When the simple model wins
For my graduate thesis I ran a benchmark study on the FinMultiTime dataset, asking a question that sounds like it has an obvious answer: does stock prediction improve when you fuse modalities — price time series, news sentiment, and SEC filings?
What I built
I evaluated 10+ models across standalone, multimodal, and ensemble categories. The architecture I explored used LSTM for prices, FinBERT for sentiment, TabNet for filings, a graph neural network for inter-sector relationships, and attention-based late fusion to combine them.
The surprising result
The intuition going in was "more signal must mean better predictions." But a standalone LSTM on price data alone stayed highly competitive with the fused multimodal models. The added modalities did not deliver the clear, consistent gains the complexity implied they should.
The lesson
Complexity has to be justified by clear empirical gains — not by how sophisticated the pipeline looks on a diagram. A multimodal system carries real costs: more data plumbing, more failure modes, more things to maintain. If a single-modality model matches it, that simplicity is a feature, not a shortcoming. I packaged the findings into a research paper and a technical report.