Sentiment, price structure, and momentum disagree often enough that fusing them beats trading any one of them alone.
workedThe first trading research in this ledger, and the project where my markets side and my engineering side stopped being separate careers.
The model fuses three families of signal: FinBERT sentiment over news flow, AlphaVantage market data, and classic technical structure (MACD, RSI, Donchian channels). A multi-factor fusion layer weights them differently per trading style: swing, scalping, and intraday each get their own signal blend.
I backtested it over five years and judged it the way a risk desk would, on Sharpe ratio and maximum drawdown rather than headline return. It averaged 2.81% monthly over the window.
What worked is recorded above. What I remain skeptical of: backtests flatter everyone, and regime change is undefeated. The model earns its ‘worked’ tag on the test window, not on a promise about the future.