Abstract
We study daily stock-return prediction in European defence equities using a rolling, walk-forward LSTM evaluated under strict time ordering. The model forecasts next-day log returns for Rheinmetall AG (RHM.DE) and BAE Systems (BA.L) over 2020–2025 from stationary, returns-based features (multi-horizon returns; 5/20/60-day volatility; high–low range; volume change and volume/20-day mean; centered RSI; normalized MACD; 5/20-day momentum). At each refit we use a 3-year rolling window, a 60-day lookback, z-score features on the train fold only, Huber loss with Adam (5e-4), dropout 0.2, early stopping, and no shuffling. We convert forecasts into trades with a single entry threshold (τ) calibrated once on the first half of the out-of-sample (OOS) period to maximize Sharpe under turnover-based 10 bps round-trip costs, then held fixed on the second half. On the post-calibration test half, Rheinmetall achieves ~100.6% total return with Sharpe 2.24, exceeding same-dates buy-and-hold (~95.7%, Sharpe 1.76). For BAE Systems, results are marginal relative to buyand-hold (LSTM ~22.5%, Sharpe ~1.05 vs ~37.5%, Sharpe ~0.94). The findings show that a small, transparent LSTM combined with a calibration-then-test design and realistic transaction-cost accounting can yield tradable signals for some assets while highlighting asset-dependence of predictability.