Preregistration: Sequential Reflective Memory and Behavioral Heuristics in Autonomous LLM Trading Agents on the Warsaw Stock Exchange (Season 2)

Abstract

Official preregistration for Season 2 of the autonomous agent trading benchmark on the Warsaw Stock Exchange (GPW), investigating the emergence of human-like behavioral heuristics in Large Language Models (LLMs) endowed with sequential reflective memory. The benchmark evaluates 5 frontier language models (GPT-6.1 Sol, Claude Opus 5.5, Gemini 3.8 Flash, Mistral Large 4, DeepSeek V4.1 Flash) across 20 WIG20 equities over 12 weeks (N=2,200 panel observations). Using a longitudinal Difference-in-Differences (DiD) framework against a parallel memoryless control arm, this study preregisters four confirmatory behavioral hypotheses rooted in canonical behavioral economics and decision theory (Kahneman & Tversky 1971, 1973, 1974, 1979; Bell 1982; Samuelson & Zeckhauser 1988; De Bondt & Thaler 1985): P1 (Intertemporal Path-Dependency): Emergence of choice hysteresis breaking Markovian independence beyond stochastic within-model noise. P2 (Regret Aversion & Post-Loss Omission Bias): Asymmetric shift toward cash status quo following directional loss. P3 (Heuristic Overreaction & Gambler's Fallacy): Directional position reversal following forecast error seeking mean-reversion. P4 (Inaction Inertia & Status Quo Trap): Self-reinforcing inertia of unpenalized cash positions. This deposit freezes the experimental design, formal epistemic operators, prompt templates (SHA-256 verified), and deterministic memory-rendering specifications prior to the live trading window commencing October 12, 2026.