on 99.6% of the published9 windows (Cohen’s kappa 0.993), and gpt-6-luna follows the rule in 99.7% of 4,691 windows (kappa10 0.985); gpt-oss-20b agrees with gpt-6-luna on 99.4%. Removing same-day expiries gives 14.7%, not11 the published rate. With relative labels gpt-6-luna detects 80.7% of the 2024 baseline windows, against12 81.2% with dates. The complete series shows no post-2023 increase that the data distinguish from13 zero

This document is currently not available here.

Share

COinS
 
Nov 20th, 2:00 PM

1339-GRP Persistent Dealer-Gamma Regimes Under Temporal Obfuscation: A Re-Run on Complete Inputs with a Controlled LLM Re-T

https://campus.kennesaw.edu/colleges-departments/ccse/events/computing-showcase/fa26-cday-program.php

Temporal obfuscation tests whether a large language model (LLM) reasons about numbers1 or recalls market history. Our earlier work reported that an LLM detects persistent dealer-gamma2 regimes in 30-day windows of obfuscated daily gamma exposure (GEX) of SPY, with a sharp rise3 after 2023. We re-run that test on the complete data. The published 2024–2025 inputs lacked the put4 leg of the options chain. We assembled six daily series (recovered, archived or rebuilt), replayed the5 deterministic rule, bridged o4-mini to gpt-6-luna on the published inputs, reran the 2024 and 20206 baselines with relative day labels, and repeated seven arms with an open-weights model. On the7 published inputs the rule detects all 501 windows of 2024–2025; on the complete series it detects 608 of 502 (12.0%), and gpt-6-luna 63 of 502 (12.5%). The two models agree on 99.6% of the published9 windows (Cohen’s kappa 0.993), and gpt-6-luna follows the rule in 99.7% of 4,691 windows (kappa10 0.985); gpt-oss-20b agrees with gpt-6-luna on 99.4%. Removing same-day expiries gives 14.7%, not11 the published rate. With relative labels gpt-6-luna detects 80.7% of the 2024 baseline windows, against12 81.2% with dates. The complete series shows no post-2023 increase that the data distinguish from13 zero