Stock Market Prediction with Long Short-Term Memory Networks: A Multi-Market Performance Analysis
Härkönen, Veeti (2025-05-14)
Stock Market Prediction with Long Short-Term Memory Networks: A Multi-Market Performance Analysis
Härkönen, Veeti
(14.05.2025)
Julkaisu on tekijänoikeussäännösten alainen. Teosta voi lukea ja tulostaa henkilökohtaista käyttöä varten. Käyttö kaupallisiin tarkoituksiin on kielletty.
avoin
Julkaisun pysyvä osoite on:
https://urn.fi/URN:NBN:fi-fe2025053056206
https://urn.fi/URN:NBN:fi-fe2025053056206
Tiivistelmä
This thesis investigates the use of Long Short-Term Memory (LSTM) models in financial trading, assessing their performance in both individual stock level and portfolio allocation across three markets: the New York Stock Exchange (NYSE), India’s National Stock Exchange (NSE), and Finland’s OMX Helsinki (OMXH). The study evaluates whether LSTM-driven strategies can outperform passive benchmarks such as buy-and-hold and equally weighted portfolios.
Results show that the LSTM model-based trading algorithm underperformed buy-and-hold strategies in individual stock trading across most markets, mainly due to limited predictive accuracy and difficulty capturing sustained trends. However, when applied to portfolio allocation, specifically through Rank-Based Allocation (RBA) and Exponentially Rank-Based Allocation (ERBA), the models consistently outperformed the equally weighted benchmark. This suggests that while the LSTM model struggled with absolute price prediction, it was more effective at ranking stocks by relative performance. When factoring transaction costs into the active allocation methods, costs must remain below 0.005%-0.01%, depending on the market and method, to outperform the passive benchmark.
The study also examines whether market characteristics like size and economic development influence the model’s outcomes. Although predictive accuracy tended to decline in the larger and possibly more efficient NYSE stock market, allocation strategies performed best in it, indicating that prevailing market trends may matter more than structural factors.
Lastly, the research explores the effect of signal sensitivity, finding that increased trading signal frequency generally improved prediction accuracy, particularly in OMXH and NSE, though its impact on returns remained limited.
In conclusion, while LSTM-based models are not consistently effective for individual stock trading, they show promise in portfolio allocation. The study highlights limitations such as a short backtesting period and exclusion of real-world events like slippage, recommending that future work explore enhanced model configurations, alternative allocation methods, and longer testing horizons.
This study utilised AI-based tools: Grammarly was used to check and improve the language of the thesis, while ChatGPT assisted with coding tasks, debugging, and explaining programming concepts related to the trading models.
Results show that the LSTM model-based trading algorithm underperformed buy-and-hold strategies in individual stock trading across most markets, mainly due to limited predictive accuracy and difficulty capturing sustained trends. However, when applied to portfolio allocation, specifically through Rank-Based Allocation (RBA) and Exponentially Rank-Based Allocation (ERBA), the models consistently outperformed the equally weighted benchmark. This suggests that while the LSTM model struggled with absolute price prediction, it was more effective at ranking stocks by relative performance. When factoring transaction costs into the active allocation methods, costs must remain below 0.005%-0.01%, depending on the market and method, to outperform the passive benchmark.
The study also examines whether market characteristics like size and economic development influence the model’s outcomes. Although predictive accuracy tended to decline in the larger and possibly more efficient NYSE stock market, allocation strategies performed best in it, indicating that prevailing market trends may matter more than structural factors.
Lastly, the research explores the effect of signal sensitivity, finding that increased trading signal frequency generally improved prediction accuracy, particularly in OMXH and NSE, though its impact on returns remained limited.
In conclusion, while LSTM-based models are not consistently effective for individual stock trading, they show promise in portfolio allocation. The study highlights limitations such as a short backtesting period and exclusion of real-world events like slippage, recommending that future work explore enhanced model configurations, alternative allocation methods, and longer testing horizons.
This study utilised AI-based tools: Grammarly was used to check and improve the language of the thesis, while ChatGPT assisted with coding tasks, debugging, and explaining programming concepts related to the trading models.