Predicting stock prices using a TensorFlow LSTM (long short-term memory) neural network for times series forecasting
View the Project on GitHub JordiCorbilla/stock-prediction-deep-neural-learning
Price-level forecasts must be compared against the last-observation baseline:
prediction(t) = actual(t-1)
The package reports:
RMSE skill = 1 - RMSE(model) / RMSE(naive)
Positive skill indicates lower RMSE than the naive predictor; zero is equivalent; negative values are worse.
Random train/test shuffling is inappropriate for this project. Use chronological splits and, when comparing research variants, expanding-window walk-forward evaluation.
quant_forecast_lab.validation.expanding_window_splits provides deterministic fold boundaries while leaving model training policy to the caller.
Directional accuracy is measured relative to the information available immediately before the forecast:
sign(predicted_price[t] - actual_price[t-1])
vs
sign(actual_price[t] - actual_price[t-1])
The historical stochastic trajectories are scenario simulations obtained by perturbing a point forecast with noise estimated from recent history. They should not be interpreted as calibrated prediction intervals.
quant_forecast_lab.uncertainty provides a symmetric conformal interval based on held-out residuals. The inference script applies one-step residuals to recursive future forecasts and labels the result a one-step residual band, since horizon-specific coverage has not been established. Quantile heads are also nominal until empirical coverage is checked on untouched observations.
Prediction error and trading value are different questions. When a forecast is translated into a directional strategy, report transaction costs, turnover, Sharpe-like diagnostics and drawdown in addition to forecasting metrics.