Think Twice Before Plugging Variables into Model

With the development of artificial intelligence, an increasing number of AI models are being applied in the financial sector. The Long Short-Term Memory (LSTM) model, as an AI model for processing time-series data, has achieved promising results in the investment field. Currently, many studies use LSTM models with inputs mainly consisting of variables such as prices, returns, and volatility, while some studies also include additional variables to improve prediction accuracy. However, these studies lack sufficient discourse on why these variables are chosen and what variables should be inputted. This is due to the lack of interpretability of the relationships between variables in AI models, resulting in a decreasing emphasis on the theoretical connection between input data and prediction results. In this study, we use LSTM models to predict stock returns, with both return and price-to-earnings ratio (P/E ratio) sequences as inputs. Based on the change in LSTM model prediction accuracy resulting from different input data, we suggest that providing more variables without selection may not necessarily lead to better prediction results. For the LSTM model, the momentum effect of the input variable sequence is related to its prediction accuracy, and grouping stocks according to P/E ratio indicators can improve the predictive performance of the LSTM model.