Abstract
This chapter presents Multi Source Attention Long Short Term Memory (MSA-LSTM), a deep learning framework for real time financial forecasting that fuses three heterogeneous data streams historical OHLCV price data, financial news sentiment, and macroeconomic indicators within an Attention enhanced Long Short Term Memory (LSTM) network. A sliding window procedure converts these streams into synchronized temporal sequences, and the attention layer dynamically re weights time steps and features so the model can be inspected as well as used for prediction. The model performance is approximately 73% and 39% better than an ARIMA baseline and attention free LSTM respectively. Based on these results and an ablation analysis presented later in the chapter, as well as a statistical significance assessment performed later in the chapter, we can conclude that sentiment and macro economic context and attention based temporal modeling can be combined in a way that results in a measurable and reproducible improvement over single-sources baselines.

This publication has 12 references indexed in Scilit: