Meta Learning Strategies for Comparative and Efficient Adaptation to Financial Datasets

This research proposes a Meta learning framework for financial time series forecasting, designed to rapidly adapt to novel market conditions with minimal retraining. The framework operates in two stages: 1) pretraining on a diverse set of financial datasets, including stocks (e.g., MSFT, AAPL) and cryptocurrencies (e.g., BTC, ETH), and 2) fine-tuning on recent data to adapt to new markets. The model utilizes XGBoost with dynamic feature engineering, which adjusts technical indicators (e.g., Relative Strength Index, Bollinger Bands) to account for evolving market conditions. Experimental results demonstrate that the proposed framework achieves significant improvements in Root Mean Squared Error (15%) and Mean Absolute Percentage Error (10%) compared to traditional methods, such as simple moving averages and exponential smoothing. These findings highlight the framework’s robustness, scalability, and ability to manage dynamic market behaviors, making it an effective tool for both short-term traders and long-term investors. Compared to LSTM-GARCH, the proposed Meta learning model achieves an RMSE of 0.82 (versus up to 10.11), an MAE of 0.61 (versus up to 8.39), and a DA of 67.33% (versus up to 50.44%).