Machine learning research
Research on LLM evaluation, learning from simulated data, and time-series prediction.
I’m an independent machine learning researcher and a Computer Engineering undergraduate at Middle East Technical University (METU). All research conducted independently, without an advisor. Independent research: Spring 2026 - Present.
Papers, experiments, and open research data
Research study · First author · 2026
Designed a study of 11 LLMs and ~99,000 text assessments to test whether 0-10 writing scores reflected deliberate reductions in text quality.
Compared five scoring corrections. Probability-weighted scoring increased correct rankings of better versus degraded text from 51.8% to 64.9% across nine models.
7th ACM International Conference on AI in Finance (ICAIF 2026)
Accepted after peer review; selected for oral presentation · Not yet published; forthcoming in ICAIF 2026 proceedings
Built a method to estimate what a crowd believed before an event, separating that belief from distortions caused by trading behaviour. Combined models of trader behaviour according to their estimated errors.
Reduced worst-case regret across unseen market mechanisms by about 80% (from 8.8% to 1.8%) relative to direct black-box prediction.
On 1,500 simulated markets from ten types absent from training, recovered beliefs with 12.8% less error than averaging prices after each trader's first trade.
Across 98 real markets, recovered beliefs with 10.2% less error (RMSE) than averaging the prices after each trader's first trade. Tested against pre-trading surveys from projects excluded from training.
7th ACM International Conference on AI in Finance (ICAIF 2026)
Accepted after peer review; selected for oral presentation · Not yet published; forthcoming in ICAIF 2026 proceedings
Built a benchmark to test whether randomness in price changes helps predict future returns. Compared neural-network, linear and tree-based models across 16 market series.
Ran 113 tests across individual markets and years, selecting models on earlier data and measuring their forecasts and trading returns in the following year.
7th ACM International Conference on AI in Finance (ICAIF 2026)
Accepted after peer review · Not yet published; forthcoming in ICAIF 2026 proceedings
Trained a Transformer on simulated markets to learn patterns hidden behind observed prices, then adapted it to forecast price variability with five real training examples.
Used a Gaussian-process model to choose simulations likely to improve real forecasts. Compared identical models trained with and without those extra labels.
Against this control, reduced volatility forecast error (MSE) by 4.7% to 7.0% in three of five markets.
2nd International Online Conference on Risk and Financial Management (IOCRF 2026)
Accepted; selected for oral presentation · November 2026
Developed models that weight recent rate changes more heavily to predict whether central banks will cut, hold or raise rates, using past policy decisions alone.
Evaluated an exploratory pipeline on 2,490 post-hold forecasts across 37 economies, achieving 0.614 macro-F1 for cut/hold/hike prediction.
Dataset manuscript under review at Scientific Data
Built a public ML dataset from 2,872 Bank of England and Federal Reserve documents, with graphs of linked information and numerical text representations.
Checked records for consistency and linked them to source documents so others could verify the data.
I’m completing a B.E. in Computer Engineering at Middle East Technical University, with graduation expected in 2027. My coursework includes machine learning, statistics, scientific computing, and a graduate-level Probability Theory course currently in progress.
All research conducted independently, without an advisor.
Coursework: Graduate-level Probability Theory (in progress); Machine Learning; Statistics for Computer Engineers; Linear Algebra; Calculus I-III; Scientific Computing; C Programming; Data Structures; Algorithms.
Programming: Python, C++, SQL, TypeScript.
ML and analysis: PyTorch, NumPy, pandas, SciPy, scikit-learn; model training, simulation, controlled comparisons, and statistical evaluation.
Tools: Git, Linux, LaTeX.
I’m open to research collaborations and roles in machine learning, forecasting, and AI safety.
idrees.aziz@metu.edu.tr ↗