Deep learning for financial time-series forecasting

What it is

Built and evaluated LSTM models forecasting daily returns for US equities. Primary programmer in a team of three — I owned data cleaning, feature construction, model training and the walk-forward evaluation loop.

What I took from it

• An in-depth look at where LSTMs fail on financial series (non-stationarity, low signal-to-noise) and which adaptations actually help.
• Model testing, visualisation and analysis for the final paper, which received an A* rating from the supervisor.