Manipulating stock graphs using LLMs : Influencing predictions of stock trends in candlestick graphs

The thesis studies how humans interact with stock graphs that are manipulated using Generative AI. The thesis contributes to research on human decision-making based on data visualizations and interactions between humans and Generative AI visualizations by attempting to answer the research question: Can large language models be used to generate manipulated stock price visualizations that influence human estimates of future prices? The study design is exploratory, and the goal is to contribute with new methods and initial information on the subject. A novel application for manipulating candlestick graphs is created. A stock prediction task is formed around the manipulated images to study whether predictions for these images (positive, neutral, negative) differ from the predictions for images without manipulations (clean). The stock prediction experiment recruitment yielded a small diverse sample of participants (N=12). Due to the small sample size and scope of the thesis, its contribution is limited to effect sizes and their practical implications. The study found a systematic underestimation in the predicted next candlestick compared to the last shown candlestick. This effect could be a result of a central tendency in predictions, as most stocks showed positive development during the visualization period, and for some stocks with negative development, there appeared to be an overestimation compared to the last candlestick shown. The effect sizes for the manipulations compared to the clean condition were estimated using Cohen’s d for the mean difference to clean condition. The strongest effect was found for images including positive manipulation and the weakest for images with negative manipulation. The estimates for images with neutral manipulation were similar to the positively manipulated images. Similar results were found when analyzing the frequency of estimates above and below the clean reference, where the positive and neutral manipulations had higher odds of being above clean condition, and the negative manipulation had higher odds for being below the clean condition. Analyzing the effect sizes, it appears that the predictions could be influenced by the manipulations created by LLMs. However, this study was unable to show the effects using robust statistical models, possibly due to small sample size, broad participant base, and inconsistencies in manipulations. Future research should expand on these results by improving quality control in the manipulation application, using a more targeted participant recruitment, and using a larger sample. Future studies on understanding how humans interact with data visualizations and generative AI generated content will be needed as the visualization complexity and generative AI deployment increases.

Grantor dc:publisher

Helsingin yliopisto

Year dc:date.issued

2025

Author dc:creator

• Laitinen, Jussi Arto Oskari

• visualisointi
• vuorovaikutus
• kuvamanipulaatio

Statement dc:rights

• In Copyright 1.0

Language dc:language.iso

eng

Handle dc:identifier.uri

http://hdl.handle.net/10138/601114

OAI identifier oai:identifier

oai:helda.helsinki.fi:10138/601114