Why LSTMs Can Mislead in Stock Price Forecasting

Summary

This article argues that LSTM models trained on historical prices can appear to forecast stocks accurately because predicted daily prices often remain close to the previous day’s price. When plotted against actual prices, this persistence can look impressive, even though small daily errors may make the predictions unusable for profitable trading. The article cautions that fitting price levels is not the same as finding a tradable signal, and questions claims of near-perfect predictive accuracy.

It links this critique to weak and semi-strong forms of market efficiency, while acknowledging that it does not claim markets are fully efficient and leaves high-frequency trading as an open question. As an alternative use for LSTMs, it describes analyzing company reports and subjective research alongside financial measures, then evaluating resulting portfolios through tracking and backtesting. The account reports initial positive results but gives no quantitative evidence or detailed validation procedure. Its conclusions are the author’s view, not proof that price-based models cannot work or that report analysis will outperform.

Key ideas

• Price-level forecasts can look accurate because they closely track the latest observed price.
• Visual similarity between forecast and actual prices does not establish that a signal can earn money.
• The article treats weak and semi-strong market efficiency as reasons for skepticism about simple historical-price prediction.
• It proposes applying LSTMs to company reports and analyst research alongside financial measures.
• The claimed initial portfolio results are not quantified, and the article supplies no detailed validation evidence.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.