Feature Engineering Blunders in Deep Learning Price Prediction
How Bad Features Guarantee Poor Forecasting Performance
Question whether technical indicators add information beyond raw prices. Most oscillators derive from simple moving averages and price differences already present in your sequence. Redundant features increase dimensionality without improving predictions.
Calculate indicators using only past data at each timestamp. Implementations that compute moving averages across entire datasets leak future information. Your training accuracy looks impressive but deployment predictions fail immediately.
Standardize features using rolling statistics, not global parameters. Markets shift volatility regimes frequently. Features normalized using 2019 parameters become meaningless in 2023 conditions. Compute means and standard deviations over trailing windows.
Avoid correlated feature pairs that confuse gradient descent. Including both close prices and simple moving averages creates collinearity. Your model struggles to assign proper weights during training.
Document the rationale behind each engineered feature. Freelancers often add dozens of indicators hoping something works. This shotgun approach produces brittle models sensitive to minor market changes.
Test feature importance through ablation studies. Remove features individually and measure validation performance degradation. Features that barely impact results add computational cost without benefit.
model = Sequential([ LSTM(128, return_sequences=True), Dropout(0.2), Dense(1, activation='linear') ])