Reproduce “VertiFuseX: Generalizable Financial Forecasting via Multi-Stream Temporal Fusion” on OpenResearch

About this paper

Aashish Bohra, Vivek Vijay

VertiFuseX forecasts next-day equity-index closing prices by jointly combining intermediate representations from LSTM, bidirectional LSTM, stacked LSTM, and feed-forward streams. Using a fixed configuration and chronological evaluation, it outperforms directly implemented recurrent and ARIMA baselines across four U.S. indices, while cross-market tests show lower reported errors across 33 comparable dataset–metric cases.

Key insights

• Where fusion occurs matters: learned combinations of intermediate features retain more useful temporal information than averaging branch predictions or fusing only final-layer outputs.
• Different recurrent branches provide complementary views of the same price history, while the DNN stream adds a non-sequential representation of the lag window.
• Saliency maps tend to emphasize intermediate lags, roughly 9–16 trading days before prediction, but this indicates model sensitivity rather than causal market structure.

Summary by alphaXiv.