Key points are not available for this paper at this time.
Background Developing economies offer compelling investment opportunities but are often overlooked due to perceived instability. This study challenges the bias toward developed markets by exploring whether deep learning, applied to denoised high-frequency data, can unlock superior intraday returns in developing economies. Methods We compared three developed markets (the USA, Japan, and Singapore) against three developing markets (India, Brazil, and Malaysia). A standardised pipeline retrieved minute-level data for ~20 stocks per market, selected via K-means clustering based on market capitalisation and beta. Final portfolio inclusion was validated based on model performance (adjusted R 2 scores and minimal overfitting). Data was denoised using wavelet transforms, autoencoders, and Kalman filters sequentially, with hyperparameters optimised via Optuna to filter microstructure volatility through noise. We trained multiple deep learning architectures, including LSTM, Bi-LSTM, GRU, TCN, Transformers, Conv1D-LSTM and CNN-LSTM, employing walk-forward validation to forecast the next 150-minute prices. Results The models showed high predictive power, with adjusted R 2 values ranging from 78% to 95% (average 85%) and minimal overfitting (
Sami et al. (Fri,) studied this question.