"It Looks All the Same to Me": Cross-index Training for Long-term Financial Series Prediction

Battle Count: 133

Directly relevant to quantitative trading as it demonstrates that models trained on one market index can predict other market indexes with comparable accuracy, supporting cross-market signal generation. The 30-day prediction horizon is useful for medium-term trading strategies. The finding that larger indexes 'explain' smaller ones (NIKKEI/NASDAQ vs DOW/DAX) has practical implications for cross-market arbitrage and hedging. However, the study focuses on index-level prediction rather than individual stock trading, and the MAPE values (5-10%) may be too high for direct trading application without additional risk management.

Standard ANN architectures (LSTM, GRU, CNN, ReLU, Logistic) are readily available in MATLAB Deep Learning Toolbox. However, custom implementations (KGate, GMDH, SCNN, RBF) require manual coding from scratch. The cross-training framework with session partitioning, model parameter resets, and Wilcoxon statistical testing adds complexity. MATLAB-specific implementation may require porting for Python-based workflows. The circular cross-training across four indexes with 34 sessions each multiplies computational requirements.

GitHub repository with source code provided (https://github.com/Selitskiy/LTTS). Data sourced from tradingeconomics.com. Specific hardware (Tesla K80 GPUs, QuadroPro K6000), software (MATLAB 2022a, R 4.2.1), and hyperparameters (adam optimizer, 0.01 learning rate, batch size 32, 1000 epochs) are documented. Custom ANN implementations (KGate, GMDH, SCNN, RBF) are coded from scratch and available. However, MATLAB dependency and specific hardware may limit exact reproduction.