An Evaluation of Deep Learning Models for Stock Market Trend Prediction
An Evaluation of Deep Learning Models for Stock Market Trend Prediction
This paper evaluates deep learning models for short-term stock market trend prediction using daily and hourly closing prices from the S&P 500 index and the Brazilian ETF EWZ. The authors propose xLSTM-TS, an adaptation of the xLSTM architecture that incorporates exponential gating and novel memory structures (sLSTM and mLSTM) to better capture long-term dependencies. They compare xLSTM-TS against five state-of-the-art models (TCN, N-BEATS, TFT, N-HiTS, TiDE) implemented via the Darts library. A key preprocessing step is wavelet denoising, which smooths the price signal and reduces noise, leading to consistent improvements in directional prediction accuracy across all models.
Experiments show that xLSTM-TS outperforms all benchmarks, achieving the highest test accuracy (72.82% on EWZ daily) and F1 score (73.16%). Denoising is shown to boost performance by 2-5% on average. The study also finds that daily data yields better predictions than hourly data, and that the Brazilian ETF (EWZ) is more predictable than the S&P 500, likely due to higher volatility and market inefficiency. The paper provides a comprehensive comparison of modern deep learning architectures for financial time series, highlighting the potential of xLSTM-based models and the importance of data preprocessing.
Highlights
- 1Introduces xLSTM-TS, an adaptation of the xLSTM architecture for time series forecasting, which outperforms state-of-the-art models.
- 2Applies wavelet denoising as a preprocessing step, significantly improving prediction accuracy for stock market trend forecasting.
- 3Evaluates multiple deep learning models (TCN, N-BEATS, TFT, N-HiTS, TiDE) on daily and hourly data from S&P 500 and EWZ.
- 4Achieves a test accuracy of 72.82% and F1 score of 73.16% on the EWZ daily dataset using xLSTM-TS.
- 5Demonstrates that denoising consistently enhances directional prediction performance across all models.
Methods
- MWavelet denoising using discrete wavelet transform (DWT) with Daubechies wavelet (db4) and soft thresholding.
- MxLSTM-TS model: an extended LSTM with exponential gating and sLSTM/mLSTM blocks, adapted for time series.
- MBenchmark models from Darts library: TCN, N-BEATS, TFT, N-HiTS, TiDE, with hyperparameter tuning.
- MEvaluation metrics: Accuracy, Recall, Precision, F1 Score, MAE, RMSE, RMSSE, MASE.
Results
- RxLSTM-TS consistently outperforms all other models across datasets, achieving highest accuracy and F1 scores.
- RWavelet denoising improves directional prediction accuracy for all models, with average gains of 2-5%.
- ROn EWZ daily, xLSTM-TS achieves 72.82% accuracy and 73.16% F1 score; on S&P 500 daily, 68.45% accuracy.
- RN-BEATS and TFT show competitive performance but lag behind xLSTM-TS, especially on volatile EWZ data.
- RHourly data predictions are less accurate than daily, with xLSTM-TS achieving around 60% accuracy on hourly datasets.
Analyze Paper
Generate insights from "An Evaluation of Deep Learning Models for Stock Market Trend...".