Ahmed Mahrous
Ph.D. candidate in Computer Science
King Abdullah University of Science and Technology (KAUST)
Thuwal, Saudi Arabia
[email protected]
I work on artificial intelligence and decentralized finance. I study how learning algorithms, incentives, and strategic behavior shape markets, while also exploring how AI models can make reliable predictions.
I am a Ph.D. candidate in Computer Science at KAUST, advised by Roberto Di Pietro. My research combines mathematical modeling, empirical analysis, and machine learning, with work on AI methodology as well as applications to stablecoins, prediction markets, financial manipulation detection, and investor sentiment.
[CV] [Google Scholar] [ORCID] [Research group] [GitHub]
Selected publications and accepted papers.
Ahmed Mahrous and Roberto Di Pietro
NeurIPS 2026 · AcceptedICORE A*
Under the stated sensitivity and curvature assumptions, a reflexivity matrix with spectral radius below one guarantees convergence of repeated retraining. Above one, we construct noisy systems where every local-information learner incurs expected cumulative excess loss growing linearly for horizons up to order 1/(ρ(Γ) − 1).
Ahmed Mahrous, Maurantonio Caprolu, and Roberto Di Pietro
International Review of Economics & Finance, 111, 105665ABDC rank: AIF 7.615% acceptance
A systematic review of stablecoin research, connecting economic, technical, and policy questions and identifying open challenges.
Ahmed Mahrous, Jens Schneider, and Roberto Di Pietro
ACM Transactions on the Web, 20(3), Article 35Q1IF 4.9
On about 528,000 emoji-containing StockTwits posts, emoji-only sentiment models reach F1 ≈ 0.75, versus ≈ 0.88 for text-and-emoji models, with substantially lower computational cost. Financial emoji usage also differs from general social-media usage.
Ahmed Mahrous and Roberto Di Pietro
Findings of ACL 2026
Across financial communities in four languages, emoji frequencies vary while meanings and sentiment polarity remain largely stable. Including emojis reduces sentiment-transfer gaps relative to text-only models, though cross-language transfer remains the hardest setting.
Ahmed Mahrous and Roberto Di Pietro
ACM International Conference on AI in Finance (ICAIF 2026)Oral
Across 802,863 resolved Polymarket markets, launch-time contract information predicts trading activity, but demand prediction weakens sharply across periods of platform change. Dispute scores help rank groups of markets; the low dispute rate limits predictions for individual contracts.
Ahmed Mahrous and Roberto Di Pietro
IEEE International Conference on Blockchain and Cryptocurrency (ICBC 2026)18% acceptance
Using over 280,000 Telegram posts, we benchmark pump-announcement detection and extraction of the target coin and exchange. LightGBM reaches F1 = 0.79 and BGE-M3 reaches 0.83; LLMs perform best on target extraction.
Ahmed Mahrous and Roberto Di Pietro
ACM International Conference on AI in Finance (ICAIF 2025)Oral15.5% acceptance
A classifier trained only on synthetic pump-and-dump time series achieves 79.75% accuracy on real data. Adding 20% real training data raises accuracy to 88.1%, while simple Gaussian sampling outperforms the deep generative models tested.
Ahmed Mahrous, Maurantonio Caprolu, and Roberto Di Pietro
IEEE International Conference on Blockchain and Cryptocurrency (ICBC 2025)18.2% acceptance
A structured account of economic and technical research on stablecoins.
Ahmed Mahrous and Roberto Di Pietro
International Conference on Blockchain Computing and Applications (BCCA 2025)
Across over 160,000 NFTs in 32 collections, collection identity inflates apparent visual price predictability. Controlling for it leaves weak generalization across collections, although visual features retain modest predictive power within some collections.
Lanouar Charfeddine and Ahmed Mahrous
Finance Research Letters, 67, 105861ABDC rank: AIF 7.1
An empirical study of coin-level and market-level factors associated with cryptocurrency pump-and-dump schemes.
Ahmed Mahrous, Jens Schneider, and Roberto Di Pietro
Fourth International Conference on Intelligent Data Science Technologies and Applications (IDSTA 2023)Best Student Paper Award
A study of emojis in sentiment analysis of financial microblogs.
ICORE: 2026 conference ranking. ABDC: 2025 Journal Quality List. Q1: Journal Citation Reports 2025 edition. IF denotes Journal Impact Factor; figures are publisher-reported and were checked in October 2026. Acceptance rates are overall rates, except ICAIF 2025, where 15.5% refers to oral papers.
Sources: ICORE · ABDC · ACM TWEB · IREF · Finance Research Letters.
Teaching assistant for Introduction to Cybersecurity and Introduction to Cryptography, M.Eng. program, KAUST.
Conference reviewer: ICAIF, NeurIPS, and ICBC.
Journal reviewer: Financial Innovation; Blockchain: Research and Applications; Computer Networks.