Year: 2025
Type: dissertation
Abstract: While anomaly detection in the stock market is a topic of critical importance, it poses quite a few challenges due to the volatility, non-stationarity, and label scarcity in financial time series data. Prior work in diffusion-based anomaly detection focuses on reconstruction and imputation techniques, and practical applications in the stock market remain underexplored, with limited evaluation. We study an unsupervised, forecast-based detector for financial instruments: a conditional denoising diffusion forecaster with deterministic and stochastic DDIM sampling, and we compare its performance with LSTM and Temporal Convolutional Network baselines on daily S&P 500 OHLCV data. With labels scarce, we develop a synthetic anomaly injection protocol focused on point and contextual anomalies, run k=20 repeats, and report average precision, recall, and F1 scores. For a robust evaluation and comparison of the models, we run the experiments for both stationary(daily returns) and non-stationary (closing price of the next day) target variables with various anomaly injection procedures. We report that training on the daily returns (stationary) outperforms the close (non-stationary), with the best stationary-trained model being the LSTM with attention, achieving F1=0.516, against the best model trained on the closing price (stochastic DDIM) reaching F1=0.482. As a result, we conclude that the increased forecasting power of the diffusion models does not necessarily translate to enhanced anomaly detection, indicating data stationarity as a critical factor. Our results demonstrate the strong potential of denoising diffusion models as forecasters, and under the right target choices, as detectors. At the same time, future work should refine the injection protocol to better reflect realistic market microstructures, while more sophisticated thresholding techniques could further increase model performance.
Author: Tsouderos, Georgios
Institution: Uppsala University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7111891777
