Leveraging AI for Detecting and Exploiting Inefficiencies in Financial Markets: A Focus on Emerging Markets

Abstract

This thesis investigates the application of deep learning models to predict directional returns in the context of financial time series forecasting for emerging markets, with a specific focus on the United Arab Emirates equity market. Using a dataset comprising daily OHLCV data for the twenty most actively traded UAE stocks from 2018 to 2025, we construct a rich set of features encompassing technical indicators, statistical measures, and structural signals. Two labeling methodologies are employed: a binary classification based on next-day returns and an event-driven approach using the Triple Barrier Method.
We evaluate the performance of a Deep Belief Network coupled with a Multilayer Perceptron (DBN+MLP) and the Tabular Prior-Data Fitted Network (TabPFN), benchmarking against established models such as CatBoost, TabNet, and LSTM. Results show that TabPFN achieves superior performance in both classification accuracy and trading profitability under the next-day return labeling scheme. Furthermore, we implement a meta-labeling strategy in which the predictions of a primary model are filtered by a secondary model trained on the triple barrier labels. This dual-layer setup demonstrates improved Sharpe ratios and win rates, particularly when the secondary model is a probabilistic architecture like TabPFN or CatBoost. Our findings suggest that combining deep learning with meta-labeling offers a robust framework for building interpretable, risk-aware alpha-generating strategies in emerging markets.

Date of Award	2025
Original language	American English
Supervisor	Yerkin Kitapbayev (Supervisor)