Greek stock market forecasting: a news-based strategy using LLMs

This thesis determines if Large Language Models (LLMs) can forecast short-term stock returns in Greece through categorizing real-time news headlines into a signal to translate into systematic trading strategies. This study utilized the Reuters headlines obtained from LSEG with the Only Important Relevance Filter (402 items from July 18, 2024 to July 1, 2025) and the one-minute transaction prices for Athens Stock Exchange (ATHEX) stocks for those which were aligned in Europe / Athens Time. The headlines were categorized using GPT-4 (deployment gpt-4-0613) who had a training cut-off of September 2021; thus, this application is a strict out-of-sample evaluation. Model outputs were mapped to YES / NO / UNKNOWN, with unknowns removed from trading and retained solely for descriptive purposes. The principal assessment employed trading simulations that convert classifications into actionable positions based upon two different execution mechanisms: intra-day signals are executed one minute after they are published and then closed at the same-day closing price; while after-hour signals are executed at the next trading day's opening price and closed at the same day's closing price. There are three trading strategy implementations - long-only (YES), short-only (NO), and long-short (YES/NO), each of which have been developed in both single trade per day compounding and fixed-capital-per-trade without compounding. Per-trade statistics and risk adjusted metrics were used to assess performance with robustness checks for transaction costs. Although the simulations assume zero market impact and frictionless short selling, and do not include dynamic or state dependent transaction costs, the reported economic performance will be considered as a theoretical upper limit when assessing these real world frictions. In addition to the simulation analysis, a high frequency event study examined both pre-event drift (with raw returns to determine if there is any information leakage) and post-event price discovery (with market adjusted abnormal returns), with segmentation of classification and session, and with bootstrap confidence intervals. Results demonstrate that the long-short trading strategy yields the highest performance, with returns that are stable even to moderate levels of transaction cost. Importantly, a timing asymmetry has emerged where after-hours signals possess greater predictive accuracy and per-trade returns than intra-day signals. The single-trade-per-day (compounding) rule implicitly selects for these strong after-hours signals, generating superior efficiency (100.44% over 154 trades) compared to the no-compounding variant (109.12% over 294 trades). In conjunction with the trading results, the event study demonstrated a statistically significant negative drift in response to NO signals and a robustly positive Long-Short (YES-NO) CAR. The evidence clearly indicates a dominant timing asymmetry: after-hours signals are highly significant, while intra-day signals are statistically indistinguishable from random noise. These findings support evidence consistent with a multilayered asymmetry in sentiment, effect, and timing. This thesis presents four primary contributions: MSc Thesis AUEB ii Empirical: The first intraday LLM-news forecasting study in ATHEX with strictly out-of-sample classifications and produces economically significant and statistically significant returns. Methodological: A unified evaluation that combines execution-aware trading simulations with an intraday event study and explicit transaction-cost robustness up to 15 bps in a high-friction market. Analytical: A "Triple Asymmetry" framework—combining sentiment (the potency of NO signals), effect (the volume of YES signals), and timing (the dominance of after-hours signals)— which organizes all of the findings. Extension of Contextual: A small-market, microstructure-aware confirmation and refinement of López-Lira & Tang (2023), demonstrating how opening auction price discovery conditions the effectiveness of LLM-generated signals.