A VLM-Based Framework For Technical Analysis

Technical analysis studies whether recurring, visually recognizable geometric patterns in price series are associated with future returns. Most large-scale empirical studies to date rely either on hand-coded pattern templates or on black-box predictive models with limited interpretability. We introduce a pattern-discovery framework that uses a modern vision–language models (VLM) to generate structured semantic descriptions of price charts, which are then organized into geometric pattern families through unsupervised clustering. Applied to CRSP equities from 2000–2024, the resulting pattern families are visually coherent and economically interpretable without fine-tuning. We identify several VLM-discovered patterns whose associated return differentials remain statistically significant under conservative out-of-sample evaluation. More broadly, this approach illustrates a workflow for empirical financial modeling that treats large language models as descriptive instruments for extracting structure from noisy market data.