Datasets for: "Synapar: A Large Decision Model Utilizing a Hierarchical Temporal-Spatial Architecture for Financial Time-Series Trading Decisions"

Abstract Financial time-series trading decisions are persistently challenged by the highly non-stationary nature and low signal-to-noise ratio. This manuscript introduces Synapar, a Large Decision Model (LDM) utilizing an innovative hierarchical temporal-spatial (HTS) architecture, designed to address these challenges. We propose a novel training paradigm that transforms imitation learning into a supervised learning problem. This methodology first utilizes an "oracle-view" optimal strategy solver, based on dynamic programming, to generate theoretically optimal expert trajectories from complete historical K-line data. Subsequently, the Synapar model is trained to imitate these expert behaviors. The core of the model is its HTS architecture, which is engineered to perform deep feature extraction and model the dynamic evolution of market states across both spatial and temporal dimensions. This manuscript elaborates on the complete methodology, from expert trajectory generation and feature engineering to the model architecture and training paradigm. In preliminary inference validation, the v0.1.0-mini-alpha version of the model achieved average end-point returns of 34.68% and 49.67% (using maximum context lengths of 256 and 512, respectively) on 1-day frequency data from the main test set. On an independent CSI 300 test set, it achieved average end-point returns of 14.00% and 43.92% under the same lengths. While the model is currently closed-source, a public inference demonstration page https://synapar-0-1-0-mini-alpha-frontend.vercel.app and a public API endpoint are provided for further research and evaluation. Dataset for Synapar v0.1.0-mini-alpha This repository contains the complete dataset for the training, validation, and inference-testing of the Synapar v0.1.0-mini-alpha model. This dataset is released in support of the research paper: "Synapar: A Large Decision Model Utilizing a Hierarchical Temporal-Spatial Architecture for Financial Time-Series Trading Decisions". The data is organized into a primary training set, a primary test set, and several independent test sets to evaluate model generalization, including CSI 300, S&P 500, and China futures continuous contracts.