Core Insights - The article discusses the development and application of the Kronos model, a Time-Series Foundation Model (TSFM) specifically designed for financial market data, particularly K-line data [3][9][17] - Kronos aims to address the challenges of low signal-to-noise ratio and strong non-stationarity in financial time series data, which often hinder the performance of general-purpose models [3][9] - The model employs a two-phase framework: K-line tokenization and autoregressive pre-training, allowing it to effectively learn the complex "language" of financial markets [12][13][17] Summary by Sections Introduction to TSFM - TSFMs have emerged from the success of large-scale language models in NLP and CV, focusing on pre-training on diverse time series data to create a general-purpose model adaptable to various tasks [2][6] - The key advantages of TSFMs include their generalization and transfer learning capabilities, enabling them to learn universal time patterns and trends from vast datasets [2][6] Overview of Kronos Model - Kronos is tailored for financial K-line data, utilizing a "domain pre-training + fine-tuning" approach to deeply understand financial market characteristics [3][9] - The model's architecture includes a specialized tokenizer and a large autoregressive Transformer model, which learns the syntax and dynamics of financial data [9][12][17] Performance Evaluation of Kronos - Initial tests of the Kronos standard model on major A-share indices showed a high correlation between predicted and actual closing prices, with a Spearman correlation coefficient of 0.732 for the 5-day forecast [4][19] - The model's predictive performance improved significantly when fine-tuned, achieving a Spearman correlation of 0.856 for the same forecast [4][39] Application of Kronos in Timing Strategies - The article explores the application of Kronos in constructing timing strategies based on predicted closing prices, specifically for the CSI 1000 index [30][33] - The strategy generated positive returns, but it missed significant upward trends since July 2025, indicating a reliance on prior index reversal logic [30][33] Enhanced Performance with Fine-Tuning - A fine-tuned version of Kronos demonstrated a 33.9% return in 2025, with an annualized excess return of 9%, outperforming the original method by over 20 percentage points [5][42] - The fine-tuning process involved adjusting model parameters and rolling adjustments to better adapt to market conditions, leading to improved predictive accuracy [34][42] Conclusion - Kronos represents a significant advancement in financial time series forecasting, effectively capturing the complexities of financial data and translating predictions into actionable investment strategies [17][42]
中金 | 大模型系列（5）：大语言时序模型Kronos的A股择时应用