BiLSTM Stock Forecasting & Dynamic Trading Strategy
This peer-reviewed study compared LSTM and bidirectional LSTM models for stock-price forecasting, then tested whether the forecasts could support a dynamic portfolio-selection strategy. The experiments covered 11 stocks in 8 sectors and multiple training horizons.
Research design
| Question | Can bidirectional sequence modeling improve stock forecasts and support a stronger dynamic allocation rule? |
| Universe | 11 stocks across 8 sectors |
| Models | LSTM and BiLSTM |
| Horizons | Six-month, one-year, and two-year training windows |
| Strategy tests | 3-day, 5-day, 7-day, and 14-day rebalancing |
| Publication | 2024 5th International Conference on Machine Learning and Human-Computer Interaction (MLHMI) |
The trading experiment began with a $100,000 portfolio. At each rebalancing point, the dynamic strategy used model forecasts to select a top-five portfolio and was compared with a static version of the portfolio.
Strategy results
| Rebalancing window | Static comparison | Dynamic strategy |
|---|---|---|
| 3-day | $119,694 | $123,630 |
| 5-day | $115,444 | $117,425 |
| 7-day | $112,918 | $126,078 |
| 14-day | $109,754 | $123,306 |
The dynamic strategy finished above the static comparison in every tested window, with the largest differences in the 7-day and 14-day experiments. This wording replaces an earlier unsupported summary claim and reports only the outcomes shown in the paper.
Interpretation and limitations
The project links model evaluation to a decision rule rather than stopping at forecast error. The results remain a historical simulation: they should be interpreted with market-regime, transaction-cost, and execution assumptions in mind and are not investment advice.
Publication
Jiayi Liu, Yufeng Yang, Teng Lin, Chuanhui Peng, and Yancong Deng. “Enhancing Stock Price Forecasting and Trading Strategy through Bidirectional LSTM Integration.” 2024 5th International Conference on Machine Learning and Human-Computer Interaction (MLHMI), pp. 22–25.
Attribution: Teng Lin is a co-author of the five-author paper.