查看文章

Stock prices are characterized by high volatility and abrupt changes in trends, posing challenges for traditional forecasting models. Long Short-Term Memory (LSTM) networks are commonly considered state-of-the-art models for such predictions. However, they may struggle to handle sudden and drastic price trend shifts. Moreover, the open, high, low, and close (OHLC) prices of stocks have a few inherent constraints that have not been extensively studied in the literature. We argue that predicting the OHLC prices for the next day is more informative than predicting the overall trends, as trends are usually derived from these OHLC prices. Our focus is on Buy-Today Sell-Tomorrow (BTST) trading, and we propose using Autoencoders (AEs) pre-trained with stock prices to address the problem. We present a novel framework where a pre-trained encoder is integrated into a multi-task predictor network, creating a hybrid network that can handle OHLC constraints and capture sudden price changes effectively. The experiments include recommending the most profitable and overbought stocks for the next day (using William's% R), and our model’s recommendations achieved no losses during a test period of 300 days for multiple Indian companies.