Reproduce “StockBot 2.0: Vanilla LSTMs Outperform Transformer-based Forecasting for Stock Prices” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shaswat Mohanty
A study from Stanford University determined that vanilla LSTMs generally achieve higher predictive accuracy and more stable trading decisions compared to various Transformer-based and attention-augmented models for stock price forecasting. This robust performance was observed when models were evaluated under common default hyperparameters and deployed autoregressively within a trading simulation.
Key insights
- Vanilla LSTMs, with their strong recurrent inductive bias, are particularly well-suited for financial time-series forecasting in data-limited and noisy environments.
- Architectural simplicity and stability can be more critical than theoretical expressiveness, especially when models are deployed recursively and hyperparameter optimization is constrained.
- Effective trading decisions require stable predictions with minimal error accumulation, rather than merely high raw predictive accuracy from isolated predictions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.