Stock return prediction is one of those problems where every new paper claims progress against baselines while the market itself remains stubbornly efficient. A new study from MSA University in Egypt proposes a hybrid LSTM-XGBoost architecture for multi-horizon return forecasting across 14 U.S. equities. The approach is clean and the authors are honest about their results, but the numbers tell a familiar story about the limits of purely technical models.
The Two-Stage Architecture
The pipeline has two stages. First, a two-layer stacked LSTM with 64 hidden units processes 60-day sliding windows of five sequential market features: Volume Ratio, daily return, RSI, MACD, and MACD signal line. The LSTM compresses this sequence into a 64-dimensional temporal embedding. Second, this embedding is concatenated with 14 hand-crafted technical indicators to form a 78-dimensional feature vector fed to an XGBoost regressor.
The 14 technical indicators span five groups: moving averages (5-day, 10-day, 15-day, 30-day SMA, and 9-day EMA), momentum oscillators (RSI, MACD line, MACD signal), volatility (20-day rolling standard deviation of returns), momentum features (10-day and 21-day rate of change), and volume features (one-day volume change, 20-day volume ratio, MACD histogram). All indicators are delayed by one day to prevent data leakage and scaled to the range [-1, 1] using MinMaxScaler fitted on training data only.
The LSTM uses Adam optimization with learning rate 5e-4, L2 weight decay of 1e-4, gradient clipping at norm 1.0, and dropout of 0.5 between layers. Training runs up to 300 epochs with early stopping after 40 epochs without validation improvement. XGBoost is tuned via 3-fold cross-validation grid search over n_estimators (100 to 400), learning rate (0.001 to 0.05), max depth (4 to 10), and gamma (0.001 to 0.02). One independent XGBoost model is trained per prediction horizon.
The dataset covers 14 U.S. equities from six sectors (technology: AAPL, NVDA, MSFT, GOOGL; financials: JPM, GS; healthcare: JNJ, PFE; energy: XOM, CVX; consumer: AMZN, WMT; industrials: BA, CAT) using daily OHLCV data from 2010 onward. All stocks are pooled for training with strict chronological splits (70% train, 15% validation, 15% test) and per-stock MinMax scaling.
Results: Marginal Gains Over XGBoost Alone
On the 30-day horizon, the hybrid model achieves a test RMSE of 0.0949, compared to 0.0951 for XGBoost-only and 0.2799 for LSTM-only. The hybrid roughly triples the LSTM's performance but barely edges out XGBoost alone. The R-squared values are telling: the hybrid scores -0.003, XGBoost scores -0.008, and LSTM scores -7.73. All are negative, meaning none of the models beats a simple mean predictor for return magnitude.
The per-stock breakdown shows the hybrid wins on 7 of 14 stocks by RMSE and ties on one, while XGBoost wins on the remaining 7. Where the hybrid excels is on high-volatility growth stocks: NVDA (0.1603 vs 0.1628), AMZN (0.1246 vs 0.1270), and cyclical names like WMT (0.0705 vs 0.0714) and JPM (0.0781 vs 0.0787). Where it does not help is low-volatility defensive equities: JNJ (0.0522 vs 0.0469 for XGBoost) and PFE (0.0680 vs 0.0660 for XGBoost). The benefit is selective and sector-dependent rather than uniform.
Directional Accuracy: The Base Rate Problem
Directional accuracy rises with horizon length: 57.0% at 30 days, 64.4% at 90 days, 84.9% at 252 days, and 97.6% at 365 days. The authors benchmark this against a naive always-positive predictor, since the test period is dominated by a bull market. The base rate of positive returns is 59.9% at 30 days, 72.4% at 90 days, 91.2% at 252 days, and 97.8% at 365 days.
The model does not outperform the naive predictor at any horizon. The gap is negative across all four windows: -2.9 percentage points at 30 days, -8.0 at 90 days, -6.3 at 252 days, and -0.2 at 365 days. At 365 days, 12 of 14 stocks reach above 90% directional accuracy, with several hitting 100%. The authors are explicit that this largely tracks the high base rate of positive long-horizon returns in the sample, not genuine forecasting skill. They treat the above-base-rate gap at short horizons as the more informative signal, though even there it is negative.
Investment Scoring
The paper proposes a composite investment score combining multi-horizon predictions: S = 20 times the 30-day predicted return + 30 times the 90-day return + 50 times the 252-day return + a consistency bonus of plus or minus 10 when all three short-to-medium horizon predictions agree in sign, minus 100 times the 20-day realized volatility. At the test period end in mid-2016, the scoring ranks Goldman Sachs and Microsoft as strong buys, while flagging NVDA and AMZN as strong sells. The authors note this corresponds to mid-2016, before NVDA's GPU-driven growth surge, illustrating that purely technical models cannot anticipate regime changes.
The Honest Limitations
The persistently negative R-squared values across all models and horizons are the most important finding. None of the models beats a simple mean predictor for return magnitude. The directional accuracy at long horizons is inflated by the bull market base rate. The evaluation window ends in mid-2016, preceding structural shifts that would have invalidated many of the predictions. Long-horizon samples overlap, so effective independent observations are far smaller than the raw test-set count suggests. No transaction-cost-adjusted backtest is provided.
The LSTM temporal embeddings carry complementary information not captured by technical snapshot features alone, but the global-level margin over XGBoost is small. The hybrid benefit is best characterized as selective and sector-dependent. For practitioners, the takeaway is that combining sequential deep learning with gradient boosting can help on specific stock profiles, but the gains are modest and the fundamental challenge of beating a mean predictor on return magnitude remains unsolved.