FinChart-Multimodal: A Dataset for Context-Injected Financial Chart Understanding with Aligned OHLCV Time Series

Understanding financial charts requires integrating visual patterns with numerical time series, but existing datasets lack pixel-level alignment between chart images and underlying data. We introduce FinChart-Multimodal, a dataset of 100,000 financial charts with pixel-aligned OHLCV time series data. The key innovation is deterministic generation providing exact pixel-to-timestamp and pixel-to-price mappings without OCR. Charts are generated from real market data using a reproducible pipeline with extensive visual diversity (23+ color schemes, 30+ backgrounds, 16 grid styles, 15-20 font families) while maintaining alignment guarantees. The dataset spans 105 tickers across 10 sectors, 5 chart types, and 8 time periods (1mo to 5y), with content-based uniqueness checking incorporating visual parameters ensuring < 0.001% collision probability. We propose three benchmark tasks: temporal-visual alignment learning, context injection effectiveness, and crossmodal retrieval. The dataset, generation pipeline, and benchmark task definitions will be publicly released.